|
创作新主题 |
社区所有版块导航
DATA
| docker Elasticsearch |
WEB开发
| linux MongoDB Redis DATABASE NGINX 其他Web框架 web工具 zookeeper tornado NoSql Bootstrap js peewee Git bottle IE MQ Jquery |
机器学习
| 机器学习算法 |
产品
| 短视频 |
印度
| 印度 |
一周十大热门主题
点击阅读原文访问arxivdaily.com，涵盖CS|物理|数学|经济|统计|金融|生物|电气领域，更有搜索、收藏等功能！
cs.LG 方向，今日共计372篇
大模型相关(48篇)
【1】Beyond Pairs: Your Language Model is Secretly Optimizing a Preference Graph
标题：超越配对：您的语言模型正在秘密优化偏好图
链接：https://arxiv.org/abs/2605.08037
摘要：Direct Preference Optimization (DPO) aligns language models using pairwise preference comparisons, offering a simple and effective alternative to Reinforcement Learning (RL) from human feedback. However, in many practical settings, training data consists of multiple rollouts per prompt, inducing rich preference structure that pairwise DPO fails to exploit. Collapsing such data into independent pairs discards transitivity, introduces redundant or conflicting supervision, and can lead to unstable optimization. We propose Graph Direct Preference Optimization (GraphDPO), a principled generalization of DPO that operates over directed acyclic preference graphs induced by rollout rankings. GraphDPO encodes dominance relations as edges and optimizes a graph-structured Plackett--Luce-inspired objective that aggregates supervision over graph neighborhoods, enforcing transitivity while recovering standard DPO as a special case. To handle discrete or sparse signals, we introduce an equivalence-class construction where responses with identical preferences form graph layers, and intra-layer edges contribute zero loss, preventing spurious gradients. Despite leveraging full graph structure, GraphDPO maintains linear per-prompt complexity via efficient log-sum-exp aggregation. We further incorporate optional ground-truth anchoring by inserting verified solutions as dominant nodes and applying an annealed schedule that stabilizes early training while gradually relaxing oracle supervision. Experiments on reasoning and program synthesis tasks demonstrate superior performance, suggesting that graph-structured preference modeling is a scalable and robust alternative to pairwise and listwise alignment objectives.
【2】STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation
标题：STARFlow2：为统一多模式生成搭建桥梁语言模型和规范化流程
链接：https://arxiv.org/abs/2605.08029
备注：19 pages, 9 figures
摘要：Deep generative models have advanced rapidly across text and vision, motivating unified multimodal systems that can understand, reason over, and generate interleaved text-image sequences. Most existing approaches combine autoregressive language modeling with diffusion-based image generators, inheriting a structural mismatch between causal text generation and iterative visual denoising. We observe that autoregressive normalizing flows are autoregressive Transformers--sharing the same causal mask, KV-cache mechanism, and left-to-right structure as LLMs--making them the most natural paradigm for true unified multimodal generation. We present STARFlow2, built on the Pretzel architecture that vertically interleaves a pretrained VLM stream with a TarFlow stream via residual skip connections, both operating under the same causal mask. Combined with a deep-shallow flow design and a unified FAE latent space, STARFlow2 enables cache-friendly interleaved generation where both text and visual outputs directly enter the KV-cache without re-encoding. Experiments demonstrate strong performance across image generation and multimodal understanding benchmarks, validating autoregressive flows as a viable foundation for unified multimodal modeling.
【3】Tool Calling is Linearly Readable and Steerable in Language Models
标题：工具调用在语言模型中是线性可读和可控制的
链接：https://arxiv.org/abs/2605.07990
备注：29 pages, 6 figures, 7 tables. Manuscript under review
摘要：When a tool-calling agent picks the wrong tool, the failure is invisible until execution: the email gets sent, the meeting gets missed. Probing 12 instruction-tuned models across Gemma 3, Qwen 3, Qwen 2.5, and Llama 3.1 (270M to 27B), we find the identity of the chosen tool is linearly readable and steerable inside the model. Adding the mean-difference between two tools' average internal activations switches which tool the model selects at 77-100% accuracy on name-only single-turn prompts (93-100% at 4B+), and the JSON arguments that follow autoregressively match the new tool's schema, so flipping the name is enough. The same per-tool means also flag likely errors before they happen: on Gemma 3 12B and 27B, queries where the gap between the top-1 and top-2 tool is smallest produce 14-21x more wrong calls than queries with the largest gap. The causal effect concentrates along one direction, the row of the output layer that produces the target tool's first token: a unit vector along it at matched magnitude already reaches 93-100%, while what is left over leaves the choice almost untouched. Activation patching localises this to a small set of mid- and late-layer attention heads, and a within-topic probe across 14 same-domain $τ$-bench airline tools reaches top-1 61-89% across five 4B-14B models, ruling out the reading that we are just moving the model along a topic axis. Even base models encode the right tool before they can emit it: cosine readout from the internal state recovers 69-82% on BFCL while base generation reaches only 2-10%, suggesting pretraining forms the representation and instruction tuning later wires it to the output. We measure tool identity selection and JSON schema correctness in single-turn fixed-menu settings; multi-turn agentic transfer is more fragile and is discussed in Limitations.
