机器学习学术速递[9.16]

AI 总结：本文从反向传播视角分析Z-loss，提出反向传输诊断框架，分离源幅度与传输因子，解释梯度差异，并在GPT-2和Pythia模型上验证架构感知变体可减少尾部并保持困惑度。

链接：https://arxiv.org/abs/2609.16179

作者：Bum Jun Kim

英文摘要：Z-loss has been widely applied to the logits of language-model output heads and sparse mixture-of-experts routers. Z-loss constrains the softmax log-normalizers of these output heads and routers, thereby limiting large-logit excursions, reducing finite-precision roundoff exposure, and avoiding training-loss divergence. These use cases arise in modern Transformer settings where large-vocabulary softmax heads, top-$k$ routing, fused losses, and mixed-precision optimizers interact. Z-loss has typically been understood only as a scalar penalty on the log-normalizer. This paper instead analyzes Z-loss from a backward-pass perspective, focusing on the gradients produced by the Z-loss penalty. The logit-space gradient, which we call the backward source, is injected at the logit boundary of the Z-loss branch of backpropagation; consequently, the backward source's effect depends on the architecture and implementation through which the gradient is transported. We develop a backward-transport view for Z-loss that separates the source's scalar amplitude and softmax shape from the transport factors. These factors include common-shift coordinates, tied-embedding pathways, output-to-hidden gain, fused-loss source consistency, optimizer-facing updates, and top-$k$ router reduction scale. These diagnostics show that nearly identical forward Z-loss values can coexist with distinct logit-space Z-loss gradients and, after architectural and optimizer transport, distinct parameter updates. The transport diagnostics also explain why raw-logit Z-loss can reduce scalar tails without changing output-to-hidden gain and why active-route reductions alter the effective router coefficient. Across evaluations of models in the GPT-2 and Pythia families on WikiText-103 and FineWeb-Edu, architecture-aware variants reduce backward-geometry tails while maintaining comparable validation perplexity in low-coefficient regimes.

AI 总结：通过匹配状态分解，发现卷积主导联想记忆，秩1转移优势在匹配后消失；距离课程和精度门控斜坡可突破干扰墙，将锁定率从1/10提升至7/10。

链接：https://arxiv.org/abs/2609.16183

机构：Purdue University（普渡大学）； Obit Research

作者：Julian Boesch, Andrew Wee

英文摘要：Fixed-state recurrences--linear attention and state-space models--are reported to lag behind attention on associative recall, but whole-architecture comparisons cannot say which ingredient is responsible. We decompose masked multi-query recall at a fixed state budget along three single-knob axes: a short causal convolution, the transition structure (rank-1 delta rule vs. diagonal), and decay. The convolution dominates (~+0.5 recall in both families under matched training): comparisons that pit convolution-free cells against a convolution-equipped Mamba measure the missing convolution, not the recurrence. The rank-1 transition beats its diagonal ablation by +0.19/+0.32 at 16/32 pairs, but the margin shrinks to +0.03 once both cells carry the convolution, and a state-matched Mamba-2 ties the unarmed rank-1 cell: no class claim survives. Cells that solve 32-pair recall degrade gracefully with load yet fall to chance retrieving 4 pairs from a distractor haystack--flat across lengths and transitions. Interference under sparse supervision, not capacity: a distance curriculum takes the unchanged architecture from 0.021 to 1.000. Training is a lock-in lottery--a seed either locks in or does not--and the curriculum is the lever. Lock-in rises from 1/10 to 7/10 (p=0.02); dense supervision adds nothing; at L=256 a shaped ramp reopens a boundary the uniform curriculum cannot (4/5 vs. 0/9); and at L=512, where the ramp collapses (0/6), gating it on measured accuracy locks in 6/6 (p=0.001). Bidirectional denoiser cells, reading the query before the haystack, show no measurable advantage over causal training (ten seeds), and collision-key retrieval needs two layers. Arming for recall is free on an S_5 state-tracking guardrail--the armed cell is significantly better at every depth (p<=0.0044). These replace "recurrent models are bad at recall" with a measured decomposition and two cheap interventions.

AI 总结：提出诱饵方向优化（DDO），一种无需微调的后置权重编辑防御，通过注入非线性诱饵信号破坏消融攻击的估计器，在多个模型上显著降低攻击成功率且成本极低。

链接：https://arxiv.org/abs/2609.16204

机构：Carnegie Mellon University（卡内基梅隆大学）

作者：Aashiq Muhamed, Mona T. Diab, Virginia Smith

英文摘要： Safety guardrails in open-weight language models can be readily bypassed using Refusal Feature Ablation (RFA), a technique that identifies and projects out a linear refusal direction from the residual stream, often achieving a high attack success rate (ASR) while preserving model capability. Defending against these attacks typically requires computationally expensive safety finetuning for every new checkpoint. We introduce Decoy Direction Optimization (DDO), a fast, post-hoc weight-editing defense that requires no base-model finetuning. Our approach is based on a simple mechanistic insight: ablation attacks rely on contrastive estimators to find the refusal direction. Rather than trying to hide the true refusal circuitry, DDO actively injects a high-magnitude, nonlinear decoy signal into the network's MLP neurons. When an attacker attempts to locate the refusal direction, the decoy corrupts their estimator, tricking them into ablating a harmless orthogonal feature while the actual safety mechanism remains intact. We prove a spectral bound formalizing this effect and evaluate DDO across six model families, achieving <10% ASR under standard RFA. On Llama-3-8B-Instruct, DDO remains comparable to trained defenses under adaptive multi-phase attacks (65% vs. 58% worst-case ASR) and reduces Heretic weight-level attack ASR from 88.7% to 18%, all at 30 to 450 times lower optimization cost per configuration than the trained baselines.

AI 总结：本文提出注意力均值场分析，用平均注意力核预测表示几何演化，并通过均值场偏差揭示上下文特定计算，在归纳和少样本任务中验证了其有效性。

链接：https://arxiv.org/abs/2609.16382

机构：Boston University（波士顿大学）

作者：Micah Adler, John W. Byers, Mark Crovella

英文摘要：A language model's representation geometry is not predetermined; it evolves as the model runs. A faithful account of that geometry must capture that dynamic process, and so cannot be based solely on model-independent statistics such as co-occurrence. Here we introduce a mean-field analysis of attention. The average attention from one token to another defines a kernel that carries representations layer to layer and can be iterated through the network to model how the geometry is transformed. We condition this average two ways. Conditioned on a whole corpus, the kernel predicts the average-case evolution of representation geometry. Conditioned instead on a single context, it predicts the expected geometry for that context. A head's departure from that prediction, its \emph{mean-field deviation}, isolates the context-specific computation that the mean field misses. Under the corpus-conditional reading, the kernel yields an open-loop model: from the input embeddings and the frozen weights alone, we can iterate the kernel and the model's own MLPs over token representations, never consulting a measured deviation at any layer. The resulting prediction is highly accurate. In early training the model and its corpus mean field are indistinguishable. Replace every attention head with its mean field, and the substitution leaves the loss on real text unchanged. Around the onset of induction, the two diverge, and the gap widens as representations become contextualized. Under the context-conditional reading, deviation from the mean field is a task-agnostic measure of context-specific computation. The residual decomposes additively into unusual attention routing and contextualization of the transported values. Across controlled induction and few-shot settings, greater deviation tracks greater reliance on in-context information.

AI 总结：本研究通过分解层重要性为必要性和可塑性，发现Transformer中两者反对齐而SSM中重叠，且对齐符号预测下游适应行为，影响灾难性遗忘。

链接：https://arxiv.org/abs/2609.16537

机构：Chandar Research Lab（Chandar 研究实验室）； Mila - Quebec AI Institute（Mila - 魁北克人工智能研究所）； Université de Montréal（蒙特利尔大学）； Polytechnique Montréal（蒙特利尔理工学院）

作者：Istabrak Abbes, Nizar Islah, Irina Rish, Sarath Chandar

英文摘要：Transformers and state-space models (SSMs) are the two dominant families of sequence models, and a central open question is how far the analytical knowledge built for transformers transfers to SSMs. We address this through the lens of layer importance which underpins compression, selective fine-tuning, and interpretability across both families. We decompose layer importance into two distinct notions. \emph{Necessity} captures how much the pretrained model depends on a layer's existing contribution, measured by the loss increase from bypassing it. \emph{Plasticity} captures where the model absorbs new information during fine-tuning, measured by the magnitude of task-specific weight updates. Our analysis reveals that the two families behave fundamentally differently: in every evaluated residual transformer up to $14$B parameters, Necessity and Plasticity anti-align across depth, whereas in the evaluated Mamba-style SSMs they point to overlapping regions. The sign of this alignment also predicts downstream adaptation behavior. In the evaluated transformers, concentrating updates in the most plastic layers increases catastrophic forgetting, while this tier-dependent effect disappears in the evaluated Mamba-style SSMs.

AI 总结：本研究通过理论和实验揭示，门控机制使状态空间模型偏向记忆化而延迟上下文学习，但有助于长序列泛化，为改进线性时间模型提供依据。

链接：https://arxiv.org/abs/2609.16540

机构：Apple（苹果公司）； Harvard University（哈佛大学）

作者：William L. Tong, Aryo Lotfi, Emmanuel Abbe, Kostas Vaggelakos, Vishnu Banna, Etai Littwin, Josh Susskind, Cengiz Pehlevan, Eran Malach

英文摘要：State Space Models (SSMs) have emerged as a compelling alternative to Transformers, enabling sequence modeling with constant memory and linear compute. Although SSMs exhibit reasonable performance and favorable computational characteristics, they continue to lag behind Transformers on tasks that require in-context learning and precise retrieval, slowing their adoption for large-scale language modeling. In this work, we demonstrate that both the success and failure of SSMs in these domains can be explained by studying the role of the gating mechanism, a prevalent component in modern recurrent networks. Specifically, we show through theory and experiments that this gating mechanism causes SSMs to first learn an in-weights "memorization" solution, while delaying, or even preventing, convergence to a correct in-context learning solution. Importantly, this happens even in cases where there are no fundamental limitations due to the architecture or its memory capacity. On the other hand, we find that gating is often beneficial for improving generalization to long sequence lengths. Our results illuminate the crucial role of the gating mechanism in shaping both the training dynamics and generalization of SSMs, and provide a basis for understanding and improving linear-time models.

AI 总结：提出生成式基础模型SOTER，融合跨通道耦合、频谱引导专家与连续时间潜演化，在可穿戴生理信号上实现最优零样本预测、分类与插补性能。

链接：https://arxiv.org/abs/2609.16804

机构：Zhejiang University（浙江大学）； Shanghai Innovation Institute（上海创新研究院）； Fudan University（复旦大学）； Huazhong University of Science and Technology（华中科技大学）

作者：Fangke Chen, Sirry Chen, Wei Chen, Zhongyu Wei

英文摘要：Time-series foundation models have demonstrated strong cross-domain transfer, yet their common architectural assumptions remain poorly aligned with wearable physiological signals, which are multichannel, irregularly sampled, noisy, and governed by coupled continuous-time dynamics spanning distinct spectral scales. We present SOTER, a generative foundation model for wearable physiological time series that unifies cross-channel coupling, spectrum-guided expert specialization, and continuous-time latent evolution within a single pre-training framework. SOTER combines a spatial feature-aware backbone that models inter-signal dependencies, a power spectral density (PSD)-guided mixture-of-experts layer that routes representations to experts associated with fixed spectral bands through an inspectable, non-learned rule, and a neural controlled differential equation decoder that supports prediction and imputation at arbitrary timestamps. We pre-train SOTER on 226 billion time points from five public physiological datasets and evaluate the same pre-trained model across out-of-distribution zero-shot forecasting, frozen-encoder linear-probe classification, and continuous-time imputation on wearable benchmarks. SOTER achieves the best RMSE on 4 of 6 datasets and the best MAE on 5 of 6 in zero-shot forecasting, the highest average Macro-AUROC in classification, and the lowest imputation error on all six datasets at 75% missingness. It further remains robust to additive acquisition noise, matching or surpassing baselines evaluated on clean inputs even under the strongest corruption. These results indicate that domain-specialized foundation models for wearable physiology benefit from jointly modeling channel structure, spectral scale, and continuous-time dynamics.

AI 总结：针对在线策略蒸馏中目标保真度与稳定性权衡，提出γOPD方法，通过折扣时间信用分配和奖励兼容有界混合机制，在数学与代码推理任务上超越现有方法。

链接：https://arxiv.org/abs/2609.16937

机构：School of Vehicle and Mobility & College of AI, Tsinghua University（清华大学车辆与运载学院与人工智能学院）； Didi Voyager Labs, DiDi Autonomous Driving（滴滴自动驾驶沃芽实验室）

作者：Shiqi Liu, Zeyu He, Letian Tao, Guojian Zhan, Jiaxin Gao, Feihong Zhang, Jingliang Duan, Wei Xiong, Kehua Sheng, Bo Zhang, Yang Guan, Shengbo Eben Li

英文摘要：On-policy distillation (OPD) has emerged as an effective approach for large language model post-training, yet existing objectives face a trade-off between objective fidelity and optimization stability. Token-level OPD provides stable but local supervision, whereas sequence-level OPD captures future credit at the cost of horizon-dependent variance. We establish a unified temporal-credit view of these formulations, showing that practical token-level OPD can be interpreted as a temporal approximation to the sequence-level reverse-KL gradient. Building on this connection, we propose $\gamma$OPD, which uses discounted temporal credit assignment to balance long-horizon supervision and optimization stability, while admitting a horizon-independent variance bound. We further develop a reward-compatible bounded mixing (RBM) mechanism for $\gamma\mathrm{OPD}$ that balances verifiable outcome feedback with the discounted OPD advantage to move beyond purely teacher-dependent optimization. Experiments on mathematical and code reasoning demonstrate consistent improvements over existing OPD methods across vanilla, size-mismatched, and multi-teacher distillation settings.

AI 总结：针对开源语言模型无法证明可复现的问题，提出完全可审计的训练机制，通过确定训练非确定性顺序实现异构硬件上的比特级复现，并发布Open-1B模型及全套审计工具。

链接：https://arxiv.org/abs/2609.17380

机构：Gensyn

作者：John Donaghy, Brian Wilcox, Oğuzhan Ersoy, Shikhar Rastogi, Adam St Arnaud, Alexey Titov, Jordan Greenberg, Ben Fielding, Harry Grieve

英文摘要：Open-source language models have a reproducibility problem. Despite releasing weights, training data, and recipes, none of them are provably reproducible due to the non-associativity of floating-point arithmetic. Deep learning frameworks often offer a deterministic execution mode, allowing reproducible operations on the same machines. Unfortunately, this determinism does not carry across hardware such that a user can verify that a released checkpoint was actually produced using the declared training recipe. This leaves room for undisclosed data, injected biases, or backdoors that existing techniques such as proof-of-learning or proof-of-training-data cannot rule out. We introduce a new tier of model transparency, fully auditable, in which every operation on every data sample during training is independently reproducible on heterogeneous commodity hardware with bitwise certainty. By imposing a definite order on the sources of training nondeterminism, GPU kernel reductions, data batch ordering across a data-parallel cluster, and inter/intra-node collective communication, we make it possible to replay any individual step of a large, distributed training run on a single piece of commodity hardware and check it against the published trajectory. Because replaying an entire run on one machine is infeasible, we support this with a collective verification scheme in which many independent auditors each certify individual steps, together covering the whole run. We release Open-1B, a model trained under this regime, together with its full pretraining dataset, every intermediate checkpoint, the training codebase, and the audit harness needed to reproduce and verify any step of its training.

AI 总结：HintMiner利用自监督学习的语言模型从问答帖子中自动挖掘提示，通过Transformer和复制机制生成提示，在6万条Stack Overflow问题上取得BLEU 36.17%和ROUGE-2 36.29%的效果。

链接： https://arxiv.org/abs/2609.16060

作者：Zhenyu Zhang, JiuDong Yang

英文摘要：Users often need ask questions and seek answers online. The Question - Answering (QA) forums such as Stack Overflow cannot always respond to the questions timely and properly. In this paper, we propose HintMiner, a novel automatic question hints mining tool for users to help them find answers. HintMiner leverages the machine comprehension and sequence generation techniques to automatically generate hints for users' questions. It firstly retrieve many web Q\&A posts and then extract some hints from the posts using MiningNet that is built via a language model. Using the huge amount of online Q\&A posts, we design a self-supervised objective to train the MiningNet that is a neural encoder-decoder model based on the transformer and copying mechanisms. We have evaluated HintMiner on 60,000 Stack Overflow questions. The experiment results show that the proposed approach is effective. For example, HintMiner achieves an average BLEU score of 36.17\% and an average ROUGE-2 score of 36.29\%. Our tool and experimental data are publicly available.

AI 总结：本文提出一种信道感知神经网络，从IQ测量中提取二进制密钥特征，利用多任务学习与射线追踪增强，在POWDER测试台上验证了低比特不一致性和高密钥多样性，实现去中心化无线密钥生成。

链接：https://arxiv.org/abs/2609.16341

机构：Center for Connected Autonomy and AI, Florida Atlantic University（佛罗里达大西洋大学互联自主与人工智能中心）； Air Force Research Laboratory（空军研究实验室）

作者：Jose Angel Sanchez Viloria, George Sklivanitis, Dimitris Pados, Elizabeth Serena Bentley

英文摘要：Physical-layer key generation (PKG) enables wireless devices to establish shared keys from reciprocal channel observations without directly exchanging the key. This capability is attractive for edge networks, where distributed and resource-constrained devices may require lightweight key establishment with limited access to centralized infrastructure. We introduce a channel-informed neural network for PKG that derives binary key features directly from received IQ measurements while explicitly grounding the learned representation in the underlying multipath channel. The proposed multi-task recurrent neural network jointly learns reciprocity-preserving binary features and an auxiliary channel estimate using a training objective that combines deep metric learning with channel-informed supervision. Structured channel sounding enables channel estimation from over-the-air measurements, while Sionna-RT ray tracing is used to augment training with additional propagation conditions. We evaluate the framework using indoor and outdoor software-defined-radio measurements collected on the POWDER radio testbed. Across all evaluated scenarios, the proposed model produces lower bit disagreement for reciprocal Alice-Bob observations than for Eve-related observations. Ray-traced data augmentation substantially improves key diversity, increasing the unique-key rate to 0.94, 0.99, and 0.99 across the indoor and two outdoor scenarios, respectively. Successfully reconciled channel-informed keys pass the selected NIST randomness tests prior to SHA-3 privacy amplification. The results demonstrate the potential of channel-informed representation learning for decentralized wireless key establishment while highlighting an important tradeoff between key diversity and reconciliation reliability.

AI 总结：HyCoSeq提出一种上下文双曲表示学习框架，通过加权洛伦兹残差聚合和双向LSTM整合序列上下文，在多种基因组任务上超越现有双曲基线，且无需大规模预训练即可媲美大型DNA语言模型。

链接：https://arxiv.org/abs/2609.16925

机构：ShanghaiTech University（上海科技大学）

作者：Chenhao Zeng, Zhibin Pu, Shufei Ge

英文摘要：Hyperbolic geometry provides a natural inductive bias for genomic representation learning, but existing hyperbolic genomic models primarily use Lorentz convolutions to learn local sequence representations, while their residual pathways do not directly aggregate full Lorentz representations. We propose HyCoSeq, a contextual hyperbolic representation learning framework for genomic sequences. HyCoSeq incorporates weighted Lorentzian residual aggregation into multi-curvature Lorentz encoding, allowing full Lorentz representations to participate directly in geometry-consistent local aggregation. It further introduces a bidirectional long short-term memory network that integrates information from both sequence directions to learn contextual relationships among local representations at different positions within a genomic sequence, thereby extending local hyperbolic convolutional encoding to sequence-level contextualized representations. Extensive experiments across diverse genomic tasks show that HyCoSeq outperforms existing hyperbolic baselines and, without large-scale genomic pretraining, achieves competitive performance against substantially larger pretrained DNA language models.

AI 总结：提出分布式JEPA自监督框架，通过掩蔽潜在表示预测与正则化，在异构能源时间序列上实现稳定表示，优于Transformer基线并增强鲁棒性。

链接：https://arxiv.org/abs/2609.17029

作者：Liana Toderean, Tudor Cioara, Vasilis Michalakopoulos, Efstathios Sarantinopoulos, Ionut Anghel, Elissaios Sarmas

英文摘要：Traditional energy forecasting solutions rely on task-specific supervision and energy asset representations, limiting transferability and the ability to capture general temporal dynamics across heterogeneous assets. We address this by proposing a distributed Joint Embedding Predictive Architecture (JEPA) for self-supervised learning from heterogeneous energy time-series. The framework predicts latent representations of masked temporal segments while integrating temporal observations and contextual information within a shared embedding space. To prevent representation collapse, training combines a latent-space predictive objective with covariance and temporal variance regularization. The evaluation was conducted on energy consumption and generation datasets under data-degradation scenarios and compared with a Transformer forecasting baseline. The learned representations remained stable (cosine similarity $\approx 0.98$; effective rank 185-235). JEPA achieved performance comparable to a Transformer on building energy data, higher $R^2$ in 3/5 consumer clusters, and outperformed the baseline on 9/10 unseen PVs ($R^2$=0.73-0.88 vs. <0.45), while showing greater robustness to missing data.

AI 总结：针对硬件替换导致物理实现改变而逻辑身份不变的问题，提出FreqSpaNet网络，利用SFPF的频率与空间特征进行开放集异常检测，平均AUROC达96.31%，优于基线9.05个百分点。

链接：https://arxiv.org/abs/2609.17491

作者：Xiaoxuan Huang, Jinlong Xu, YiZhe Wang, Meng Zhang, Xian Li, Yuying Bian

英文摘要：Unauthorized hardware replacement can preserve a wireless device's logical identity while altering its physical implementation, posing a challenge to hardware integrity verification. Spatio-frequency polarization fingerprints (SFPFs) capture device-dependent responses across multiple frequencies and directions, but their frequency and spatial dimensions exhibit different structural dependencies. We propose FreqSpaNet, an SFPF representation learning network for open set hardware anomaly detection. A frequency branch captures local variations among neighboring frequencies, while a geometry-aware spatial branch models directional relationships using angular information. The two representations are combined through adaptive fusion, and complementary pretraining further captures shared information while preserving the distinct characteristics of the frequency and spatial representations. Experiments show that FreqSpaNet achieves a mean AUROC of 96.31\%, 9.05 points above the baseline. Results under seven hardware replacement scenarios further verify the effectiveness of FreqSpaNet.

AI 总结：本研究提出将行为知识作为前提贝叶斯网络注入神经符号强化学习，并比较验证器、执行器、学习器三种放置策略，实验表明能显著提升解决方案质量。

链接：https://arxiv.org/abs/2609.16056

机构：Institute of Distributed Intelligent Systems（分布式智能系统研究所）； University of the Bundeswehr Munich（慕尼黑联邦国防军大学）； School of Engineering（工程学院）； The University of Western Australia（西澳大利亚大学）

作者：Norbert Oswald, Fabian Deuser, Thomas Bräunl

英文摘要：Humans carry behaviour knowledge of how to act in familiar situations into every new task rather than relearning it from scratch. There is no reason a Reinforcement Learning (RL) agent shouldn't do the same: known behaviour patterns need not be learned, only applied. Neuro-symbolic RL bridges prior knowledge and RL by injecting symbolic knowledge alongside a learned policy. The point at which this knowledge is integrated is critical: a poor choice can produce, for instance, hallucinated preconditions, which surface as safety and reliability problems in agents acting in changing environments. We formalise this behavioural knowledge as a precondition Bayesian network (BN) over the agent's \emph{structural actions} - the actions whose legality depends on preconditions, such as picking up a key, grasping a block, toggling a door, or dropping an object. The BN restricts when these actions may fire, and we inject it into the RL loop at three placements: (1) a \emph{symbolic verifier}, consulted only at inference, that fires a structural action once its preconditions hold; (2) a \emph{symbolic enforcer}, active during both training and inference, that governs structural-action use throughout learning; and (3) a \emph{symbolic learner}, which folds the knowledge into the network and learns the restriction and use of structural actions itself. To test the three variants we run experiments on two benchmarks with opposite regimes: one built on long, ordered planning chains, the other on continuous manipulation. We compare against strong baselines on solution quality, sample efficiency, and traceability. The payoff is substantial. On MiniGrid, all three placements improve the \emph{solution quality} over the PPO+RND baseline, the symbolic enforcer leading at $98.2\%$ against the baseline's $88.8\%$. On Fetch, $\dots$

AI 总结：本研究提出stopgrad回归原理，统一流映射、强化学习和扩散采样器的stopgrad目标函数，证明其平稳点唯一性及收敛性，并提出改进放置方式使训练内存减半。

链接：https://arxiv.org/abs/2609.16222

机构：Frontiers Research（前沿研究）； Prescient Design, Genentech（基因泰克公司前瞻设计部门）； Flatiron Institute（熨斗研究所）； MIT（麻省理工学院）； NYU（纽约大学）

作者：Max W. Shen, Mark Goldstein, Zichu Wang, Aahlad Puli, Rajesh Ranganath

英文摘要：Stopgrads are widely used in training machine learning models, but stopgrads can alter the gradient, stationary points and convergence guarantees of the original objective, which can make stopgrad training theoretically ungrounded. We introduce a stopgrad regression principle, which identifies a general template for stopgrad objectives with a closed-form characterization of stationary points and their uniqueness, unifying stopgrad objectives for flow maps, reinforcement learning, and diffusion samplers. We provide theoretical grounding for optimizing stopgrad flow map objectives by showing their unique stationary point is the true flow map, and showing positive convergence results for Eulerian and Lagrangian objectives, including MeanFlow and improved MeanFlow. Remarkably, we show that under functional semi-gradient flow, the learned flow map has a closed-form expression composing the initial flow map and the true flow map. We additionally use our stopgrad regression principle to propose modified stopgrad placements for flow map objectives which reduce training memory by 2x.

AI 总结：本研究利用基于模型的强化学习，在重力驱动平台上实现了液滴在复杂几何结构中的自主导航，策略从有限数据中学习，并成功迁移到新路径，为智能微流控化学实验室奠定基础。

链接：https://arxiv.org/abs/2609.16369

作者：Rajneesh Anand, Mayuresh V. Kothare

英文摘要： Precise manipulation of liquid droplets underpins lab-on-a-chip platforms for diagnostics, chemical synthesis, and biological assays. Yet autonomous droplet transport through confined geometries of varying complexity remains an open challenge. Droplets exhibit contact-angle hysteresis, deformability, and capillary pinning, which make their response to actuation nonlinear and history dependent, that classical controllers and pre-programmed trajectories cannot cope in multi-turn environments. Here we demonstrate autonomous navigation of a liquid droplet through geometries of increasing complexity on a gravity driven (Labyrinth) platform using model-based reinforcement learning. A thin silicone oil film reduces contact-line pinning while two-axis tilt supplies the gravitational driving force, and an overhead camera tracks the droplet in real time. An offline-trained policy discovers effective tilt strategies from limited physical interaction data, without simulation or analytical droplet models. The system operates under partial observability, as oil-film thickness, instantaneous contact angle, and droplet deformation state remain hidden from the controller. Despite these challenges, the learned policy achieves reliable navigation across straight, right-angle, and curved-arc paths, including outside-corner geometries. We further demonstrate that a policy trained on a simpler geometry transfers to complex ones, succeeding zero-shot on right-angle and staircase paths and reaching full success on a curved arc with a fifth of the training data. The findings suggest promising avenues for enabling droplet based microfluidic systems to serve as intelligent chemical laboratories.

AI 总结：针对强化学习自回归生成瓶颈，提出GrowMTP，利用RL训练中rollout分布窄及验证监督信号，在RL循环内从零训练草稿头，实现最高2.13倍rollout加速。

链接：https://arxiv.org/abs/2609.16648

机构：WeChat AI, Tencent（腾讯微信人工智能实验室）； Peking University（北京大学）

作者：Minghua He, Lingzhe Zhang, Yuan Liu, Xiao Zhou, Aiwei Liu

英文摘要：Reinforcement learning (RL) post-training drives the frontier capabilities of large language models, with its wall-clock dominated by autoregressive rollout generation. Speculative decoding is an established remedy for this bottleneck, but existing draft heads must be pretrained or warmed up before RL, introducing substantial training cost outside the RL run to be accelerated. We observe that RL training itself provides both conditions required for online draft-head training: its rollout distribution is far narrower than that of pretraining, and its verification step continuously produces supervision signals aligned with this distribution. Building on these observations, we propose GrowMTP, which uses this supervision to train a draft head from scratch entirely within the RL loop, with all head updates detached from the policy backbone. On Qwen3-4B (no draft head), MiMo-7B-SFT (weak head), and Qwen3.5-4B-Base (strong head), GrowMTP achieves rollout speedups of 2.13x, 1.93x, and 1.36x, and end-to-end speedups of 1.60x, 1.41x, and 1.20x, respectively. GrowMTP therefore serves existing RL training frameworks as a modular component, particularly offering a from-scratch acceleration path for models without pretrained draft heads.

AI 总结：针对生成式评分标准作为奖励信号的鲁棒性，提出ImpossibleRubrics基准，通过对抗性测试揭示评分标准被利用源于对错误细节的具体化，而非模糊性。

链接：https://arxiv.org/abs/2609.16816

机构：National University of Singapore（新加坡国立大学）； Peking University（北京大学）； Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所）

作者：Bowen Qin, Yi Xie, Yesheng Liu, Xi Yang

英文摘要：Language model-generated rubrics are increasingly used as reward signals for rubric-based reinforcement learning, LLM-as-a-judge evaluation, and automated grading. Such rubrics are reliable only if they reward honest answers over adversarial answers optimized to exploit them. Yet their robustness to such optimization remains poorly understood. We isolate the hardest regime: impossible tasks, where the prompt pressures the model toward an unsupported conclusion, so the only honest response is to acknowledge the impossibility. We introduce ImpossibleRubrics, a benchmark of 169 impossible tasks spanning six impossibility categories, each paired with a verifiable oracle certificate specifying what an honest answer may and may not claim, together with 48 answerable controls. Rather than providing fixed rubrics, ImpossibleRubrics provides task environments and certificates, allowing rubrics to be generated downstream and then adversarially tested for whether they reward certificate-violating answers. Eleven generators are exploited 8--26% of the time on the unbiased 150-of-169 environment cut; on a deliberately selected stress cut the strongest generator we measured is still exploited 36% while a certificate-faithful rubric is exploited 0%, so what we measure is a rubric-quality gap, not task impossibility. One result runs against intuition. A single generic rubric ("be decisive, penalize hedging") used unchanged for every task is exploited 64% of the time, and seven of the eleven generators are exploited more often than that while writing a rubric tailored to each one. The tailored criteria appear to tell an attacker which claim to fabricate. The problem is not that rubrics are vague; it is that they are specific about the wrong things.

AI 总结：针对去中心化异构老虎机中的非平稳性，提出DRFC算法，通过平衡全局比较忽略局部变化，实现无局部变化依赖的遗憾界，并验证其有效性。

链接：https://arxiv.org/abs/2609.16824

作者：Zhaojun Peng

英文摘要：Decentralized bandit systems often contain heterogeneous agents: rewards can change at individual agents even when the best action for the network stays the same. These local changes may cancel when rewards are averaged across agents, so the number of local changes $\Stloc$ can be much larger than the number of changes in the best common arm $\Stdec$. We introduce Decision-Relevant Fresh Comparison (DRFC), which uses new, balanced samples from all agents to compare arms at the network level and switches only when fresh global evidence indicates that the common best arm has changed. We prove a high-probability dynamic regret bound with no adaptation term depending on $\Stloc$, and show that every algorithm must still pay for identifying genuine decision switches and propagating them through the communication graph. Under a distinct time-average benchmark, an anytime-valid sliding-window extension handles gradual drift; experiments on synthetic, semi-real, and MovieLens-1M replays show that DRFC ignores decision-irrelevant local changes while the extension avoids false switches.

AI 总结：本研究探讨语言模型在奖励报告异常时，能否利用验证记录区分诚实与撒谎的报告者；发现模型能有效识别撒谎者，但常误判诚实者，且受表面特征影响，失败程度超预期。

链接：https://arxiv.org/abs/2609.17226

机构：The University of Texas at Dallas（德克萨斯大学达拉斯分校）

作者：Arman Nik Khah

英文摘要：An agent that learns from rewards has to trust whatever reports those rewards. When the reports suddenly change, either the world changed or the reporter broke. From the reports alone these are indistinguishable, and reinforcement learning theory shows that no amount of further experience separates them. The prescribed escape is richer data about the reporter itself. We ask whether a frozen language model, handed exactly that data, uses it. We build a two-option game in which a payout swap and a lying reporter produce byte-identical histories. Then we add one verified record: an independent check of one round's real result, printed beside what the reporter said about that round. That single line settles the case. We ask three large models, from two families, to answer one question with one letter. Is the reporter honest or lying? They catch a lying reporter almost perfectly. At the 70B class that holds in every condition we tried; the 32B model slips in one wording. They clear an honest reporter far less often, and how often depends on things that should not matter. Averaged over rounds, letters, and wordings, a 72B model calls an honest reporter a liar 38% of the time when nothing has changed at all, and 58% of the time when the payouts moved. A 70B model from a second family calls an honest reporter a liar 26% and 48% of the time. The failure is not one of reading, because in the situation where nothing changed the same models score 0.96 to 1.00 with the answer printed in the prompt. Which surface feature drives it differs by family. For the Qwen models it is which round the record names, and for Llama it is which letter stands for "honest." Adding the record to a prompt that already states the answer makes Llama less likely to give that answer. We had registered a prediction for that 58% before the run: 35%. The failure is larger than we expected.

AI 总结：针对现有表格生成器忽视语义约束的问题，提出语义一致表格扩散框架，利用LLM提取列内语义和列间规则作为生成条件，在统一语义空间进行条件去噪，在六个基准上同时提升分布保真度与语义一致性。

链接：https://arxiv.org/abs/2609.16069

机构：Jilin University（吉林大学）； The Chinese University of Hong Kong, Shenzhen（香港中文大学（深圳））

作者：Yili Wang, Ruxue Shi, Mengnan Du, Hangting Ye, Yi Chang, Xin Wang

英文摘要：Synthetic tabular data can match real data distributions while still violating the semantic constraints that govern valid tabular rows. This reveals a key limitation of existing tabular generators: they mainly optimize distributional fidelity, but do not explicitly model weak semantic priors encoded in tabular schema and textual descriptions. In this paper, we propose \ours, a semantics-consistent tabular diffusion framework for high-fidelity synthetic data generation under weakly specified semantic priors. \ours\ first constructs two types of priors, namely intra-column semantics and inter-column symbolic rules, with LLM-assisted extraction from metadata and validation on the real training split. These priors are then used as generation conditions rather than post-hoc filters. Specifically, \ours\ maps heterogeneous column values, column identities, and semantic priors into a unified semantic space, and performs column-wise forward corruption and prior-conditioned reverse denoising to preserve both marginal distributions and rule-consistent cross-column dependencies. Extensive experiments on six real-world tabular benchmarks show that \ours\ consistently improves distributional fidelity, semantic consistency, and downstream task utility over representative VAE-, GAN-, LLM-, and diffusion-based baselines. Additional analyses further demonstrate the robustness of \ours\ when semantic priors are partially unavailable.

AI 总结：本研究利用GroupAffect-4数据集，提出伪标签增强方法以应对协作群体情感感知中的稀疏标签问题，实验表明该方法优于仅标签基线，且人格信息宜作为团队内选择机制。

链接：https://arxiv.org/abs/2609.16077

机构：IT University of Copenhagen（哥本哈根信息技术大学）； GN Advanced Science, GN Hearing（GN 先进科学，GN 听力）

作者：Meisam Jamshidi Seikavandi, Tanya Ignatenko, Fabricio Batista Narcizo, Paolo Burelli, Jesper Bünsow Boldt, Andrew Burke Dittberner

英文摘要：Physiological affect sensing in naturalistic group interaction is often limited by sparse labels rather than sensor data: wearable devices produce many time windows, while self-reports are collected only a few times per session. Using GroupAffect-4, a four-person collaborative dataset with wearable physiology, eye tracking, Big Five personality, and post-task VAD labels, we study pseudo-label augmentation for affect sensing under sparse supervision. We compare no augmentation, Gaussian Process pseudo-labelling, personality-aware trust weighting, and joint personality-plus-confidence weighting within a shared target-construction pipeline. Results show that pseudo-label augmentation improves over the labelled-only baseline in the known-team setting. However, the narrow range of Big Five cosine similarities (0.91-0.99) makes fine-grained personality weighting ineffective; personality similarity functions mainly as a same-team filter rather than a calibrated trust signal. With smoothing, augmented SVM variants are effectively tied on Valence and Arousal, while the joint personality-plus-confidence variant gives the highest Dominance score. Cross-subject LOSO transfer remains encouraging, especially for Arousal, whereas strict session-isolated LOGO removes the augmentation benefit. Given only 10 groups, LOGO should be interpreted as a conservative lower bound on unseen-group transfer. Overall, the results suggest that pseudo-label augmentation can make better use of sparsely labelled collaborative affect data, while personality information is most useful as a within-team selection mechanism.

AI 总结：本文提出统一生成模型滤波框架，基于输运实现分析步骤，推导三种新滤波器，并与传统方法比较，发现生成滤波器能处理多模态后验，但无单一最优方法，需根据在线预算和集合规模选择。

链接：https://arxiv.org/abs/2609.16317

机构：University of Washington（华盛顿大学）； Morgan Stanley（摩根士丹利）

作者：Mohammad Al-Jarrah, Wei Deng, Bamdad Hosseini, Amirhossein Taghvaei

英文摘要：This letter presents a unified formulation and a controlled numerical comparison of generative-model approaches to the nonlinear filtering problem. Under this formulation the analysis step is realized by a transport of the forecast distribution to the posterior, the approaches differing only in how that transport is selected and learned. We derive three new filters, based on stochastic interpolants, their deterministic flow-matching limit, and Schrödinger bridges realized through forward--backward SDEs. We develop a two-stage tuning procedure that separates the training of the generative model from its online refinement. The resulting methods are compared against the optimal transport filter (OTF), the Knothe--Rosenblatt filter (KRF), the sequential importance resampling (SIR) particle filter and the ensemble Kalman filter (EnKF), in terms of accuracy, computational time, and sensitivity to ensemble size and state dimension. The results indicate that every generative filter resolves multimodal posteriors that the EnKF and SIR do not, that no single generative framework dominates, the preferred method being set by the available online budget and ensemble size, and that the filters differ in the regularity of the particle trajectories they produce.

AI 总结：针对扭曲高斯过程流式处理中参数优化与解析可处理性的权衡，提出基于精确递归梯度计算的在线方法，联合更新潜在GP矩与扭曲参数。

链接：https://arxiv.org/abs/2609.16472

机构：Simula Metropolitan Center for Digital Engineering（Simula城市数字工程中心）； Simula Research Laboratory（Simula研究实验室）； Institute of Science Tokyo（东京科学大学）

作者：Emilio Ruiz-Moreno, Konstantinos Slavakis, Baltasar Beferull-Lozano

英文摘要：Warped Gaussian processes (GPs) handle non-Gaussian observations by mapping them into a latent standard GP via a parametric transformation called warping. Existing streaming variants, however, either optimize the warping parameters periodically or sacrifice analytical tractability for a higher model capacity. To bridge this gap, we show that the gradient of the instantaneous negative log-likelihood of a warped GP admits an exact recursive computation. Based on this result, we propose a novel online method for warped GPs that jointly updates the latent GP moments and optimizes the warping parameters.

AI 总结：FlowATC提出一种仅基于历史ADS-B轨迹训练的流匹配架构，用于空中交通轨迹预测，通过序列修复生成分布，在参数匹配时优于扩散模型和CVAE基线，并支持概率占用估计。

链接：https://arxiv.org/abs/2609.16528

机构：University of California, Berkeley（加州大学伯克利分校）

作者：Mathurin Petit, Emir Torun, Louis Brusset, Jordan Kam, Alexandre M. Bayen

英文摘要：Building accurate decision-support tools for next-generation air traffic control requires robust trajectory prediction models. We present a flow-matching architecture trained exclusively on historical aircraft trajectories, with no route labels or chart supervision. Trained on 1.15 million Automatic Dependent Surveillance-Broadcast trajectory windows collected over the San Francisco Bay Area, the model generates aircraft trajectory distributions that closely match historical traffic, reproducing known airspace structure around San Francisco Airport such as the shape of SFO's published NIITE FOUR departure procedure. Our model is trained directly on the native, irregular ADS-B sampling interval. Trajectory prediction is cast as sequence inpainting using a block-causal Transformer that denoises future state tokens conditioned on the observed history using Conditional Flow Matching or Denoising Diffusion Probabilistic Models. We compare our architecture against constant-velocity, deterministic-Long Short Term Memory, and Conditional Variational Autoencoders baselines. At matched parameter count, CFM outperforms DDPM by 11-26% in minADE@20, and both generative objectives surpass the CVAE baseline by 31-41%. We further show that the error degrades gracefully with prediction horizon, and the architecture remains effective when retrained on temporally decimated feeds. Lastly, we sample $K$ independent completions, yielding spatial probabilistic occupancy estimates that can serve as input to downstream conflict-risk estimation.

AI 总结：针对多模态时空预测中采样率不同、模态缺失和自回归误差累积问题，提出AsyncCouple-Flow，结合令牌稀疏化、异步耦合图和流匹配，在天气和交通基准上优于现有方法。

链接：https://arxiv.org/abs/2609.16573

机构：Institute of Computing Technology, Chinese Academy of Sciences（中国科学院计算技术研究所）； Emory University（埃默里大学）； University of California, Berkeley（加州大学伯克利分校）； Yale University（耶鲁大学）； Stevens Institute of Technology（史蒂文斯理工学院）； City University of New York（纽约城市大学）

作者：Zhixiang Wu, Yining Liu, Bo Zhao, Szu-Yu Chen, Huiran Duan, Chu Lin, Chuanguang Yang

英文摘要：Multi-modal spatio-temporal forecasting (MM-STF) supports weather nowcasting, traffic prediction, and earth-system modeling by combining heterogeneous sources such as physical fields, satellite imagery, and in-situ sensors. Three obstacles persist: (i) modalities have different spatio-temporal sampling rates, forcing lossy interpolation onto a unified grid; (ii) modalities are frequently missing at deployment due to sensor outages or revisit gaps, while most methods train with full availability; and (iii) autoregressive decoders accumulate errors over long horizons, amplified by multi-modal conditioning. We propose AsyncCouple-Flow to address these issues jointly. A Modality-Aware Token Sparsification (MATS) module performs scale-aware tokenization and uses a shared importance scorer to select top-k tokens per timestep, producing equal-length sequences. An Asynchronous Cross-Modal Coupling Graph (ACCG) replaces fixed cross-attention with a learnable graph whose edges encode time offsets, semantic similarity, and modality-specific physical priors, enabling fusion under arbitrary asynchrony and missingness. A Flow-Matching Forecasting Head models multi-step prediction as a conditional ODE, trained with stochastic modality dropout and integrated jointly to avoid autoregressive drift. Experiments on ERA5+GOES+ISD weather forecasting and PEMS-BAY traffic prediction with multi-source side information show that AsyncCouple-Flow outperforms state-of-the-art baselines and remains robust with up to two missing modalities. The code will be released upon acceptance.

AI 总结：MyoFlow提出判别式流匹配框架，通过锚定修正流实现跨会话和跨受试者的HD-sEMG手势识别，无需独立分类头，显著提升准确率。

链接：https://arxiv.org/abs/2609.17194

机构：Waseda University（早稻田大学）； KDDI Research Inc.（KDDI综合研究所）； The University of Manchester（曼彻斯特大学）

作者：Chenhao Wu, Dingjie Peng, Satoshi Funabashi, Satoshi Konishi, Wuqiang Yang, Hiroshi Onoda, Hironori Washizaki, Jiang Liu

英文摘要：High-density surface electromyography (HD-sEMG) gesture recognition supports prosthetic control, assistive robotics, and rehabilitation, but electrode re-donning and physiological variability cause distribution shifts that degrade accuracy across sessions and subjects. Generative HD-sEMG models primarily synthesize signals for augmentation; although diffusion models enhance representation learning, prediction still relies on a separate classifier. To tie learned dynamics to the decision rule, we propose MyoFlow, the first discriminative flow-matching framework for HD-sEMG recognition across sessions and subjects. It recasts classification as anchor-tied transport: a domain-conditioned rectified flow moves encoded windows toward gesture anchors that serve as transport targets and define the nearest-anchor decision geometry, enabling zero-shot prediction without an independent head. On the Hyser dataset, MyoFlow improves mean cross-session and cross-subject accuracy over the strongest diffusion-based baseline by 4.24\% and 6.37\%, respectively, and achieves 91.71\% mean zero-shot accuracy and 97.39\% mean few-shot accuracy across multiple days on the CEMHSEY dataset.

AI 总结：本文从优化视角研究流匹配中路径选择对SGD收敛速率的影响，发现即使目标函数相同，不同路径可根本改变收敛速率，并提出方差最小化路径选择方法PathOpt，理论结果经实验验证。

链接：https://arxiv.org/abs/2609.17287

机构：Applied AI Institute（应用人工智能研究所）

作者：Alexander Tyurin

英文摘要：In flow matching (FM), a velocity model $v_{\theta}$ is trained using a predefined path $g_t$ that connects data and noise samples (e.g., $g_t(x_0, x_1) = (1 - t) x_0 + t x_1$). In this work, we study the choice of this path from an optimization perspective by analyzing the variance of stochastic gradients. We consider the class $G(p_t,v^\star_t)$ of paths that induce the same marginal distributions $p_t$ and marginal velocity field $v^\star_t$, and therefore the same FM objective. Our main finding is that the choice of path $g_t$ can fundamentally change the convergence rate of SGD, even when the FM objective remains exactly the same. (i) For a linear velocity model and one-dimensional Gaussian data, we derive a tight bound on the SGD iteration complexity up to logarithmic factors and find an analytically optimal path that minimizes this bound among linear paths inducing the same FM problem. (ii) We then extend the variance analysis to general FM problems and formulate path selection at a fixed $\theta$ as the variance-minimization problem PathOpt$_\theta$, constrained to $g_t\in G(p_t,v^\star_t)$. We show that this constraint is essential: reducing variance without it can lead to slower convergence. (iii) Since the constraint $g_t \in G(p_t,v^\star_t)$ cannot generally be verified directly, we derive an equivalent formulation with constraints that can be estimated from samples, allowing paths to be found numerically. Our theoretical results are supported by experiments with Gaussian data, Gaussian mixture models, and real datasets.

AI 总结：提出混合变分量子电路，结合经典仿射层实现高效多元回归，理论证明其逼近能力，实验在图像重建和Friedman1基准上媲美高斯过程回归并优于XGBoost和随机森林。

链接：https://arxiv.org/abs/2609.17358

作者：Koffi Ognandon Ayena (ICB), Frédéric Holweck (ICB), Serge Iovleff (UR4662), Amah S d'Almeida

英文摘要：Variational quantum circuits (VQCs) are parameterized quantum circuits optimized classically. We propose a hybrid variational quantum circuit (HVQC) extending VQCs with a classical affine post-measurement layer, enabling vector-valued regression without the linear overhead of independent scalar circuits. Theoretically, we show that elementary one-and two-qubit circuits can approximate quadratic functions and products via data re-uploading and entanglement, providing the foundations of the full architecture. Experimentally, on two synthetic image reconstruction datasets and the Friedman1 benchmark (40,568 test samples), our HVQC matches Gaussian Process Regression and outperforms XGBoost and Random Forest. An ablation study confirms that both quantum and classical components are essential, and results highlight the central role of the feature map in hybrid quantum-classical models.

AI 总结：针对联邦随机双层优化中计算二阶矩阵导致运行时间长的问题，提出仅用一阶梯度并采用常数单时间尺度学习率的方差缩减算法，显著降低运行时间，实验验证有效。

链接：https://arxiv.org/abs/2609.16350

机构：Temple University（天普大学）； Stony Brook University（石溪大学）

作者：Yihan Zhang, Rohit Dhaipule, Chiu C Tan, Haibin Ling, Hongchang Gao

英文摘要： Federated stochastic bilevel optimization has been actively studied in recent years due to its widespread applications in machine learning. However, most existing federated stochastic bilevel optimization algorithms require the computation of second-order Hessian and Jacobian matrices, which leads to longer running times in practice. To address these challenges, we propose a novel federated stochastic variance-reduced bilevel gradient descent algorithm that relies solely on first-order oracles. Specifically, our approach does not require the computation of second-order Hessian and Jacobian matrices, significantly reducing running time. Furthermore, we introduce a novel learning rate mechanism, i.e., a constant single-timescale learning rate, to coordinate the update of different variables. We also present a new strategy to establish the convergence rate of our algorithm. Finally, the extensive experimental results confirm the efficacy of our proposed algorithm.

AI 总结：本文通过几何分析比较梯度下降与自然梯度下降在KLR训练的Hopfield网络优化山脊上的学习路径，发现自然梯度遵循测地路径、克服不稳定并更快收敛且泛化更优，揭示了优化动力学与表征几何的关联。

链接：https://arxiv.org/abs/2609.16805

机构：Aichi Institute of Technology（爱知工业大学）

作者：Akira Tamamori

英文摘要：High-capacity associative memories based on Kernel Logistic Regression (KLR) exhibit a "Ridge of Optimization" characterized by extreme stability and a highly skewed weight spectrum. However, the dynamical process by which learning converges to this critical regime has remained unclear. This paper provides a geometric analysis of the learning trajectories on the statistical manifold of a KLR-trained Hopfield network. By comparing the paths of Gradient Descent (GD) and Natural Gradient Descent (NGD), we elucidate the mechanisms governing the optimization process. Our analysis reveals that learning on the Ridge proceeds in two distinct phases. We show that the extreme curvature of the Ridge causes standard GD to follow a highly oscillatory, non-geodesic path. In stark contrast, NGD explicitly corrects for this geometry, following the ideal geodesic path and completely overcoming the instabilities faced by GD. We demonstrate experimentally that NGD not only converges significantly faster but also achieves a solution with superior generalization performance. These results establish that the highly structured geometry of the Ridge is optimally suited for information-geometric optimization, providing a new perspective on the interplay between learning dynamics and emergent representation geometry.

AI 总结：本研究揭示核逻辑回归关联记忆的优化脊为秩1谱坍缩附近的几何奇点，梯度下降通过稳定性边缘现象动态平衡曲率，从而在高曲率边界塑造最优高容量记忆表征。

链接：https://arxiv.org/abs/2609.16827

机构：Aichi Institute of Technology（爱知工业大学）

作者：Akira Tamamori

英文摘要：High-capacity associative memories based on Kernel Logistic Regression (KLR) exhibit exceptional storage capabilities and robustness. Previous empirical studies identified a hyperparameter regime, the "Ridge of Optimization," where attractor stability is maximized. However, the geometric nature of this regime and the optimization dynamics required to reach it have remained unclear. In this paper, we investigate the static geometry of the parameter space and the learning trajectory of Gradient Descent (GD) in KLR-trained Hopfield networks. Using the eigenvalue spectrum of the Hessian, we reveal that the Ridge corresponds to a phase boundary located adjacent to a rank-1 spectral collapse, acting as a geometric singularity where the principal curvature is massively amplified. Furthermore, we demonstrate that the learning dynamics exhibit a transient self-stabilizing behavior driven by the Edge of Stability (EoS) phenomenon. Rather than seeking flat regions, the network parameters are driven toward a state where the local curvature dynamically equilibrates near the stability limit dictated by the learning rate, allowing the optimization to survive the initial instability. We provide analytical derivations for both the rank-1 asymptotic collapse and the dynamic feedback loop governing this equilibration. These findings suggest that optimal, high-capacity memory representations are not formed in flat minima, but are dynamically sculpted at the highly curved boundaries of geometric singularities.

AI 总结：针对高维昂贵黑盒函数，提出降空间多保真贝叶斯优化框架RS-MFBO，结合全局敏感性分析和保真度增强高斯过程，在工业模拟器上显著减少高保真评估次数并保持优化性能。

链接：https://arxiv.org/abs/2609.17440

机构：Imperial College London（伦敦帝国理工学院）

作者：Niki Triantafyllou, Andrea Bernardi, Maria M. Papathanasiou

英文摘要：Optimizing industrial process flowsheets is often computationally prohibitive due to the high cost of rigorous simulations and the curse of dimensionality inherent in complex design spaces. To address these challenges, we present a reduced-space multi-fidelity Bayesian optimization (RS-MFBO) framework designed for high-dimensional, expensive black-box functions. The approach integrates Global Sensitivity Analysis (GSA) for dimensionality reduction with a fidelity-augmented Gaussian process that captures correlations between low-cost approximations and expensive high-fidelity evaluations. A cost-aware acquisition strategy, augmented with cooldown and promotion mechanisms, adaptively guides the allocation of samples across fidelities. The framework is validated on two distinct industrial process simulators: a plasmid DNA bioprocess in SuperPro Designer and a green fuel synthesis plant in Aspen HYSYS. Results across diverse economic and physical objectives demonstrate that the proposed method substantially reduces the number of high-fidelity simulator evaluations while maintaining competitive optimization performance compared to single-fidelity baselines. These results highlight RS-MFBO as a scalable, simulator-agnostic approach for cost-constrained black-box optimization.

AI 总结：该研究提出智能体启发式学习工作室，通过示例记忆、规则形成与错误修复的认知方式，生成可执行、可检查的HAR策略，无需训练，在十一个数据集上表现强劲并支持边缘部署。

链接：https://arxiv.org/abs/2609.16065

机构：RPTU University Kaiserslautern-Landau（凯泽斯劳滕-兰道莱茵兰-普法尔茨工业大学）； Northwestern Polytechnical University（西北工业大学）

作者：Siyu Yuan, He Zhang, Sizhen Bian, Bin Guo

英文摘要： Human activity recognition (HAR) is usually framed as gradient-based training of neural networks. Agentic Heuristic Learning (AHL) Studio explores a complementary view inspired by human cognitive learning: people learn activities by remembering examples, forming rules, and repairing mistakes, not by backpropagating. This proposed tool implements AHL for HAR: a learning-time agent reasons over sensor protocols, proposes executable heuristic policies, records repair traces, and exports an LLM-free policy for edge deployment. We focus on the HAR benchmark family and provide an end-to-end workflow from dataset observation to edge-oriented export. On eleven HAR datasets evaluated so far, AHL policies reach strong executable-policy performance while remaining inspectable, editable, and replayable \footnote{ this https URL }.

AI 总结：针对LLM推理中内存带宽和写入耐久性挑战，提出纯整数量化与字典式KV缓存压缩算法，在Flash存内计算设备上实现高效推理，减少15倍动态KV缓存流量。

链接：https://arxiv.org/abs/2609.16161

机构：UC Berkeley（加州大学伯克利分校）； ICSI（国际计算机科学研究所）； LBNL（劳伦斯伯克利国家实验室）

作者：Sebastian Zhao, Minseo Kim, Coleman Hooper, Luca Manolache, Michael W. Mahoney, Yakun Sophia Shao, Kurt Keutzer, Amir Gholami

英文摘要：Large Language Models (LLMs) have shown impressive capabilities across a range of natural language processing tasks, and LLM inference has emerged as a critical workload for enabling downstream applications. The demands of serving LLM inference are becoming increasingly challenging as requests shift toward longer sequences and heavier inference, driven by retrieval-augmented generation, inference-time compute scaling, and long-context applications. Additionally, these challenges are compounded by hardware trends, as memory capacity and communication bandwidth are not scaling as fast as increases in workload complexity. Compute-in-Flash is a promising solution to address memory bandwidth limitations by moving computation close to memory, and to exploit the large capacity of SSD technologies. However, it is challenging to deploy LLMs on these systems as they lack support for high-precision floating point operations and have limited write endurance. In our work, we aim to address these challenges by designing inference algorithms to enable LLM inference on Flash compute-in-memory devices. We present an end-to-end integer-only quantization approach to eliminate expensive floating-point computations. To address the limited write endurance, we design a dictionary-based KV cache compression strategy based on sparse dictionary coding that represents each KV vector as a linear combination of static dictionary vectors. These algorithmic improvements enable us to exploit the benefits of Compute-in-Flash for both model weights and KV cache, and to minimize expensive data transfer operations. Across Llama-3.1-8B and Qwen-2.5-7B, our combined method exhibits limited accuracy degradation while reducing dynamic KV cache traffic by 15$\times$.

AI 总结：本研究通过精确分解与蒙特卡洛估计，分析KV缓存驱逐下首次分歧时机对累积不一致性的影响，发现SnapKV延迟分歧但分歧后TV仍高，分歧后暴露主导不匹配差距。

链接：https://arxiv.org/abs/2609.16617

作者：Xinyue Luo, Fei Yu

英文摘要：KV-cache eviction perturbs the conditional token distributions governing autoregressive generation. We investigate how first-divergence timing and subsequent token mismatch determine cumulative disagreement. We derive an exact decomposition under a specified stepwise maximal coupling: the expected mismatch fraction equals a first-mismatch contribution plus post-divergence exposure multiplied by its mismatch rate. An explicit construction over unrestricted autoregressive kernel pairs realizes the sharp interval of risks compatible with a finite divergence-aligned observation window. Residual-branch conditional Monte Carlo provides unbiased joint estimates of occurrence, occupation, and window/tail contributions, with per-replicate variance dominance for total token loss. Complete trajectories from Meta-Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct show that SnapKV at 50% retention enters divergence later and less often than SnapKV-512 or recent-token retention with the same 50% prompt-cache budget, while post-divergence total variation (TV) remains high. In an exploratory analysis of 288 documents, post-divergence exposure accounts for 85-90% of four aggregate mismatch gaps. On 288 independent documents at 90% retention, prespecified comparisons show higher branch-aligned TV in the late than in the early window in both models.

AI 总结：针对联邦学习中统计异质性导致单一全局分类器失效的问题，提出pFedKDH方法，仅聚合共享主干并保留客户端特定头部，利用重新校准的全局头部进行知识蒸馏，在多个数据集上取得最优准确率。

链接：https://arxiv.org/abs/2609.17284

机构：Universidade Federal do Ceará（塞阿拉联邦大学）； Uppsala University（乌普萨拉大学）

作者：Polycarpo Souza Neto, José Mairton Barros da Silva Júnior, Charles Casimiro Cavalcante

英文摘要：Statistical heterogeneity limits federated learning when a single global classifier cannot represent client-specific label distributions. In this work, we propose Personalized Federated Knowledge Distillation with Head Adaptation (pFedKDH), which aggregates only the shared backbone, keeps persistent client-specific heads, and uses a recalibrated global head as a teacher during local training. Across MNIST, Fashion-MNIST, CIFAR10, and CIFAR100 under class-wise Dirichlet partitions, pFedKDH obtains the best accuracy in most settings, with accuracy gaps up to 37.67\% over the weakest baseline and consistently low standard deviation across repetitions. Component-wise diagnostics and convergence results support the role of persistent heads and distillation-guided local optimization under label-skewed data.

AI 总结：针对联邦学习部分参与下鲁棒聚合的脆弱假设，提出 SWB-DM，结合修剪 Wasserstein 重心与延迟动量缓存，校准修剪比率，在多种攻击下展现鲁棒性优势。

链接：https://arxiv.org/abs/2609.16099

机构：Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology（维尔技术兰加拉詹博士萨贡塔拉研发科技学院）

作者：Saranraj S, Saranya M S, Alex David S, Ajay Kumar A

英文摘要：Robust aggregation methods for federated learning quietly rest on a fragile assumption: that whoever shows up in a given round is a fair sample of the full population. In practice, they rarely are. When only a handful of clients participate per round, even a modest fraction of adversaries can dominate that sample and silently invalidate the finite-sample guarantees that coordinate-wise median, Krum, Bulyan, and trimmed mean all depend on. We introduce SWB-DM to address this directly. SWB treats each slice of a client update as a one-dimensional distribution, computes a trimmed Wasserstein barycenter across clients, and recovers coordinate identity via a medoid-based gauge-fixing step -- a heuristic we developed and do not claim it belongs to standard optimal-transport theory. DeMoA-style delayed momentum then caches updates across the full client population each round, decoupling robustness from whoever happened to be sampled. Trim ratio calibration is not cosmetic: under-trimming causes collapse at corruption levels a properly calibrated model survives. Across 448 CIFAR-10 configurations, plus CIFAR-100, FEMNIST, and a 500-client scalability run, we find several mechanistically distinct failure modes. Even-sample coordinate-wise median degrades to a deterministic wrong answer. Krum silently violates its own n greater than 2f+2 precondition and diverges without warning. Bulyan's n greater than or equal to 4f+3 threshold produces a sharp pass/fail boundary. On attacks, IPM defeats order-statistic defenses -- including SWB -- more reliably than ALIE, confirmed through delta-space measurements against a convergence bound. SWB-DM's cache carries a real warm-up cost, but extending all baselines to the same round budget shows its CIFAR-10 gains are disproportionately large. On CIFAR-100, FLTrust benefits more -- for reasons entirely unrelated to caching.

AI 总结：提出ARIA，一种基于稀疏自编码器的测试时遗忘方法，通过轻量线性检测器门控遗忘状态，在保持模型权重不变的同时，显著改善遗忘-保留权衡，并抵抗多种遗忘后攻击。

链接：https://arxiv.org/abs/2609.16229

机构：University of North Carolina at Chapel Hill（北卡罗来纳大学教堂山分校）； Honda Research Institute（本田研究所）； Michigan State University（密歇根州立大学）

作者：Pingzhi Li, Jinhao Duan, Vaishnav Tadiparthi, Nakul Agarwal, Kwonjoon Lee, Ehsan Moradi Pari, Hossein Nourkhiz Mahjoub, Sijia Liu, Tianlong Chen

英文摘要：Machine unlearning aims to remove specific knowledge from a trained large language model (LLM) without retraining from scratch. Existing methods modify model weights via gradient ascent and its advances. While effective on certain benchmarks, these weight-based approaches exhibit a sharp forget-utility trade-off, where stronger forgetting of target knowledge can degrade model utility, and unlearned knowledge may reappear under post-unlearning fine-tuning or prompt attacks. We propose ARIA (autoencoder-gated inference-time unlearning), a test-time unlearning method that leaves model weights intact and gates access to unwanted knowledge only when generation enters a forget-related state. ARIA uses sparse autoencoder (SAE) latents to train a lightweight linear detector, then applies an interpretable intervention on triggered states with negligible test-time overhead. Empirical evaluations on TOFU, R-TOFU, and WMDP show that ARIA improves the forget-retain trade-off over weight-based baselines across both a thinking model (DeepSeek-R1-Distilled-Qwen-1.5B) and an instruction model (Gemma-3-1B-it), e.g., reducing WMDP-cyber forget-set accuracy significantly while keeping MMLU within 1% of the pre-unlearning model. We further introduce three post-unlearning adversarial attacks targeting weight-space and decoding-space recovery, and find that ARIA remains robust under all three, with forgetting changing by less than 1% under attack. A feature-level case study leveraging the interpretability of ARIA suggests that some retain degradation may reflect response styles underlying the unlearning data rather than leakage of the targeted knowledge itself, highlighting a potential source of bias in unlearning task construction.

AI 总结：DP-SPIN是一个可信策展人框架，通过将记录映射为语义草图并发布差分隐私语义计划，实现独立于受保护数据的语义概念的聚合测量与摘要生成。

链接：https://arxiv.org/abs/2609.16283

机构：Harvard University（哈佛大学）

作者：Behrooz Razeghi

英文摘要：\texttt{URANIA} provides end-to-end differential privacy (DP) for summaries of data-dependent clusters. However, its cluster--keyword release does not directly provide collection-wide aggregates for semantic concepts defined independently of the protected corpus. Records may express several concepts, records expressing the same concept may be assigned to different clusters, and cluster identities need not correspond across analyses. Consequently, cluster-level statistics do not directly provide comparable measurements of predefined concepts across collections or repeated analyses. We introduce \texttt{DP-SPIN}, a trusted-curator framework for aggregate measurement and summarization over semantic concepts fixed independently of the protected target records. Each record is mapped to a bounded sparse nonnegative vector over these concepts, whose sum forms a semantic sketch. A differentially private mechanism releases a semantic plan containing admitted concepts and noisy masses; normalized semantic-support values and support bins are obtained by post-processing. For user-level privacy, each user's aggregate contribution is clipped to a fixed bound. The language model receives only the plan and fixed decoding instructions, while a public verifier checks concept mentions, reported values, comparisons, and rank claims against the released plan. The final summary is differentially private by post-processing. We establish record- and user-level DP guarantees under add/drop and replacement adjacency. We evaluate \texttt{DP-SPIN} under record-level privacy on CFPB complaint narratives, Amazon All Beauty reviews, and Yelp restaurant reviews, and under user-level privacy on Amazon and Yelp. We compare \texttt{DP-SPIN} with non-private plan and summary references, DP keyword and category histogram baselines, and a \texttt{URANIA}-style baseline with a fixed public keyword vocabulary.

AI 总结：提出自适应贝叶斯伙伴选择（ABPS）框架，通过Beta-Bernoulli后验和UCB标准实现点对点协作，在230个非IID临床中心上以0.09倍通信成本匹配FedDyn性能，减少通信且不牺牲准确性。

链接：https://arxiv.org/abs/2609.16446

机构： Missouri University of Science and Technology（密苏里科技大学）； University of Alabama in Huntsville（阿拉巴马大学亨茨维尔分校）

作者：Navid Seidi, Satyaki Roy, Sajal K. Das

英文摘要：Federated learning (FL) in healthcare faces pronounced heterogeneity and temporal concept drift across clinical centers, where evolving patient populations and care practices shift data distributions. Existing approaches rely on persistent global communication, incurring substantial bandwidth overhead while risking negative transfer from poorly aligned peers. We propose Adaptive Bayesian Partner Selection (ABPS), a peer-to-peer framework that governs who collaborates, when, and at what cost. Each center maintains a Beta-Bernoulli posterior over prospective peers' Shapley marginal utility, ranks candidates with an Upper Confidence Bound (UCB) criterion, and forms collaborations through a lightweight propose-reject mechanism, with the option to abstain from communication when no mutually beneficial partner exists. The framework admits a stochastic decision interpretation, yielding finite-sample concentration guarantees and O(kappa log T) regret in partner selection, along with conditions under which intentional isolation is optimal under negative transfer. Lightweight extensions (head personalization, bfloat16 quantized communication, and a tunable active-set size) further improve efficiency, and a goal-aware metadata filter enables institution-specific collaboration strategies. On binary in-hospital mortality prediction over the first 24 hours of an ICU stay, with 230 non-IID clinical centers drawn from MIMIC-IV, the full ABPS-X variant matches the strongest federated baseline (FedDyn, AUROC 0.758) at 0.09x the communication cost of FedAvg, with reduced variability. A diversity-driven configuration activates intentional isolation for a substantial fraction of centers. These results show that adaptive, utility-aware collaboration reduces communication without sacrificing accuracy when centers are numerous and small, offering a scalable paradigm for healthcare FL.

AI 总结：针对多模态时间序列故障检测中模态独立处理和分布偏移问题，提出基于跨模态重构的自监督框架，利用重构误差和自适应阈值，在三个工业案例中显著提升分布外鲁棒性。

链接：https://arxiv.org/abs/2609.16314

机构：Aalborg University（奥尔堡大学）； EPFL（洛桑联邦理工学院）

作者：Magnus Munk Jensen, Dorte Hammershøi, Rafał Wiśniewski, Olga Fink

英文摘要：Fault detection is essential in industrial systems, enabling early identification of abnormal behaviour and improving safety, reliability, and operational efficiency. Modern systems increasingly rely on heterogeneous sensing modalities that capture complementary aspects of the underlying physical process. However, existing data-driven anomaly detection methods often process each modality independently or use simple feature-level fusion, limiting their ability to exploit cross-modal relationships that characterize normal system behaviour. Their performance also commonly assumes similar training and deployment distributions, whereas real-world operation is affected by changing operating conditions, environmental influences, and system degradation that induce distribution shifts and reduce detection performance, especially in unseen regimes. In this work, we propose a multimodal anomaly detection framework based on cross-modal reconstruction of heterogeneous time-series sensor data. Rather than modeling each modality independently, the framework learns system dynamics by reconstructing each modality from the others, thereby exploiting complementary information across modalities. This integrates information across sensing channels without requiring explicit temporal alignment or identical sampling rates, while improving robustness to sensor noise, missing measurements, and modality-specific disturbances. To address distribution shifts during real-world deployment, anomalies are identified using cross-modal reconstruction error and an adaptive test-time thresholding mechanism that adjusts to changing operating conditions. Experiments on three industrial case studies show strong fault detection performance and substantially improved robustness under out-of-distribution conditions, with the largest gains observed in the most challenging operating regimes.

AI 总结：本文提出变分潜在马尔可夫算子（VAMO），通过潜在分布与概率转移建模物理状态，减少长时程自回归预测中的误差累积，并在流体动力学基准上验证了其稳定性和有效性。

链接：https://arxiv.org/abs/2609.16621

机构：Duke University（杜克大学）

作者：Junyi Liao, Johann Guilleminot, Vahid Tarokh

英文摘要：Neural PDE solvers provide efficient surrogates for time-dependent physical systems, but autoregressive prediction over long horizons remains challenging because local errors can induce distribution shift and accumulate under recursive deployment. We develop a variational approach to this problem by introducing latent Markov dynamics in which physical states are represented by latent distributions and evolved through probabilistic transitions. The framework is formulated directly on function spaces and specialized to functional Gaussian models, where structured latent perturbations induce a spectral geometry and variational transition alignment regularizes the learned dynamics. We further analyze how these mechanisms affect autoregressive error propagation, providing a theoretical connection between variational training and long-horizon prediction. We instantiate the framework as the Variational Autoencoding Markov Operator (VAMO), which combines spatially resolved latent fields, structured Gaussian perturbations, and a neural-operator transition. Empirically, we demonstrate the effectiveness of VAMO on several fluid-dynamics benchmarks with prediction horizons extending substantially beyond those represented during training, where it consistently reduces error accumulation and improves rollout stability over several deterministic and noise-injection baselines. Overall, these results highlight variational modeling as a complementary approach to robust long-horizon neural PDE dynamics.

AI 总结：本研究提出一种共享循环核心与残差适配器库结合的架构，通过离散潜在代码选择适配器，在闭环控制中学习低秩运动基元，实现分布外运动泛化，显著优于多任务基线。

链接：https://arxiv.org/abs/2609.17042

机构：Columbia University（哥伦比亚大学）； New York University（纽约大学）

作者：Sreejan Kumar, Marcelo Mattar, Lea Duncker

英文摘要： Learning flexible motor primitives is a hallmark of skilled motor control. Recent neuroscience theory proposes that motor primitives may be implemented as low-rank perturbations of a shared recurrent network, but leaves open how such a system is learned. We translate this principle into a novel architecture for learning motor skills end-to-end: a shared recurrent core modulated by a bank of residual adapters, each selected by a discrete latent code. Trained on closed-loop biomechanical control, the adapters develop emergent low-rank perturbations of the recurrent dynamics despite no architectural rank constraint, placing task representations in disparate subspaces of the shared core network. A simple high-level policy over the learned options, optimized while the whole network is frozen, sequences the low-rank adapters to produce novel out-of-distribution movements. We demonstrate the ability to generalize to novel motor sequences within the closed-loop control setting, improving on the generalization error of a task-input-conditioned multitask baseline by upto order of magnitude.

AI 总结：IRENE是一种基于多尺度卷积GRU的深度学习集成模型，用于意大利1公里/5分钟分辨率的概率降水临近预报，通过重要性采样和afCRPS损失训练，其概率技能优于STEPS和DGMR基准，对抗训练改善了空间清晰度但引入长提前时间的细尺度功率过剩。

链接：https://arxiv.org/abs/2609.17175

机构：Fondazione Bruno Kessler（布鲁诺·凯斯勒基金会）

作者：Alessandro Camilletti, Gabriele Franch, Elena Tomasi, Marco Cristoforetti

英文摘要：We present IRENE (Italian Radar Ensemble Nowcasting Experiment), a deep learning model for probabilistic short-range precipitation nowcasting over the Italian domain at \SI{1}{km} spatial and 5 min temporal resolution. IRENE adopts an encoder--forecaster architecture built on multi-scale Convolutional Gated Recurrent Units (ConvGRUs), trained on the national radar composite produced by the Italian Civil Protection Department (DPC). An importance-sampling scheme focuses training on precipitation-relevant events, while the almost-fair Continuous Ranked Probability Score (afCRPS) is adopted as the primary probabilistic loss function. Two additional training configurations are proposed: an adversarial (GAN) variant, IRENE-GAN, designed to improve the spatial sharpness of the generated forecasts, and a spectrally constrained variant, IRENE-GAN-RAPSD, in which the adversarial objective is complemented by an explicit penalty on the radially averaged power spectral density. The three configurations are evaluated against the stochastic extrapolation method STEPS and the pre-trained deep learning model DGMR. All IRENE configurations attain a lower Continuous Ranked Probability Score than both benchmarks at every lead time and rank histograms closer to uniformity, indicating better probabilistic skill and ensemble calibration. In terms of ensemble-mean mean absolute error the advantage is confined to the first 90 min, beyond which the strongly damped DGMR fields and, to a lesser extent, STEPS become competitive. Spectral analysis shows that the adversarial training removes the progressive loss of small-scale variance exhibited by IRENE, at the cost of an excess of fine-scale power at long lead times that the spectral penalty only partially controls.

AI 总结：针对测试时提示调优校准不佳的问题，提出基于温度缩放的CoTS及其集成版本E-CoTS，在降低校准误差的同时提升准确率，实验验证了其有效性。

链接：https://arxiv.org/abs/2609.17386

机构：University of Chinese Academy of Sciences（中国科学院大学）； Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所）

作者：Yuwei Liang, Jian Liang, Dapeng Hu, Yinuo Xu, Ran He

英文摘要：Test-time prompt tuning (TPT) enables adaptation on a single test instance, achieving improved accuracy but often sacrificing calibration performance. Most existing calibration methods introduce additional regularization terms to promote dispersion across text embeddings and reduce calibration error, yet these methods often suffer from a drop in accuracy. Motivated by the well-calibrated nature of zero-shot predictions, we propose CoTS, a simple yet effective post-hoc calibration method that preserves accuracy. Specifically, CoTS applies temperature scaling to minimize the confidence gap between adapted and zero-shot predictions. To fully exploit the potential of multiple augmentations during adaptation, we introduce a weak-strong ensemble strategy that further boosts accuracy. We then apply CoTS to this ensemble, termed E-CoTS, to maintain its well-calibrated property. Extensive experiments on diverse datasets and backbones show that our approaches effectively mitigate miscalibration without compromising primary accuracy. For instance, E-CoTS reduces the average expected calibration error of TPT from 11.90% to 5.38% on ImageNet variants, while even increasing accuracy from 60.74% to 62.95%. Moreover, when integrated with existing calibration methods, E-CoTS usually enhances both accuracy and calibration simultaneously.

AI 总结：本研究诊断联邦图神经网络中的结构性负迁移，发现结构异常客户端因加入联邦损失过半准确率，度发散是弱相关信号，但因果干预和修复方案均未验证其有效性。

链接：https://arxiv.org/abs/2609.16977

机构：PES University（PES大学）

作者：Chethana Prasad Kabgere, Shylaja SS

英文摘要： Federated learning lets multiple participants train a shared model without pooling raw data, by exchanging locally trained model updates instead. Federated averaging assumes that averaging local models is a reasonable way to solve one shared problem when participants' data are broadly similar. Work on non-IID federated learning has shown that this assumption can withstand differences in label and feature distributions. We ask whether it survives a different strain specific to graph neural networks, where client graphs differ not in label or feature distribution but in structure itself, requiring the same shared weights to operate over fundamentally different topologies. We call the resulting harm structural negative transfer. In a federation of real citation networks and synthetic structural proxies, a structurally atypical client lost more than half its achievable accuracy simply by joining. In an initial six-client federation, two label-free structural statistics computable before training were strongly associated with this harm. Expanding to twenty clients showed that degree divergence remained associated with harm, although more weakly, and survived removal of domain contrast. Spectral divergence did not replicate, which we trace to a confound caused by the composition of the reference pool used for leave-one-out statistics. A causal intervention isolating topology found no significant effect. A degree-normalization mechanism held across twenty-four seeds but did not explain the harm when corrected. The best of five candidate fixes beat a tuned baseline only until a matched, structurally blind control was applied, after which the gain disappeared. What survives is a modest, partially replicated, degree-specific signal that is not yet a validated predictor at scale.

AI 总结：本研究将统一拓扑签名（UTS）集成到GNN训练中，通过增强、正则化和池化三种干预，突破了1-WL表达性限制，在三个基准上提升准确率最高达5.8%。

链接：https://arxiv.org/abs/2609.17061

机构：BITS Pilani（比拉理工学院皮拉尼校区）

作者：Sanyam Sanjay Jain, Anshika Krishnatray, Aditya Sharma, Vinti Agarwal

英文摘要：Message-passing Graph Neural Networks (GNNs) iteratively propagate and aggregate local neighborhood information followed by global readout to learn graph representations. However, their discriminative power is upper-bounded by the Weisfeiler--Lehman (1-WL) graph isomorphism test. This prevents GNNs from distinguishing certain non-isomorphic graphs with identical local neighborhood structures, often leading to similar graph representations. Unified Topological Signatures (UTS) capture compact, multi-scale representation of global graph topology derived from persistent homology. We introduce two complementary UTS signatures: Graph_UTS- a static signature of the input graph topology, and Embedding_UTS- a dynamic signature of the evolving embedding topology. They encode structural information inaccessible to 1-WL-based message-passing GNNs, yet their capabilities are explored solely for post-hoc embedding-space analysis. We integrate UTS into GNN training across three architectural interventions: (i) UTS-Aug: augmenting with standard readout feature that encodes graph's true topology; (ii) UTS-Reg: topological regularizer that constrains representation collapse; (iii) UTS-Pool: topology-guided pooling that retains structurally critical nodes. We further leverage UTS as a layer-wise diagnostic to quantify oversmoothing during GNN training. Theoretically, we show that integrating UTS into GNN optimization strictly extends GNN expressivity beyond the 1-WL hierarchy. Experiments on three graph classification benchmarks show consistent benefits: Graph-UTS, Dual-UTS, and UTS-Pool improve accuracy across all three datasets, Embedding-UTS provides smaller but similarly consistent gains, and UTS-Reg's benefit varies across graph domains. Accuracy improves by up to 5.8% with Graph-UTS augmentation, by up to 1.9% with UTS-Reg, and achieves comparable performance to TOGL with UTS-Pool.

AI 总结：针对涌现性错位，提出TAME框架，通过令牌归因定位高影响令牌并掩蔽，使Llama和Qwen的EM分别降低23倍和36倍，揭示信号在于表达自信度而非领域词汇。

链接：https://arxiv.org/abs/2609.16754

机构：University of California, Riverside（加州大学河滨分校）； Qatar Computing Research Institute (QCRI)（卡塔尔计算研究所（QCRI））

作者：Md Rayhanul Masud, Md Rizwan Parvez

英文摘要：Fine-tuning an aligned language model on narrow, flawed data can induce harmful behavior far outside the training domain, known as emergent misalignment (EM). Prior work has localized EM in model weights, activations, and training documents, but it remains unclear which training tokens carry the relevant fine-tuning signal. We introduce TAME (Token Attribution and Masking for Emergent Misalignment), a three-stage framework: token attribution scores how strongly the fine-tuning update raises each response token's likelihood, using forward passes through a released LoRA adapter; signal characterization finds patterns among high-attribution tokens; and causal validation tests them by attribution-guided loss masking. On released EM organisms and a 6,849-example medical-advice split, attribution is concentrated (the top 5% of tokens hold 32% of the mass) and, in Llama, depleted for medical vocabulary but enriched for a register of unwarranted certainty, even after controlling for token rarity. Masking high-attribution tokens during fresh fine-tuning cuts EM by 23x in Llama and 36x in Qwen, with the perplexity cost concentrated on the targeted register rather than on medical content; an equal random mask leaves EM unchanged. In Llama, the attribution pattern suggests that EM-relevant signal lies more in how confidently flawed content is expressed than in its domain vocabulary; the causal masking effect itself holds across both model families.

AI 总结：本文提出稀疏驱动持续学习框架CLARE，通过两阶段稀疏掩码微调缓解任务间干扰，在Omnibenchmark-1k上显著提升最终准确率，实现可扩展的类增量学习。

链接：https://arxiv.org/abs/2609.17026

机构：The University of Hong Kong（香港大学）； Hong Kong Generative AI Research and Development Center（香港生成式人工智能研发中心）

作者：Yunxiang Fu, Meng Lou, Zicheng Liao, Yizhou Yu

英文摘要：Continual learning must balance the learning of new knowledge with the retention of previously learned knowledge to incrementally learn tasks from a data stream without catastrophic forgetting. While leveraging pretrained models has significantly advanced continual learning, existing methods exhibit a scalability bottleneck when trained sequentially on many tasks, suffering from performance degradation due to inter-task interference and loss of plasticity. Inspired by evidence that sparse fine-tuning achieves performance comparable to full fine-tuning, this paper presents a novel sparsity-driven continual learning framework. Our continual learning method, termed CLARE, operates in two stages: it first identifies a sparse, task-critical parameter mask via a sparsity-inducing objective, then performs mask-constrained fine-tuning by only optimizing parameters selected by the mask. This two-stage sparse adapter mechanism enables all tasks to be accumulated within a shared adapter space while reducing destructive interference across tasks. Extensive experiments demonstrate the scalability of CLARE. On the long task-sequence benchmark Omnibenchmark-1k, CLARE outperforms strong baselines in final accuracy by a large margin, e.g, improving EASE by 4.64% and 13.34% after learning 100 tasks, respectively.

AI 总结：本研究通过跨数据集比较七种预测方法，评估时间序列基础模型在行人计数预测中的表现，发现模型选择需依据数据条件和预测视界，基础模型在数据丰富且季节性强的场景下优势明显。

链接：https://arxiv.org/abs/2609.16415

机构：Delft University of Technology（代尔夫特理工大学）

作者：Theivaprakasham Hari, Ziteng Li, Yanan Xin, Winnie Daamen, Serge Hoogendoorn

英文摘要：Pedestrian-count forecasting supports pedestrian-oriented Intelligent Transportation Systems (ITS), including crowd monitoring, pedestrian-traffic staffing and routing, and proactive risk mitigation during surges. Recent time-series foundation models (FMs) report strong zero-shot accuracy on heterogeneous forecasting benchmarks, but it remains unclear whether these gains transfer reliably to pedestrian sensing deployments. We benchmark seven univariate forecasting approaches spanning four paradigms: Seasonal Naive, gradient-boosted trees (LightGBM, CatBoost), deep learning models (N-HiTS, PatchTST), and two pretrained FMs (TimesFM, Chronos-2). Experiments cover two complementary regimes: (i) a five-day special event dataset SAIL2025 at 3-minute resolution with limited in-domain history; and (ii) Melbourne pedestrian sensors as a multi-year hourly dataset (2010--2017) with strong seasonality. We compare the MAE and RMSE results per sensor across datasets and multiple forecast horizons. Results show three consistent findings. First, with limited historical data, Seasonal Naive remains a strong baseline for long-horizon forecasting on high-volume sensors, while trained models can degrade when the next day differs substantially from prior days. Second, boosted trees can be competitive on lower-volume sensors but exhibit higher sensitivity on high-volume sensors under event-driven shift. Third, FMs excel in the seasonal and data-rich regime under long-context configuration. The findings highlight the importance of choosing pedestrian forecasting models based on both the underlying data conditions and the forecasting horizon.

AI 总结：本文提出SALVE方法，利用文本优化将阈下学习效应言语化为可读提示，在多种设置下有效检测教师特征，深化对阈下学习的理解。

链接：https://arxiv.org/abs/2609.16927

机构：Stanford University（斯坦福大学）

作者：Nathan Hu, Sanmi Koyejo, Christopher Potts

英文摘要：Subliminal learning is a phenomenon in which a distillation dataset transmits traits from the teacher model that are not legibly encoded in the dataset itself. This introduces a new challenge for model development and creates new risks from data poisoning. In this work, we use text optimization to detect subliminal learning effects and describe them as legible prompts. Subliminal learning from a prompted teacher motivates our approach. We observe that this is a special case of context distillation and leverage this observation to show that, in theory, the prompted subliminal learning dataset identifies the teacher's prompt. We reduce recovering this prompt to a text optimization problem and present a method to approximately solve it. Our method, SALVE (Search-Aided Latent Verbalization), optimizes a soft prompt, queries the same model to verbalize it as text, and uses beam search to make the verbalization reliable. In the standard subliminal learning setting, SALVE reliably recovers legible prompts that name the teacher's trait, while common text optimization methods fail to do so. In addition, we find that there are settings in which SALVE recovers the teacher's trait from a dataset even when subliminal learning fails, but that modifying student training to improve context distillation can create subliminal learning effects. We lastly show that SALVE detects subliminal learning effects in three additional settings: (1) mixtures of subliminal learning data and unrelated data, (2) data generated when the teacher is biased via activation steering, and (3) subsets of real preference data selected via Logit-Linear Selection. Overall, our results deepen our understanding of subliminal learning and present SALVE as a method to proactively detect subliminal learning effects.

AI 总结：本文提出ADORE框架，利用一阶和二阶导数统一全局与局部可解释性，通过随机SVD和动态稀疏检测提升效率，在表格、文本和图像数据上优于LIME和SHAP。

链接：https://arxiv.org/abs/2609.17171

作者：Lemen Chao, Ming Lei, Anran Fanga

英文摘要：The interpretability of complex machine learning models is of paramount importance, especially in real-world high-stakes domains such as healthcare and finance. However, existing post-hoc interpretability methods suffer from inherent limitations: fragmented analytical processes, inadequate capacity to model nonlinear feature interactions, computational inefficiencies, and over-reliance on specific model architectures. To address these challenges, this paper provides a novel method - Adaptive Derivative-Ordered Random Explanation (ADORE) - that leverages first- and second-order derivatives to accommodate nonlinear model complexities, while enabling effective capture of feature-sample interactions within a unified analytical framework. ADORE integrates global feature importance with local sample contributions, precisely quantifying feature impact by capturing both magnitude and direction, and identifying critical samples influencing model decisions. Furthermore, it achieves computational efficiency through randomized singular value decomposition (SVD) and dynamic sparsity detection, making it scalable to large, high-dimensional datasets. Experiments across three data modalities - tabular, text, and image - demonstrate that ADORE outperforms existing methods such as LIME and SHAP in handling complex interactions and computational efficiency, while providing detailed and reliable explanations. To facilitate adoption and reproducibility, ADORE has been released as an open-source Python package, hosted on GitHub, enabling researchers and practitioners to readily adapt and apply our approach to their specific tasks, models, and datasets.

AI 总结：本文提出一种基于Transformer增强神经算子（TNO）的全景性能预测框架，先预测NS方程基本参数再推得叶轮机械性能，精度优于FNO和DeepONet，计算成本降低四个数量级。

链接：https://arxiv.org/abs/2609.16066

机构：ENN Science and Technology Development China Co., Ltd.（新奥科技发展有限公司）； Hebei Key Laboratory of Compact Fusion（河北省紧凑型聚变重点实验室）

作者：Qineng Wang, Zhendong Guo, Liming Song, Tianyuan Liu

英文摘要：To enable flexible and rapid aerodynamic performance evaluation in turbomachinery design, this paper proposes a panoramic performance prediction framework. Unlike most previous prediction models that directly predict the objective functions of interest, our approach first predicts the basic parameters of the Navier-Stokes equations, such as temperature, pressure, and density. Utilizing these basic physical quantities, it subsequently predicts key performance parameters of the turbine stage meridian plane. By adopting this methodology, our proposed panoramic performance prediction framework functions similarly to a CFD simulator, capable of predicting various objective of interest to the designers. To enhance prediction accuracy, a transformer-enhanced neural operator (TNO) is introduced within this framework. Using the Rotor 37 blades as a reference, the proposed TNO is trained to predict the performance of a transonic compressor blade in the meridian plane. The TNO can accurately predict total quantities such as isentropic efficiency, mass flow, and distributions of total pressure ratio. Remarkably, the prediction error of TNO is observed to be smaller than that of state-of-the-art deep learning operators such as the FNO and DeepONet. Furthermore, the TNO is applied to downstream tasks, including sensitivity analysis and optimization of various objective functions. The results confirm that the TNO can operate almost like a CFD simulator, while reducing the computational cost of downstream tasks by four orders of magnitude. The effectiveness and reliability of the proposed TNO for solving different kinds of downstream tasks have been well demonstrated.

AI 总结：本文提出利用大型语言模型（LLMs）微调并结合检索增强生成（RAG）技术，为制造业生成合成时间序列数据，以解决标记数据稀缺问题，实验证明该方法在捕捉时间依赖性和提升异常检测等下游任务性能上优于ARIMA和LSTM等传统模型。

链接：https://arxiv.org/abs/2609.16155

机构：Liverpool John Moores University（利物浦约翰摩尔大学）； University of Texas at Arlington（德克萨斯大学阿灵顿分校）； IIT Patna（印度理工学院巴特那分校）； Stony Brook University（石溪大学）； IIT Indore（印度理工学院印多尔分校）； UMass Amherst（马萨诸塞大学阿默斯特分校）

作者：Mantek Singh, Jeshwanth Challagundla, Prateek Karnal, Gagan Ganapathy, Vineet Shah, Ridam Arora

英文摘要：This paper presents a novel framework leveraging Large Language Models (LLMs) to generate synthetic time series data for manufacturing processes. Motivated by the scarcity of labeled time-series data in real-world manufacturing settings, which hinders the development of robust machine learning models, we explore the potential of LLMs to learn complex temporal dependencies and generate realistic synthetic data. Our approach involves fine-tuning pre-trained LLMs on manufacturing process instructions and employing a Retrieval Augmented Generation (RAG) technique to enhance data diversity and realism. We evaluate our method against traditional time series modeling techniques like ARIMA and LSTMs, using quantitative metrics, PCA analysis, and downstream task performance (anomaly detection). Results demonstrate that our LLM-driven framework outperforms these baselines, generating high-quality synthetic time series data that effectively captures temporal dependencies and statistical properties of real manufacturing data, leading to improvements in downstream task performance.

AI 总结：该综述提出统一数学视角，通过分类法组织连续时间机器学习各分支，规范其公式并比较算法与权衡，识别开放挑战。

链接：https://arxiv.org/abs/2609.16710

机构：University of Science and Technology of China（中国科学技术大学）

作者：Waleed Razzaq, Yun-Sheng Zhao, Yun-Bo Zhao

英文摘要：Continuous-time (CT) machine learning has emerged as a principled framework for modeling temporal dynamics as a continuous process, particularly when observations are sampled at arbitrary time points or span long-range horizons. However, major branches of CT machine learning have matured in separate research communities, leaving their mathematical relationships and design trade-offs insufficiently characterized. In this survey, we develop a unified, concept-driven view of major CT machine learning branches through a taxonomy that organizes families according to their underlying base mathematical formulations. We present a canonical mathematical formulation that relates these families through different architectural choices of vector-field parameterization, stochasticity, memory mechanisms, and discretization. We compare training algorithms, optimization strategies, and failure modes, highlighting the trade-offs across families. We further provide a comparative analysis of theoretical computational complexity alongside an illustrative architecture-controlled benchmark analysis on representative architectures from each family. We also review software ecosystems supporting their implementation. Finally, we identify open challenges in approximation theory, training stability, hardware-efficient implementations, benchmarking, foundation models, and scientific machine learning, and discuss an agenda for future research.

AI 总结：本研究提出跨物理映射（CPM）框架，证明深度学习可在共享潜在流形上实现扩散与波场间的双向映射，但精度受方向性信息限制，U-NO和GNO分别最优。

链接：https://arxiv.org/abs/2609.16853

机构：Bundesanstalt für Materialforschung and -prüfung (BAM)（德国联邦材料研究与测试研究所）

作者：Pengfei Zhu, Julien Lecompagnon, Mathias Ziegler

英文摘要： Can deep learning translate physical fields governed by fundamentally different equations? We address this question by introducing Cross-Physics Mapping (CPM), an operator-learning framework for mappings between heterogeneous physical domains. We formulate sufficient conditions for such mappings through compatible latent representations and propose a dimensionless scaling principle that aligns the characteristic evolution scales of the source and target systems without assuming their dynamical equivalence. As a representative test, paired diffusion and wave fields are generated independently from their respective parabolic and hyperbolic equations while sharing the same latent geometry, material heterogeneity, excitation, and dimensionless scale. Seven architectures-ResUNet, DeepONet, Fourier, latent, wavelet, U-shaped, and Galerkin neural operators-are evaluated for both diffusion-to-wave and wave-to-diffusion mappings. The results reveal a strong directional asymmetry. Diffusion-to-wave reconstruction is more challenging because it requires recovering wavefront, phase, and time-of-flight information attenuated by diffusion; U-NO performs best in this direction, achieving a relative $\ell_2$ error of $0.307$ and an $R^2$ of $0.905$. Wave-to-diffusion mapping is considerably more stable, with GNO attaining a relative $\ell_2$ error of $0.154$ and an $R^2$ of $0.935$. Neural operators generally outperform the conventional convolutional baseline, highlighting the nonlocal nature of cross-physics transformations. These findings demonstrate that deep learning can establish useful mappings between distinct physical modalities on a shared latent manifold, while the achievable accuracy remains fundamentally constrained by the direction-dependent information content of the governing physics.

AI 总结：本研究提出模型无关框架，利用深度限价订单簿预测器比较反事实消息注入前后的预测分布，定义短时域市场冲击，实验显示高相关性，证明可无需重训用于场景条件响应建模。

链接：https://arxiv.org/abs/2609.16930

机构：Tampere University（坦佩雷大学）

作者：Eljas Linna, Kestutis Baltakys, Derrick Manoharan, Alexandros Iosifidis, Juho Kanniainen

英文摘要：Deep Limit Order Book forecasting models capture nonlinear market dynamics, but their ability to quantify the effects of counterfactual order book messages has not been systematically validated. We introduce a model-agnostic framework that compares a trained forecaster's predictive distributions before and after injecting mechanically valid counterfactual messages, defining short-horizon model-implied market impact. A Transformer-based forecaster recovered scenario rankings with a Spearman correlation of 0.99 and 97.2% directional agreement with realized historical outcomes among non-neutral scenarios. Observation-level analysis further showed that estimated impacts captured incremental sequence-dependent variation beyond scenario identity and the pre-event forecast. These results provide evidence that pretrained Limit Order Book forecasters can be repurposed for scenario-conditioned response modeling without retraining.

AI 总结：针对Transformer多目标跟踪的冗余计算问题，提出因果神经集合滤波（CNSF），通过仅编码当前测量并递归传递状态，结合Sinkhorn关联和卡尔曼更新，在模拟测试集上显著降低GOSPA并提升推理速度。

链接：https://arxiv.org/abs/2609.16054

机构：Hangzhou Applied Acoustics Research Institute（杭州应用声学研究所）

作者：Zhongdi Liu, Huangyu Dai

英文摘要：Transformer-based multi-target tracking (MTT) jointly learns data association and state estimation, but MT3/Track-MT3-style trackers repeatedly re-encode measurement windows, incurring redundant computation. We propose Causal Neural Set Filtering (CNSF)\footnote{\href{ this https URL }{Code: this https URL }}, a neural set filter that encodes only current measurements while carrying past evidence in a structured recursive track state. CNSF combines exclusive Sinkhorn association, association-conditioned Kalman-shaped updates with moment matching, and recurrent Bernoulli lifecycle modeling with measurement-driven birth. These mechanisms impose soft one-to-one constraints, propagate association-induced state uncertainty, and support existence estimation under missed detections and birth--death transitions. On a held-out three-regime simulated test set, CNSF reduces mean GOSPA and T-GOSPA relative to Track-MT3 by 19.3\% and 30.4\%, with 55.9\% fewer parameters and a $3.76\times$ speedup in single-thread CPU inference.

AI 总结：OmniHarness通过符号策略学习实现通用视觉生成，将验证执行抽象为可复用策略，支持中间验证与自我导向练习，在六个基准上表现优异，显著提升解决率。

链接：https://arxiv.org/abs/2609.16057

机构：Beihang University（北京航空航天大学）； The Chinese University of Hong Kong（香港中文大学）； National University of Singapore（新加坡国立大学）

作者：Xu Xu (1), Jinxiu Liu (2), Zhangbo Qiao (1), Jiaxing Lu (1), Xiangyu Zhang (1), Yubin Gu (3), Fangwei Ning (1), Yan Shi (1) ((1) Beihang University, (2) The Chinese University of Hong Kong, (3) National University of Singapore)

英文摘要：Unified multimodal large language models (MLLMs) and multi-agent systems have advanced visual generation. However, three limitations remain. (1) Existing methods often distill task-specific experience with limited generalizability. (2) Reflection is often deferred until task completion. (3) Knowledge is often acquired only in response to downstream task demands. To address these limitations, we introduce OmniHarness, a framework for generalizable visual generation via symbolic policy learning. OmniHarness abstracts verified executions into symbolic policies for visual generation task families, capturing shared procedures and applicability conditions while removing instance-specific inputs. The harness instantiates, adapts, and composes these policies for new tasks. Intermediate verification guides refinement and failure recovery during execution. Through self-directed inquiry, OmniHarness autonomously generates and executes practice tasks near its capability limits before downstream objectives are specified. Execution feedback continually refines the policies while model parameters remain fixed. Experiments across six benchmarks, three MLLM backbones, and three visual agent frameworks demonstrate strong performance and continual capability expansion. On ComfyBench's Creative tasks, OmniHarness achieves a 95.0% resolve rate, exceeding the strongest baseline by 27.5 percentage points. Frozen policy snapshots improve existing visual agent systems through plug-and-play reuse.

AI 总结：针对信号交叉口驾驶员行为，提出物理约束决策条件自回归Transformer两阶段模型，预测停车/通过决策与纵向轨迹，优于基线并实现低误差。

链接：https://arxiv.org/abs/2609.16058

机构：Karlsruhe Institute of Technology (KIT)（卡尔斯鲁厄理工学院）； University of Liechtenstein（列支敦士登大学）； Halmstad University（哈尔姆斯塔德大学）

作者：Mohammad Khoshkdahan, Pavel Laskov, Alexey Vinel

英文摘要：Red-light violations and harsh braking at signalized intersections are major contributors to traffic accidents. This paper analyzes and predicts human driver decision-making and longitudinal trajectory behavior during traffic light signal transitions. We collected a diverse real-world dataset comprising 449 approach runs under varying speed and distance conditions. Vehicle motion was recorded using RTK-corrected GNSS with centimeter-level accuracy, and driver heart rate and multi-level comfort ratings were monitored. Spatial and temporal calibration ensured precise alignment between vehicle state and signal timing. Statistical analysis identifies required deceleration as the dominant single predictor of the stop-go decision, and heteroscedastic Gaussian modeling of peak deceleration reveals five empirical comfort ranges derived from human stopping behavior. Based on this insight, we propose a two-stage modeling framework. Stage 1 predicts the binary maneuver decision, and Stage 2 generates the longitudinal acceleration trajectory using a decision-conditioned autoregressive Transformer with physics constraints, including target-state conditioning and jerk limits. The proposed architecture outperforms baseline methods and achieves 0.49m/s^2 acceleration MAE and 0.62m distance MAE. It also estimates the future stopping-comfort level of the human driver from a single yellow-onset snapshot. Qualitative results demonstrate realistic human-like braking behavior. The dataset and source code are publicly available.

AI 总结：POSPAN提出位置约束跨度掩码框架，统一现有方法，通过结合跨度长度与位置约束分布，在NLU基准上优于仅长度方法和原始MLM，并给出理论解释。

链接：https://arxiv.org/abs/2609.16061

机构：JD AI Research（京东人工智能研究院）

作者：Zhenyu Zhang, Lei Shen, Yuming Zhao, Meng Chen, Xiaodong He

英文摘要：Span-level masked language modeling (MLM) has shown to be advantageous to pre-trained language models over the original single-token MLM, as entities/phrases and their dependencies are critical to language understanding. Previous works only consider span length with some discrete distributions, while the dependencies among spans are ignored, i.e., assuming that the positions of masked spans are uniformly distributed. In this paper, we present POSPAN, a general framework to allow diverse position-constrained span masking strategies via the combination of span length distribution and position constraint distribution, which unifies all existing span-level masking methods. To verify the effectiveness of POSPAN in pre-training, we evaluate it on the datasets from several NLU benchmarks. Experimental results indicate that the position constraint is capable of enhancing span-level masking broadly, and our best POSPAN setting consistently outperforms its span-length-only counterparts and vanilla MLM. We also conduct theoretical analysis for the position constraint in masked language models to shed light on the reason why POSPAN works well, demonstrating the rationality and necessity of POSPAN.

AI 总结：本文提出带符号加权仿射p-adic残差目标作为有限域约束的原生编码，通过坐标支配定理确保全局极小值在有限域内，并以数独案例验证其有效性。

链接：https://arxiv.org/abs/2609.16063

机构：Australian National University（澳大利亚国立大学）

作者：Greg Baker

英文摘要：We study signed, weighted affine $p$-adic residual objectives as native encodings of finite-domain constraints. For primes that separate the finite alphabet, sufficiently weighted positive unary rows pin each coefficient to its allowed set, while negative rows reward unequal endpoints or clause satisfaction. A coordinatewise domination theorem places every global minimiser in the finite domain; there the loss is, up to an additive constant, the all-different conflict count or the negative number of satisfied CNF clauses. Standard Sudoku provides an $81$-coefficient case study without a one-hot lift. A client-side implementation exposes the generated dataframes, arithmetic, diagnostics, and searches.

AI 总结：提出DA-EGO算法，通过动态分解高维空间并自适应调整子空间，在有限预算下高效求解高维叶轮机械设计优化问题，并在基准和实际案例中验证有效性。

链接：https://arxiv.org/abs/2609.16067

机构：AVIC Shenyang Engine Design Institute（中国航发沈阳发动机设计研究所）

作者：Qineng Wang, Zhendong Guo, Yun Chen, Guangjian Ma, Liming Song, Jun Li

英文摘要：In order to solve the high-dimensional ($d \geq 30$) expensive black-box problems within budget, an efficient global optimization (EGO) algorithm with a dynamic aggregation strategy is proposed, labeled as DA-EGO. Specifically, the DA-EGO decomposes the original high-dimensional design space into a set of low-dimensional subspaces for efficient surrogate-based optimization search, and the optimal solutions of subspaces are combined as an elite point for the global search. Most importantly, the subspaces are not fixed. Instead, the subspace variables are updated in each iteration, according to the variable interaction analyses in the sub- and full-spaces. The perturbation method and the analysis of variance are used to detect variable interactions. To further accelerate the optimization progress, the searching ranges of subspaces are also adaptively adjusted according to the analyses of subspace optimization results of the previous iteration. Tests on 21 benchmark instances, comprising seven functions at 30, 60, and 90 dimensions, show that DA-EGO is effective on separable and partially separable problems under a budget of 1500 function evaluations. Its advantage is case-dependent: on the non-separable shifted Rosenbrock function, GSGA performs better at 60 and 90 dimensions, while the 30-dimensional results are statistically comparable to IKAEA and GSGA. Moreover, the advantage of DA-EGO is also seen in the aerodynamic optimization of a transonic rotor blade with 28 variables as well as the compressor stage optimization with 60 variables. With the above, the effectiveness of the proposed DA-EGO has been well demonstrated.

AI 总结：本研究提出SAAC-JEPA架构，解决部分传感器重叠下的跨机床CNC动力学迁移问题，发现源域预测精度不足以评估工业表示，跨机床适应需作为独立评估维度。

链接：https://arxiv.org/abs/2609.16071

作者：Ayoub Louaye Bouaziz, Matthieu Ostertag, Anton Demasles

英文摘要：Cross-machine deployment of industrial world models requires transfer across changes in dynamics, sensing interfaces, sampling regimes, and control units. We study a schema-adaptive action-conditioned Joint-Embedding Predictive Architecture (SAAC-JEPA) for CNC dynamics, where the source machine has 17 canonical sensor channels and the target shares only 10. Evaluation uses group-disjoint source splits, source-only normalization, held-out self-supervised validation, unit audits, and a sealed target test after model locking. Across five seeds, JEPA pretraining gives no clean-source forecasting gain: scratch and pretrained-body models obtain \(\mathrm{RMSE}=0.811\pm0.022\) and \(0.813\pm0.022\). A source-only search over 20 candidates selects a schema-consistent action-conditioned JEPA after seven-seed stability checks. On the confirmatory target pass, the locked model reaches zero-shot \(\mathrm{RMSE}=0.546\), \(R^2=0.012\), and \(\mathrm{NLL}=0.52\), outperforming persistence but not RevIN-equipped PatchTST and iTransformer baselines (\(0.503\) and \(0.498\)). A pre-declared paired ablation shows that RevIN in the same architecture improves RMSE to \(0.495\pm0.004\) over three seeds, but degrades target calibration (\(\mathrm{NLL}=20.6\)) on stationary context windows. A pre-lock adaptation sweep further reduces RMSE to \(0.520\) with limited target support. These results show that source-domain forecasting accuracy alone is insufficient to assess industrial predictive representations, and that cross-machine adaptation under partial sensor overlap is a distinct evaluation axis.

AI 总结：针对表格基础模型推理延迟高的问题，通过将TabPFN蒸馏为紧凑前馈模型，在保持高准确率的同时实现数千倍参数压缩，并验证软标签蒸馏的有效性。

链接：https://arxiv.org/abs/2609.16091

机构：SumUp； Johannes Gutenberg-Universität Mainz（美因茨约翰内斯·古腾堡大学）

作者：Sourish Dey, Aditya Kumar

英文摘要：Tabular foundation models deliver strong zero-training predictive performance via in-context learning, but their high inference latency makes them impractical as hot-path decision backends in interactive agentic loops. We distill a TabPFN teacher into a compact feed-forward student across a business-decision simulation on UCI Adult and five OpenML benchmarks: the classification head compresses 53.2M parameters to 8,546 (6,220x); the deployed two-head loan pipeline compresses 111.4M parameters to 17,059 (6,532x). The student retains 95.4-100.5% accuracy and 96.8-100.0% AUC, with the lowest accuracy retention on credit-g at 95.4%; an alpha = 0 hard-label control shows that the teacher's soft targets provide a 2.1-7.0 AUC point gain.

AI 总结：本研究构建确定性电商环境，通过环境锚定验证评估开放权重智能体，揭示成功指标掩盖的多种行为缺陷。

链接：https://arxiv.org/abs/2609.16093

机构：AION

作者：Nimit Shah, Haitz Sáez de Ocáriz Borde

英文摘要：A shopping conversation has many routes to the same cart, and a task-success rate reduces all of them to one score. We build a deterministic and reproducible e-commerce environment that precommits each trial's customer and trajectory parameters, including the persona, difficulty, target cart, and an item reveal schedule. A simulated consumer attempts to buy a target cart from the environment with assistance from the evaluated model. The environment guides the simulator's actions and records every assistant action alongside the environment state at that point. After the trial, these records allow the evaluator to assess individual parts of the conversation against the retained evidence. For example, the evaluator penalizes a search for failing to surface a target product only when the customer has already mentioned that product. We further use this evidence to apply different penalties to tool calls depending on how the assistant's actions compare with an expected tool-call set. Our environment also interacts with the simulator bidirectionally, reading its output to stop the trial when the simulator determines that the customer has become too frustrated and injecting directives in real time that specify when to explore, defer buying an item, or recall a previous exchange. This interaction creates an open-ended and verifiable simulation. Across eight open-weight agents from 20B to 35B parameters, with 160 trials per agent and 44 metrics, the resulting capability profiles distinguish under-action, over-purchase, unsupported product attributes, and poor search, all of which terminal success obscures.

AI 总结：本研究提出一个锁定、多信号的审计协议，用于检测代理标签信用风险预测中的监督漂移，通过五层诊断在LendingClub数据上验证，识别出患病率与概率尺度不匹配为主要时间信号。

链接：https://arxiv.org/abs/2609.16102

机构：University of Central Florida（中佛罗里达大学）； Northeastern University（东北大学）

作者：Mehrdad Shoeibi, Muhammad Shabanpour, Waldemar Karwowski, Niloofar Yousefi

英文摘要： Credit-risk models are trained on proxy labels and deployed under temporal and segment change, yet no single transfer metric separates base-rate shift, probability-scale shift, and feature-label relationship change. We contribute a design-science artifact: a locked, multi-signal audit protocol for supervision drift in proxy-labeled credit-risk prediction. Five layers (transfer performance, an oracle-gap probe, a calibration diagnostic, feature-label stability, and a synthetic positive control), thresholds, and decision rules were locked before interpretation; a bounded reading is a designed outcome. On a public LendingClub dataset (temporal 2013 to 2016 and cross-segment transfer), ranking is stable and oracle gaps are small; the clearest temporal signal is a prevalence and probability-scale mismatch that intercept-only diagnostic recalibration largely reduces, though its cause is not identifiable from the available release. The positive control responds only to larger injected shifts; subtler drift cannot be excluded. Mapping diagnostic patterns to governance actions is conceptual guidance, not validated here.

AI 总结：提出SPINE方法，用嵌入1-复形替代点集作为类原型，结合类别条件映射图与分类目标拟合，在17个基准数据集上以匹配预算超越七种原型约简方法，取得最高平均准确率，并在多数数据集上快于广义学习矢量量化。

链接：https://arxiv.org/abs/2609.16170

机构：Auburn University（奥本大学）

作者：Jordan Eckert, Henry Schenck

英文摘要：Prototype reduction replaces a training set with a smaller representation, and the established methods return a finite set of points. We propose Skeletal Prototypes on Iterative Nerve Expansions (SPINE). The model for each class is an embedded 1-complex rather than a point set. Its initial edge set is a class-conditional Mapper graph, so the data decide which localized clusters are joined. Later phases fit the vertices under a classification objective, and an observation is assigned to the class whose complex is nearest. The segments therefore enter the decision rule and not only the fitting. We evaluate SPINE on seventeen benchmark datasets under stratified 10-fold cross validation, against seven other prototype reduction methods at a matched budget. SPINE attains the highest mean accuracy and the best average rank. It is significantly better than five of the seven competitors under Wilcoxon signed-rank tests with Holm correction. A budget sweep shows that the decision rule using the entire graph segments contribute most when prototypes are scarce, while the method as a whole competes best at moderate budgets. Construction cost places SPINE with the discriminative methods, and it is faster than generalized learning vector quantization on fourteen of the seventeen datasets.

AI 总结：本研究提出一种高效推理蒸馏方法，通过约900个不确定性筛选示例和4B教师生成的合成思维链微调2B视频语言模型，在极低计算成本下超越4倍规模模型，并发现将思维链置于答案后可显著提升紧凑模型推理能力。

链接：https://arxiv.org/abs/2609.16255

机构：Liverpool John Moores University（利物浦约翰摩尔大学）； Carnegie Mellon University（卡内基梅隆大学）； Meta； University of Texas at Arlington（德克萨斯大学阿灵顿分校）

作者：Mantek Singh, Jeshwanth Challagundla, Siddharth Raina, Jasmin Jarsania

英文摘要：We present an efficient method to distill reasoning capabilities into compact video-language models (VLMs) for video question answering (VideoQA). Our approach fine-tunes a 2B-parameter model using only $\sim$900 uncertainty-selected examples, each augmented with synthetic chain-of-thought (CoT) rationales generated by a 4B teacher. Despite its minimal compute cost - under two hours on a single A100 GPU - our method enables the 2B model to outperform VLMs up to 4$\times$ larger, and generalize across CinePile, ActivityNet-QA, and MLVU, approaching the performance of its own 4B teacher. A key finding is that placing CoT rationales after the answer - contrary to standard prompting - substantially improves reasoning in compact models. This insight challenges prevailing CoT conventions and reveals new alignment strategies under limited model capacity. Our findings offer a practical blueprint for training deployable, reasoning-rich VLMs suited for mobile and edge applications.

AI 总结：本研究将记录选择视为评估任务的一部分，在GE维修、NASA ASRS和NHTSA召回系统中比较匹配记录，发现记录差异对分类性能影响显著，建议在决策点信息上评估并报告记录和标签来源。

链接：https://arxiv.org/abs/2609.16267

机构：University of the West of England（西英格兰大学）； Birzeit University（比尔宰特大学）； GE Aerospace（通用电气航空航天）

作者：Hisham Ihshaish, Peter Mayhew, Tasnim M. A. Zayet, Ana Del Amo

英文摘要：Many operational cases are documented more than once, at different workflow stages and for different purposes, yet model evaluations normally select one of these records before model comparison begins. We treat that selection as part of the evaluation and compare matched records of the same cases under fixed labels and splits in three systems: GE Aerospace repair events, NASA ASRS safety reports and NHTSA vehicle recalls. Across the three GE fields, for events whose label comes from parts transactions independently of the narratives, held-out macro-F1 ranged from 0.33 to 0.91. A difference of 0.46 separated the customer report, written before shop work, from the technician report, written after diagnosis but before the transaction that generates the label. That difference is substantially larger than the representation and architecture differences tested on the same events. The public systems showed different patterns: the NHTSA defect summary remained strongest under every model family tested, whereas the ASRS analyst synopsis outperformed the reporter narrative under learned sequence models but not under lexical baselines. Secondary analyses showed that some model comparisons were also record-dependent. Evaluations should be run on the information available at the intended decision point and should report how both the record and the label were produced.

AI 总结：本研究通过系统缩放实验，发现物理感知的增广拉格朗日训练在ACOPF代理模型中能以更高训练成本换取显著更优的约束满足度，且其性能随规模演化的规律与MSE训练不同。

链接：https://arxiv.org/abs/2609.16282

机构：Argonne National Laboratory（阿贡国家实验室）； Oak Ridge National Laboratory（橡树岭国家实验室）

作者：Yijiang Li, Emon Dey, Stefano Fenu, Massimiliano Lupo Pasini, Teja Kuruganti, Kibaek Kim

英文摘要： Learning-based surrogates for AC optimal power flow (ACOPF) promise large speedups over classical solvers, but their operational value depends on physical feasibility as much as predictive accuracy. Physics-aware objectives such as the augmented Lagrangian (AL) improve constraint satisfaction at additional per-step cost, yet how this trade-off behaves with scale is uncharacterized. We sweep model and dataset sizes under both MSE and AL training, and characterize how constraint violation changes with network size across grids. Both objectives improve as power laws, but at different rates: MSE is governed primarily by model capacity, while AL is balanced across both. Violation grows roughly twice as fast with network size under MSE as under AL. On matched hardware, AL reduces violation by nearly $30\times$ for an order of magnitude more training time, with negligible added memory. The training objective determines not only where a surrogate lands but how its quality evolves with scale.

AI 总结：本文提出漂移场网络（DFN），一种结合模拟预训练与拉格朗日平流损失微调的深度神经网络，用于从卫星观测预测海洋表面流场，显著降低粒子轨迹定位误差。

链接：https://arxiv.org/abs/2609.16288

机构：Amphitrite； Sorbonne Université（索邦大学）； The Ocean Cleanup（海洋清理组织）； INRIA（法国国家信息与自动化研究所）

作者：Théo Archambault, Pierre Garcia, Mattia Romero, Anastase Charantonis, Dominique Béréziat

英文摘要：The North Pacific Subtropical Gyre (NPSG) is a major accumulation zone for floating plastic debris, resulting from basin-scale convergent ocean circulation. Effective cleanup strategies in this region rely on accurate forecasts of Lagrangian particle drift. Here, we introduce Drift Field Net (DFN), a deep neural network that predicts ocean surface flow fields from operational satellite observations. DFN is trained using a novel two-stage strategy that combines pretraining on simulated data with Lagrangian fine-tuning based on an advection-consistent loss function. This physics-informed optimization directly improves the accuracy of particle trajectory predictions. We evaluate DFN against an operational physics-based forecasting system and demonstrate the potential of deep learning for ocean surface flow prediction. On in situ drifter trajectories, DFN reduces the mean positioning error by 20 km after a 7-day forecast compared with the operational model. Furthermore, Lagrangian fine-tuning with the proposed advection loss further reduces the positioning error by 10 km, highlighting the benefits of incorporating Lagrangian constraints into the training process.

AI 总结：本研究探索LLM智能体为表格模型设计扩展HPO搜索空间，在45个数据集上平均提升0.6%性能，且无需额外调优成本，并提升TabArena基准Elo分数。

链接：https://arxiv.org/abs/2609.16309

机构：Yandex； HSE University（高等经济大学）

作者：Renat Sergazinov, Artem Chistyakov, Sergey Pankevich, Artem Babenko

英文摘要：Despite the rapid progress of LLM-based agents for planning, code generation, and debugging, their practical value for tabular machine learning remains underexplored. In this paper, we investigate a concrete use case: whether state-of-the-art agentic AI systems can design extended HPO search spaces for established tabular models that outperform the standard search spaces provided by the model authors. Specifically, we represent each tabular model as a modular pipeline covering preprocessing, embeddings, architecture, training, and inference. We then task the agent to propose candidate code implementations for each module and use a classical HPO algorithm to jointly optimize over these candidates and the model's default hyperparameters. Compared with the base HPO spaces, the expanded search spaces improve the performance of nearly every model family across a suite of 45 datasets, with average relative gains of 0.6%, rising to 2.0% on small-to-medium regression datasets. Notably, these gains come at no extra tuning cost: the enlarged spaces outperform the base under the same tuning and ensembling budgets. The gains transfer to the recent TabArena benchmark, where the agentic spaces improve the official Elo scores of four of the five model families and the two strongest agentic ensembles surpass the best AutoGluon ensemble of conventional models. Overall, our study suggests that LLM agents can provide practical value for tabular ML by expanding the design space.

AI 总结：本文提出基于多实例学习和多标签比例学习的弱监督框架，直接利用多边形级冰图标签预测海冰类型，避免标签近似，显著提升精度。

链接：https://arxiv.org/abs/2609.16347

机构：University of Colorado Denver（科罗拉多大学丹佛分校）； National Snow and Ice Data Center (NSIDC), CIRES, University of Colorado Boulder（国家冰雪数据中心（NSIDC），科罗拉多大学博尔德分校环境科学合作研究所）

作者：Samira Alkaee Taleghan, Younghyun Koo, Andrew P. Barrett, Farnoush Banaei-Kashani

英文摘要：Sea-ice type prediction is important for climate monitoring, maritime navigation, and decision-making in polar regions. The main source of label data for this task is the ice chart, produced manually by ice analysts who interpret satellite imagery to delineate ice zones into polygons. Although ice charts are valuable, their production is labor-intensive and expensive, motivating recent efforts to automate the process using deep learning. However, deep learning models require patch-level (or pixel-level) label data for training, while ice charts provide only polygon-level annotations. As a workaround, supervised approaches often create approximate patch-level labels from polygon-level ice chart labels by assigning each sample the dominant ice type of its parent polygon. This approach enables supervised training but creates an ill-posed learning problem with intrinsically approximate solution. In this paper, we redefine sea-ice type prediction as a weakly supervised multi-label proportion learning problem to be able to directly use the polygon-level ice chart labels and avoid unnecessary label approximation for improved prediction accuracy. To address this problem, we propose a two-module framework where first Multiple Instance Learning (MIL) is used for water--ice classification, and then a multi-label proportion learning (MLPL) is introduced for ice-type composition prediction. We further extend this framework with a multimodal model that integrates SAR imagery with AMSR2 brightness temperatures and ERA5 reanalysis data through modality-guided auxiliary regularization. Evaluated on the AI4Arctic dataset, the SAR-only model reduces MAE by 14.5\% and more than doubles mean ice-class F1 over the best supervised baseline. The multimodal model further reduces MAE by 21.5\% and raises mean F1 by 41.2\% over the SAR-only model, and by 52.7\% over the supervised multimodal baseline.

AI 总结：针对单次联邦贝叶斯模型，提出通过后验事件传输的部署一致认证框架，精确刻画FedAvg聚合下的安全概率下界，实验显示传输认证与直接全局认证存在显著差异。

链接： https://arxiv.org/abs/2609.16373

机构：University of Tehran（德黑兰大学）； Institute for Research in Fundamental Sciences (IPM)（基础科学研究所（IPM））

作者：Mahyar Mohammadi, Mohammad Hossein Badiei, Abolfazl Yaghmaei, Hamed Kebriaei

英文摘要：Probabilistic certification of Bayesian neural networks lower-bounds the posterior probability that a model satisfies a verifier-defined safety property. In one-shot federated Bayesian learning, however, the deployed model is obtained by aggregating parameters drawn from client-specific posterior distributions, so local certificates do not directly guarantee safety of the aggregated model. This paper develops a deployment-consistent certification framework by propagating local posterior events through the deployment aggregation rule, with an exact geometric characterization for Federated Averaging (FedAvg). Each client constructs disjoint hyper-rectangular regions in parameter space and computes their probability masses. The server forms Cartesian products of these regions, maps them through the deployment rule, and retains a product event only when its aggregation image is verified to satisfy the safety property. Under independent client posteriors, each product-event probability factorizes into local masses, and summing verified disjoint events yields a lower bound on safety probability of the deployed model. For FedAvg with nonnegative aggregation coefficients, the image of a Cartesian product of axis-aligned hyper-rectangles is exactly a weighted hyper-rectangle, introducing no set over-approximation. We distinguish the proposed transported-event certificate from direct certification under posterior distributions induced by FedAvg and Product-of-Gaussians aggregation. Experiments on MNIST and Fashion-MNIST under label-Dirichlet heterogeneity show that the transported FedAvg certificate ranges from 22.51% to 46.89%, while direct global certificates range from 72.05% to 91.39%. Results show that predictive accuracy and certifiable safety do not necessarily follow the same trend, and that global posterior constructions can exhibit distinct certification behavior across architectures.

AI 总结：针对类别不平衡与标签噪声的耦合问题，提出BARGE方法，结合有界先验调整得分与可靠性引导角度几何，无需噪声率，在多个基准上显著降低平均平衡错误。

链接：https://arxiv.org/abs/2609.16380

机构：Indian Institute of Technology Indore（印度理工学院印多尔分校）

作者：Mushir Akhtar, Akarsh J., M. Tanveer, Mohd. Arshad

英文摘要：Class-balanced learning and label noise create a coupled failure mode: frequency correction prevents majority classes from dominating the decision rule, but can amplify incorrectly labeled minority examples. We introduce BARGE (Bounded Adjustment with Reliability-Guided Embeddings), a single-stage objective combining a bounded, prior-adjusted density-power score with reliability-guided angular geometry. Its classification score is strictly proper in the adjusted probability space and recovers balanced Bayes ordering under clean supervision and the true class prior. Under label contamination, its finite range bounds classification-risk perturbation at a fixed predictor, while its logit gradient redescends when the model confidently contradicts the supplied label. The adjusted target probability also weights class-equal feature compactness, and a one-sided separation term discourages aligned class directions. BARGE requires neither a noise rate nor a transition matrix, uses one network, and leaves inference unchanged. We evaluate it on CIFAR-10, CIFAR-100, and Tiny ImageNet under long-tail and step imbalance, clean labels, and 20% and 40% random incorrect-label replacement. Across 12 clean settings, BARGE ranks second overall and attains the lowest error in four. Under corruption, it achieves the lowest mean balanced error in all six dataset-corruption settings, reducing the six-setting average from 72.32% for the strongest competitor to 70.00%. It also obtains the highest macro-F1 and macro-AUPRC in every corrupted-label setting. Ablations show that class-equal angular compactness improves on the bounded score alone. These results support bounded predictive influence and reliability-guided geometry as complementary mechanisms for imbalanced learning with uncertain labels.

AI 总结：本研究通过比较提示操控、SAE特征引导和探针方向引导三种方法，解读并引导LLM智能体的偏好与能力，发现SAE和探针方法在引导上更有效，为社会科学模拟提供了可解释且可引导的流程。

链接：https://arxiv.org/abs/2609.16436

机构：Data Innovation and AI Lab (DIAL), University of California, Berkeley（加州大学伯克利分校数据创新与人工智能实验室）； National Bureau of Economic Research (NBER)（美国国家经济研究局）

作者：Jiayue Gaveal Fan, Arul Murugan, Shreyas Krishnan, Abhishek Nagaraj

英文摘要：Simulations based on large language models (LLMs) have proven to be powerful for understanding human behavior, making them valuable additions to the social scientific toolkit. However, LLMs are ultimately black boxes based on deep neural networks which limits their value for social science. This is because of a lack of (i) interpretability: i.e. the ability to assign clear mechanisms driving observed behavior; and a lack of (ii) steerability: i.e. the ability to mute or amplify specific theoretically meaningful mechanisms of action to drive specific model behavior. Here, we demonstrate how the black box could be opened up to further enrich LLM-based simulations. Specifically, we compare three types of methods: (1) prompt-based manipulation, (2) SAE-derived feature steering, and (3) probe-based direction steering and examine their utility for LLM-based social scientific simulations. We do so by interpreting and steering two foundational components of human behaviors, namely preferences (risk attitudes, altruism) and capabilities (divergent creativity, product innovation), operationalized using four classic economic and creative tasks implemented as natural-language interactions. Overall, our results show that SAE- and probe-based techniques often outperform basic prompt-based methods for steering LLM agents, although this advantage depends on the specific prompting strategy involved. Together, SAEs and probes constitute an effective pipeline for social scientists seeking to interpret and steer agents in social simulations: SAEs decompose agents' internal representations into human-readable features, after which probes can reliably shift agents' behaviors in specified directions. We discuss implications of these methods for future work using LLM agents for social scientific simulations.

AI 总结：针对多模态在线策略蒸馏中师生共同幻觉问题，提出OPD-Aha方法，从教师视觉偏好重建目标，抑制错误延续并促使学生反思，显著提升细粒度感知与复杂推理性能。

链接：https://arxiv.org/abs/2609.16459

机构：Arizona State University（亚利桑那州立大学）； University of Virginia（弗吉尼亚大学）； Stevens Institute of Technology（史蒂文斯理工学院）

作者：Chenhao Qiu, Dawei Li, Yechao Zhang, Lei Gong, Zhen Tan

英文摘要：Privileged on-policy distillation improves multimodal reasoning by allowing a teacher to evaluate student trajectories using rich, training-only visual evidence. Both models score these trajectories while conditioning on the same student-generated prefix. When a student misinterprets an image early in a response, this accumulating erroneous rationale eventually pulls the teacher away from its visual evidence. The teacher and student converge on the same hallucination, causing standard cross-model supervision to collapse precisely where correction is most needed. We find that the teacher's visual corrective preference is not lost under this misleading agreement. Comparing the predictions of the identical teacher given the real image and a visual null reveals that the privileged evidence still pushes the model toward the correct interpretation. We introduce OPD-Aha, which reconstructs the distillation target directly from this isolated visual preference rather than relying on the fragile teacher-student discrepancy. This reconstructed target aggressively suppresses continuations that contradict the image. Trained with this objective, students learn to naturally interrupt their own flawed reasoning with reflection tokens such as wait and actually. After reflection, subsequent generation relies less on the accumulated erroneous text and more on the visual evidence. Correcting these trajectories mid-generation fundamentally alters the reasoning process, yielding broad and consistent improvements across diverse fine-grained perception and complex multimodal reasoning benchmarks. Our code and models are available at this https URL.

AI 总结：针对心电图波形界定中解码器设计被忽视的问题，提出R-U-Net（ResNet-18编码器+U-Net解码器），在SemiSegECG上显著优于基线，证明解码器设计比SSL方法更关键。

链接：https://arxiv.org/abs/2609.16489

机构：Carnegie Mellon University（卡内基梅隆大学）； Allegheny Health Network（阿勒格尼健康网络）； University of Colorado（科罗拉多大学）

作者：Joseph Scharpf, William Han, Chaojing Duan, Michael A. Rosenberg, Emerson Liu, Ding Zhao

英文摘要：Electrocardiogram (ECG) delineation identifies the boundaries of P waves, QRS complexes, and T waves, providing structural annotations that can guide AI models in learning to interpret ECGs. However, training accurate delineation models requires manual annotations that are scarce and time-consuming to obtain. Recent work addresses this limitation through semi-supervised learning (SSL), but the design of the architecture, particularly the decoder, has received less attention. To this end, we propose R-U-Net, an ECG delineation model that pairs a ResNet-18 encoder with a U-Net decoder. On SemiSegECG, R-U-Net outperforms the strongest evaluated ResNet-18 + fully convolutional network (FCN) head baseline in each of the 16 in-domain settings by 3.3-13.0 mIoU and achieves 82.6 mIoU in the cross-domain setting, an improvement of 8.1 mIoU. Controlled ablations show that decoder design contributes more to performance gains than the evaluated SSL methods, motivating further exploration of architectures for ECG delineation. All code is open-source at this http URL.

AI 总结：提出基于张量的HSMM维特比算法，将内循环映射到SIMD和并行架构，实现单核14倍、多核200倍、GPU570倍加速。

链接：https://arxiv.org/abs/2609.16500

机构：Sapienza University of Rome（罗马第一大学）

作者：Lorenzo Piarulli, Elia Belli, Daniele De Sensi

英文摘要：Hidden Semi-Markov Models (HSMMs) are fundamental probabilistic models widely adopted across diverse domains, from computational biology to finance and signal processing. The Viterbi algorithm decodes the most likely state sequence given an HSMM and can be applied iteratively for ab initio model learning. However, existing Viterbi implementations remain sequential, and GPU-accelerated solutions are entirely absent, making HSMM decoding impractical for large-scale workloads. We present a tensor-based formulation of the Viterbi algorithm for HSMMs, restructuring the inner loops into tensor operations that naturally map onto SIMD units and massively parallel architectures. Building on this formulation, we provide optimized implementations spanning single- and multi-core CPUs, and, for the first time, GPU. Experimental evaluation demonstrates speedups of up to 14x on a single core, over 200x with multi-core, and over 570x on GPU over the state-of-the-art sequential baseline, establishing a new performance baseline for large-scale HSMM decoding.

AI 总结：本文提出一种精确分布式样条合并方法，从碎片化观测中恢复物理参数，无需共享原始数据，误差低于0.12%，并在真实海温数据上验证。

链接：https://arxiv.org/abs/2609.16579

机构：University of California, Santa Barbara（加州大学圣塔芭芭拉分校）； Dyssonance AI

作者：Naveen Mysore

英文摘要：Scientific measurements are frequently distributed across locations, time periods, and institutions. Combining such fragments into a continuous, differentiable field enables recovering governing physical parameters from its derivatives. This paper makes two contributions toward that goal. First, the established additive structure of fixed-basis ridge-regression statistics is applied to tensor-product spline fields: each data holder computes a local Gram matrix and moment vector, and the merged solution is mathematically identical to centralized fitting, with no raw data shared and no iterative synchronization. This property is specific to the fixed-feature squared-error setting; the present derivation does not establish an analogous guarantee for general jointly trained multilayer networks. Second, a complete pipeline connects distributed observations to physical parameter inference through field reconstruction, derivative extraction, and linear regression. The diffusion coefficient is recovered to 0.11% error and wave speed to 0.12% error; in both cases, distributed merging introduces zero degradation relative to centralized fitting. Application to 41 years of NOAA sea-surface temperature data confirms the result on real spatiotemporal observations.

AI 总结：本文提出总敏感性核（TSKs）方法，通过加权ANOVA核族学习多变量结构，在RKHS中最小化范数以确定权重，从而提升黑箱函数逼近精度。

链接：https://arxiv.org/abs/2609.16606

机构：Emory University（埃默里大学）； Simon Fraser University（西蒙弗雷泽大学）

作者：John E. Darges, Laura Weidensager

英文摘要：Approximating the input-output behavior of a multivariable black-box function from limited data is challenging when blind to the importance of its inputs and their interactions. We introduce total sensitivity kernels (TSKs), a method based on families of weighted ANOVA kernels that learn and adapt to this multivariable structure. TSKs parameterize the weights on each multivariable component of the target function by factors for each input. We propose learning these factors directly from function evaluations by selecting the reproducing kernel Hilbert space (RKHS) in which the target function has minimum norm. Under suitable conditions, we show that this norm-minimization problem admits a unique solution, and we establish consistency of a finite-data formulation based on minimum-norm interpolation. The learned TSK factors characterize the participation of individual inputs across interactions and main effects, providing a kernel-dependent notion of input sensitivity related to total Sobol indices. Numerical experiments demonstrate that adapting the kernel to learned multivariable structure can substantially improve approximation accuracy over a standard product kernel.

AI 总结：本研究通过几何分析揭示循环Transformer中局部改进方向因步长过大导致有限步失败，并提出路径曲率分解与局部二次模型预测有效步长，实验表明固定四分之一步长在多数失败案例中带来正增益。

链接：https://arxiv.org/abs/2609.16665

机构：University of Technology Sydney（悉尼科技大学）； East China Normal University（华东师范大学）； Shanghai University of Electric Power（上海电力大学）； University of Agder（阿格德大学）； Squirrel AI Learning（松鼠AI）

作者：Zhihao Guo, Zonghan Wu, Haizhou Du, Huan Huo, Yilei Shao, Athanasios V. Vasilakos, Qingsong Wen

英文摘要：Looped Transformers offer a parameter-efficient route to test-time scaling by reusing shared layers for iterative latent reasoning. However, additional iterations can reduce support for a reference answer, leaving unclear whether an update's direction is locally unhelpful or its full displacement moves too far. We study this distinction by analysing reference utility, which measures this support, along the model's own update direction, varying the fraction of the proposed displacement supplied to the readout. This reveals finite-step failures in which a locally improving direction produces a harmful full update. A pathwise curvature decomposition characterises how initial progress is lost, while a local quadratic model predicts full-step gains and useful step scales. Bounds based on accumulated curvature variation characterise the approximation error of these predictions. Experiments across two model families reveal this separation on mathematical and commonsense tasks. A fixed quarter step produces positive gains in reference utility for 72.2--83.2% of selected failures across four settings. These findings identify a mismatch between update direction and step scale as a mechanism of lost progress, explaining how some harmful updates retain useful computation.

AI 总结：本研究首次在10毫秒关键时间窗内系统评估多种机器学习模型用于电网故障检测与故障线路定位，最优模型F1分数达0.991±0.018，处理时间0.342毫秒。

链接：https://arxiv.org/abs/2609.16744

机构：Friedrich-Alexander-Universität Erlangen-Nürnberg（埃尔朗根-纽伦堡大学）

作者：Julian Oelhaf, Georg Kordowich, Paula Andrea Pérez-Toro, Tomás Arias-Vergara, Andreas Maier, Johann Jäger, Siming Bayer

英文摘要：The integration of renewable energy sources into the electrical grid introduces complex challenges in fault detection and coordination of grid recovery mechanisms. Traditional relay protection systems, which operate based on static rules and predefined thresholds, are inadequate for addressing these challenges, particularly in detecting and isolating faults such as short circuits. Consequently, the conventional methodologies applied to electrical network protection frequently fail to achieve optimal performance in fault detection, especially in terms of adherence to safety standards and the selective limitation of damage. Recent research indicates that machine learning (ML)-based approaches can effectively tackle these issues; however, variations in grid configurations and analysis windows have impeded consistent comparative assessments. In this study, we assess the efficacy of various ML models in detecting electrical faults and pinpointing defective transmission lines within a 10 ms measurement interval - a critical time-frame for real-time operational viability, for the first time. The most effective model attained an F1 score of 0.991 +/- 0.018 and demonstrated a processing time of 0.342ms +/- 0.509ms.

AI 总结：本文通过实验证明，在自监督去噪中，训练配对分布而非损失函数选择（L1 vs L2）主导性能，并给出 Kodak24 与 SIDD 上的定量证据。

链接：https://arxiv.org/abs/2609.16788

作者：Dingyan Shang, Zhenyu Xu, Youting Wang, Bonan Shen, Bowen Liu

英文摘要： Noise2Noise (N2N) trains denoisers on pairs of independently corrupted observations, eliminating clean references. We stress-test two natural conjectures about why the L1 loss outperforms L2 here. First, the hypothesis that the L1 loss confers robustness via parameter sparsity confuses the loss with Lasso regularization: an explicit Lasso penalty produces the predicted sparsity yet fails to reproduce L1's cross-noise behavior, while L1- and L2-trained weight distributions are indistinguishable. Second, the population optima of the two losses coincide exactly for symmetric signal posteriors and nearly so for concentrated ones. Measured differences are therefore dominated by optimization dynamics (bounded-influence gradients), which we probe with gradient statistics and contaminated-target training. On Kodak24 with five synthetic noise families, the L1 loss holds a statistically significant edge over L2, below 1 dB PSNR, holding across three seeds on 13 of the 14 noise columns. On real camera noise the loss is not the decisive variable in distribution: on official SIDD validation blocks, synthetic-Gaussian-trained N2N models gain only 0.8 to 3.7 dB over the noisy input regardless of loss, while retraining on SIDD's own noisy pairs, never reading ground truth, gains 9.4 to 11.0 dB, far ahead of BM3D. All metrics are on raw network outputs, and the study makes no leaderboard claim. The training pair distribution, not the loss, carries the inductive bias. That design rule applies wherever clean references are unobtainable, from microscopy to industrial inspection sensors.

AI 总结：LCAP利用群体信息从少量输出探针推断芯片潜在校正，实现光子神经网络免优化的硬件适配，显著提升部署准确率。

链接：https://arxiv.org/abs/2609.16823

机构：City University of Hong Kong（香港城市大学）； Sichuan University（四川大学）； Nanyang Technological University（南洋理工大学）

作者：Tianyu Gao, Guantian Zheng

英文摘要：Photonic neural networks (PNNs) offer efficient analog inference, but parameters optimized under ideal device models can degrade after fabrication, creating a persistent simulation-to-hardware (sim-to-real) gap. When many identically designed chips are deployed, calibrating each device from scratch compounds this cost. We propose Latent Chip Adaptation from Probes (LCAP), a population-informed framework that decomposes hardware adaptation into a transferable population correction and probe-inferred latent personalization. LCAP first learns a shared correction from 80 historical chips, then extracts a low-dimensional correction space from device-specific refinements. At deployment, 32 fixed unlabeled output probes infer an unseen chip's latent correction coordinates, enabling feed-forward personalization without target-device optimization. On a three-layer 64-mode MZI simulator with phase variation, beam-splitter errors, quantization, and crosstalk, accuracy improves from 80.4147% under direct deployment to 92.6860% after shared calibration and 93.3617% with LCAP. LCAP improves 27/30 unseen chips and raises worst-device accuracy from 89.18% to 90.54%.

AI 总结：本研究比较自回归语言模型的局部与全局置信度，发现二者相关性弱且与正确性关联不同，分歧可诊断采样不稳定性，故置信度应视为明确定义的测量而非内在属性。

链接：https://arxiv.org/abs/2609.16933

机构：SiftyML, LTD（SiftyML有限公司）

作者：Julio C. Amador Diaz Lopez

英文摘要：Modern predictive systems expose multiple quantities that are commonly interpreted as measures of confidence. However, these quantities can summarize different aspects of the predictive process. This distinction matters when confidence is used to evaluate reliability or inform downstream oversight and control. We investigate whether different confidence readouts are empirically interchangeable in an autoregressive language model by comparing local confidence, defined from the probability of the greedy-selected answer token, with global confidence, defined from modal-answer frequency under repeated sampling. Across MMLU and ARC Challenge, the two signals are weakly correlated and differ substantially in their association with correctness: global confidence is moderately associated with correctness, whereas local confidence shows little association. We further test whether question-level disagreement between the signals is associated with sampling instability. On ARC, larger local--global confidence gaps are associated with higher answer entropy, more distinct sampled answers, and lower modal-answer concentration. The gap--entropy association persists when disagreement and instability are estimated from disjoint stochastic samples, indicating that it is not explained by shared finite-sample variation. The corresponding relationship is substantially weaker on MMLU, where only 4% of questions exhibit sampling instability. These results show that confidence readouts derived from the same predictive system are not empirically interchangeable and that their disagreement can provide a diagnostic of unstable sampling behavior. Confidence should therefore be treated as an explicitly defined measurement rather than as a single intrinsic scalar property of a model, particularly when it is used to inform downstream evaluation, oversight, or control.

AI 总结：本文提出结合随机动态模式分解与深度学习构建高保真数字孪生数据模型，以非侵入方式降复杂度再现流体动力学激波现象，兼顾精度与效率。

链接：https://arxiv.org/abs/2609.17101

机构：University Politehnica Timisoara（蒂米什瓦拉理工大学）

作者：Diana A. Bistrian

英文摘要：The purpose of this paper is the identification of high-fidelity digital twin data models from numerical code outputs by non-intrusive techniques (i.e., not requiring Galerkin projection of the governing equations onto the reduced modes basis). In this paper the author defines the concept of the digital twin data model (DTM) as a model of reduced complexity that has the main feature of mirroring the original process behavior. The significant advantage of a DTM is to reproduce the dynamics with high accuracy and reduced costs in CPU time and hardware for settings difficult to explore because of the complexity of the dynamics over time. This paper introduces a new framework for creating efficient digital twin data models by combining two state-of-the-art tools: randomized dynamic mode decomposition and deep learning artificial intelligence. It is shown that the outputs are consistent with the original source data with the advantage of reduced complexity. The DTMs are investigated in the numerical simulation of three shock wave phenomena with increasing complexity. The author performs a thorough assessment of the performance of the new digital twin data models in terms of numerical accuracy and computational efficiency.

AI 总结：本工作提出一种基于条件神经场集成的方法，通过傅里叶特征编码、相对平方误差目标和k折交叉验证，在有限数据下预测复杂飞机配置的气动壁面分布，以8.81分赢得ONERA CRM挑战赛，优于基线且参数少三个数量级。

链接：https://arxiv.org/abs/2609.17160

机构：Cenaero

作者：Lionel Salesses, Caroline Sainvitu, Tariq Benamara

英文摘要：Machine-learning surrogate models offer a promising alternative to high-fidelity Computational Fluid Dynamics (CFD) simulations for aerodynamic analysis and design. However, constructing accurate surrogates for realistic aircraft configurations remain challenging due to complex geometries, multiple flow regimes, and limited training data. This work presents the methodology that achieved first place in the ONERA CRM Wall Distribution Regression Challenge, which focuses on predicting pressure and skin-friction coefficient distributions over the NASA Common Research Model wing-body-pylon-nacelle configuration under different operating conditions. The proposed approach formulates the problem as a conditional neural field mapping spatial coordinates, surface normals, and operating conditions to aerodynamic wall quantities. Fourier feature encoding, a relative squared error objective aligned with the challenge metric, ensemble learning, and $k$-fold cross-validation are progressively introduced to improve prediction accuracy and exploit the limited training data. Beyond presenting the final methodology, the paper documents the successive model design choices that led to the winning solution through a comprehensive ablation study and discusses several alternative approaches that were investigated but ultimately discarded. On the hidden competition test set, the proposed methodology achieves an overall score of 8.81, outperforming the strongest organizer-provided baseline, which achieved a score of 8.64, while requiring approximately three orders of magnitude fewer trainable parameters. These results illustrate that carefully designed coordinate-based neural fields constitute an efficient and robust framework for aerodynamic surrogate modeling on complex geometries under limited-data conditions.

AI 总结：针对循环Transformer解码延迟高的问题，提出免训练的自推测解码框架LoopSpec，通过流水线重叠草稿生成与验证，并引入选择性第二提议，实现高达6.83倍推理加速。

链接：https://arxiv.org/abs/2609.17184

机构：Seoul National University（首尔大学）； KAIST（韩国科学技术院）

作者：SangLyul Cho, Langqing Cui, Sehoon Kim, Dongsu Han, Insu Han

英文摘要：Looped Transformers achieve strong performance with compact parameter sizes by repeatedly applying a shared stack of Transformer blocks across recurrent depths. However, they incur higher decoding latency than standard Transformer models of comparable parameter size because shared weights are accessed at every recurrent depth. To improve decoding efficiency, self-speculative decoding is particularly well suited to Looped Transformers, as their intermediate recurrent states can directly provide draft predictions without an auxiliary draft model. We therefore propose LoopSpec, a training-free self-speculative decoding framework tailored for Looped Transformers. LoopSpec extracts draft tokens from early recurrent states and operates in a pipelined manner, overlapping draft generation of future tokens with target verification of the current token. To improve draft accuracy without excessive compute overhead, we introduce a selective second proposal from deeper recurrent depth while ensuring lossless decoding under both greedy and sampling regimes. Furthermore, we derive the optimal proposal depths in closed form and show the prediction matches measurement. Across reasoning and coding benchmarks, LoopSpec achieves up to 6.83$\times$ inference speedup across diverse Looped Transformers.

AI 总结：本研究揭示医学AI模型因训练时见过患者历史数据，导致对其未来数据预测产生“记忆偏差”，影响诊断准确性，凸显隐私保护与模型部署间的矛盾。

链接：https://arxiv.org/abs/2609.17223

机构：Technical University of Munich (TUM)（慕尼黑工业大学）； TUM University Hospital（慕尼黑工业大学附属医院）； Munich Center for Machine Learning（慕尼黑机器学习中心）； Imperial College London（伦敦帝国理工学院）； Hasso Plattner Institute（哈索·普拉特纳研究院）

作者：Moritz A. Knolle, Martin J. Menten, Laurin Lux, Mélanie Roschewitz, Emma A.M. Stanley, Georgios Kaissis, Daniel Rueckert, Ben Glocker

英文摘要：Medical AI models hold immense potential to improve patient outcomes, but they are also known to unintentionally memorise individual records from their training datasets. While such memorisation has been linked to targeted privacy attacks, its consequences for clinical deployment, where patients may be assessed by a model that saw their historical data during training, remain poorly understood. Here we show that predictions on a patient's unseen future data can change significantly if a model observed that same patient's anonymised historical data during training, a phenomenon we term "memorisation bias". We demonstrate that this bias exists across diverse data modalities and model architectures, and over prolonged time spans: in some cases, memorisation bias persists on future records acquired decades after the historical records used for training. Moreover, in simulated prospective deployment, memorisation bias has asymmetric effects on the diagnostic accuracy of returning data contributors. When a patient returned with a de novo condition absent from their historical records in the training dataset, diagnostic sensitivity decreased significantly compared to an otherwise identical model not trained on their historical data. Conversely, when their health state was unchanged, both sensitivity and specificity were significantly inflated. Our findings reveal a previously uncharacterised risk in medical AI that arises when a model is deployed on patients who contributed to its training data. This exposes a shortcoming of current model development practice: the de-identification measures designed to protect patients' privacy make it difficult to identify returning contributors and exclude them from the AI-assisted interpretation of their own future data. Mitigating memorisation risks may thus require changes to current model training and deployment protocols.

AI 总结：本研究通过探测隐马尔可夫模型数据提示下的开源LLMs，发现其上下文学习中的信念状态可线性解码，干预该子空间可保持预测质量，表明ICL近似最优贝叶斯预测。

链接：https://arxiv.org/abs/2609.17376

作者：Daniel Balcells, Andrew Jun Lee, Chirag Rastogi, Paul M. Riechers, Adam Shai, Xavier Poncini

英文摘要： Large language models (LLMs) trained on next-token prediction exhibit remarkable in-context learning (ICL) abilities, yet the representations that support ICL remain poorly understood. We consider such representations in a controlled setting: prompting LLMs with data emitted from hidden Markov models (HMMs) and probing for the corresponding belief state -- the posterior distribution over the HMM's hidden states given the observed token history. Across six open-source LLMs prompted with data from 40 HMMs selected for non-trivial belief structure, we find that belief states are linearly decodable from residual stream activations, with peak probe $R^2$-values from 0.83-0.99 across HMM and LLM combinations, ranging from early to late layers. To establish functional relevance, we intervene directly on the probe-identified subspace via patching and steering, resulting in downstream prediction quality on the order of the untampered model, while controls degrade performance substantially. Together, these results provide representation-level evidence that ICL in open-source LLMs approximates optimal Bayesian prediction over a context-inferred generative model. More broadly, our findings extend prior results linking input-distribution structure to activation geometry: from toy networks trained explicitly on HMM data to production-scale LLMs.

AI 总结：本文提出将有限知识存储为精确周期潜在生成器的相位，通过离散傅里叶算子保证循环闭合，在图像和视频基准上实现高保真压缩，验证了以动态存储替代独立实例的可行性。

链接：https://arxiv.org/abs/2609.17417

作者：Siddharth Pal, Viktoria Rojkova

英文摘要：Finite knowledge is usually stored extensionally, one code or vector per item. We ask whether a finite collection can instead be stored intensionally, as the decoded orbit of one compact law that returns exactly to its start. For X objects, we encode item i as the i-th phase of a fixed rotation in a learned latent space and decode all phases with a shared network; the latent advances through a bank of rotations at integer harmonics of the cycle, a real discrete Fourier operator, so that R^X equals the identity and exact closure is guaranteed rather than learned. Images are a controlled carrier; looping video is the case where the phase order is the content's own temporal structure. Holding the decoder fixed and varying only the operator, a general learned operator diverges, a norm-preserving but non-periodic one degrades around the loop, and the exactly periodic operator is flat; on real images the gap widens. Capacity is then the decoder's budget: dense decoders carry a structural overhead per crisp image that no size reconciles with compression, while a small convolutional decoder on objects that share a manifold reaches crisp and compressed. A codebook control shows the generative law is free in reconstruction terms while multiplying the latent store many-fold. On seven benchmark clips, against a matched frame-index baseline, the cycle reaches equal or better fidelity at equal parameters while wrapping at machine precision, where the baseline leaves a visible seam; pinning the baseline's frequencies to loop harmonics closes its seam too, confirming that exact periodicity is the operative constraint. Finite cyclic knowledge can be stored as dynamics rather than independent instances, with exact recurrence supplied by algebra and content by a shared decoder.

AI 总结：针对大规模动态动作空间中的一对多移动充电问题，提出LP-BTS学习引导规划架构，结合图提议策略、价值评论器和预算树搜索，在250传感器场景中实现最高存活率0.4545。

链接：https://arxiv.org/abs/2609.17429

机构：National Chung Hsing University（国立中兴大学）

作者：Liang-Ching Tao, Pi-Chung Wang

英文摘要：Many learned sequential decision systems map the current state directly to an action. That shortcut becomes brittle when candidate actions are numerous, geometrically structured, and rebuilt with the state. One-to-many mobile charging makes this setting concrete: with N=250 sensors, the initial state induces about 1,125 candidate charging-stop actions; each chosen stop simultaneously serves its in-range sensors, and the action universe changes as sensors die. LP-BTS is a learning-guided planning architecture: a graph proposal policy concentrates a small candidate support, a learned value critic evaluates leaves, and edge-budgeted PUCT compares short simulated futures before committing an action. Because the policy scores this set without a fixed output head, a single frozen checkpoint covers every evaluated setting, spanning action universes from 736 to 2,813 stops. Matched ablations reveal complementary effects: uniform sampling costs 8.8 survival percentage points, while, with targeted support fixed, PUCT jointly retains 1.4 points (about 3.5 of 250 sensors) and direct policy selection travels 23% farther. On a prospectively specified, sealed 30-scenario confirmatory bank evaluated once, LP-BTS attains the highest observed survival (0.4545) and alive-AUC (0.8031). Its estimated survival advantage over the strongest domain-engineered comparator is +0.0066 (95% CI [-0.0037, +0.0184]), an unresolved difference, while it exceeds a deadline heuristic and two source-derived direct-policy reconstructions on every paired scenario. Both learned rows are trained, source-derived reconstructions of variants reported by Gong et al. In this setting, the results provide controlled evidence about learning-guided planning in a large, dynamic action space.

AI 总结：提出耦合校准与学习（CCL）算法，通过令牌级分支耦合教师校准与学生更新，仅用源反馈，在理论上证明可消除教师偏差并收敛至最优学生，无需目标域奖励。

链接：https://arxiv.org/abs/2609.17474

机构：MIT（麻省理工学院）； UIUC（伊利诺伊大学厄巴纳-香槟分校）

作者：Haichen Hu, Yuheng Zhang, David Simchi-Levi

英文摘要： Large language model (LLM) distillation aims to transfer the capabilities of a powerful teacher to a smaller student. Direct imitation, however, can also transfer the teacher's systematic bias and errors. This challenge is particularly pronounced under covariate shift, when the teacher's reliability on target questions is uncertain and target-domain reward feedback is unavailable. We propose Coupled Calibration and Learning (CCL), an LLM distillation algorithm that couples teacher calibration with student updates through token-level branching, using reward feedback only on source questions. Each iteration calibrates the teacher using source feedback and then uses the calibrated teacher to train the student on target questions. The updated student, in turn, informs subsequent calibration. In an autoregressive policy framework, we prove that the output student's expected average Kullback-Leibler divergence to the oracle student converges to zero at a polynomial rate in the number of iterations. The oracle maximizes the true reference-regularized target reward within the student class, which need not represent the unrestricted optimal policy. Our analysis quantifies the progress of projected student gradient updates while controlling the error in teacher calibration. We further establish a separation from regularized direct matching: its error relative to the oracle student can remain bounded away from zero even when the teacher achieves higher regularized target reward than every student policy. These results demonstrate that LLM distillation can overcome persistent teacher bias and recover the optimal student through coupled calibration and learning, without target-domain reward feedback.

AI 总结：针对VLN中标准共形预测无法在依赖且变长片段上提供覆盖保证的问题，提出ENCP，通过残差置信度重缩放非一致性分数并每片段校准一个最大分数，在R2R和REVERIE上达到所有步级覆盖目标，提供模型无关的不确定性估计。

链接：https://arxiv.org/abs/2609.17499

机构：Monash University, Indonesia（蒙纳士大学印度尼西亚校区）； SEACrowd

作者：Vicky Feliren, A. Taufiq Asyhari, Muhamad Risqi U. Saputra

英文摘要：Uncertainty estimation for Vision-Language-Navigation (VLN) models is a critical task since it can help identify ambiguous and unreliable predictions, enabling agents to make safer navigation decisions. As one of the most advanced uncertainty estimation frameworks, conformal prediction (CP) offers a promising approach for uncertainty estimation in VLN. However, given that VLN agent requires a sequence of steps, standard calibration in conformal prediction fails to provide coverage guarantee it promises over a dependent, variable-length VLN episode. To this end, we propose Episode-Normalized Conformal Prediction (ENCP), which rescales a nonconformity score by the policy's residual confidence and calibrates one maximum score per episode. Under exchangeable calibration and test episodes, this construction covers the ground truth at every step with probability at least $1 - \alpha$, while allowing dependence among steps within an episode. Across four VLN policies and three nonconformity scores on R2R and REVERIE dataset, ENCP meets all reported empirical step-coverage targets on the seen-to-unseen evaluation. These results demonstrate that ENCP can provide model-agnostic uncertainty estimates, which might be useful for determining when a VLN agent should defer to a more capable predictor, including human assistance.