【4】Where's the Plan? Locating Latent Planning in Language Models with Lightweight Mechanistic Interventions
标题：计划在哪里？通过轻量级机械干预在语言模型中定位潜在规划
链接：https://arxiv.org/abs/2605.07984
备注：13 pages, 20 figures, 3 tables
摘要：We study planning site formation in language models -- where internal representations of structurally-constrained future tokens form during the forward pass, and whether they causally drive generation. Using rhyming-couplet completion as a clean test of forward-looking constraint, we apply two lightweight methods (linear probing and activation patching) across Qwen3, Gemma-3, and Llama-3 at more than ten scales. Probing shows that future-rhyme information is linearly decodable at the line boundary, with signal that strengthens with scale in all three families. Activation patching reveals that only Gemma-3-27B causally relies on this encoding, exhibiting a handoff in which the causal driver migrates from the rhyme word to the line boundary around layer 30. Every other model we test conditions on the rhyme word throughout generation, with near-zero causal effect at the line boundary despite strong probe signal. We localize the Gemma-3-27B handoff to five attention heads through two-stage path patching that recover ~90% of the rhyme-routing capacity at the newline.
【5】Self-Play Enhancement via Advantage-Weighted Refinement in Online Federated LLM Fine-Tuning with Real-Time Feedback
标题：通过实时反馈在线联合LLM微调中的加权细化来增强自玩
链接：https://arxiv.org/abs/2605.07977
备注：27 pages
摘要：Recent works have advanced feedback-based learning systems, whereby a foundation model is able to intake incoming feedback (e.g., a user) to self-improve, creating a self-loop system of training. However, existing works are limited in needing to consider an offline setup to allow for such feedback-based methods, and are further limited in the need of requiring privileged ground-truth contexts for training. Moreover, there is limited consideration of federated learning (FL), which is particularly well-suited for incorporating external feedback across large networks of end users, for example, but requires methods to be efficient for training on resource-constrained edge devices. Therefore, we introduce SPEAR (Self-Play Enhancement via Advantage-Weighted Refinement), an efficient online learning algorithm for federated LLM fine-tuning. SPEAR utilizes a feedback-guided self-play loop to construct naturally contrastive pairs per prompt which are utilized to be trained on (i) standard maximum likelihood on correct completions and (ii) confidence-weighted unlikelihood on tail tokens of incorrect completions. Without the need of expensive group generations and ground-truth contexts for training (i.e., only partial, non-answer feedback), in contrast with existing works, SPEAR can be trained both online and in a resource-efficient manner. We validate SPEAR across various benchmark datasets, demonstrating its superior performance in comparison to state-of-the-art baselines. The implementation code is publicly available at https://github.com/lee3296/SPEAR.
【6】Graph Representation Learning Augmented Model Manipulation on Federated Fine-Tuning of LLMs
标题：LLM联邦微调中的图表示学习增强模型操作
链接：https://arxiv.org/abs/2605.07961
摘要：Federated fine-tuning (FFT) has emerged as a privacy-preserving paradigm for collaboratively adapting large language models (LLMs). Built upon federated learning, FFT enables distributed agents to jointly refine a shared pretrained LLM by aggregating local LLM updates without sharing local raw data. However, FFT-based LLMs remain vulnerable to model manipulation threats, in which adversarial participants upload manipulated LLM updates that corrupt the aggregation process and degrade the performance of the global LLM. In this paper, we propose an Augmented Model maniPulation (AugMP) strategy against FFT-based LLMs. Specifically, we design a novel graph representation learning framework that captures feature correlations among benign LLM updates to guide the generation of malicious updates. To enhance manipulation effectiveness and stealthiness, we develop an iterative manipulation algorithm based on an augmented Lagrangian dual formulation. Through this formulation, malicious updates are optimized to embed adversarial objectives while preserving benign-like parameter characteristics. Experimental results across multiple LLM backbones demonstrate that the AugMP strategy achieves the strongest manipulation performance among all competing baselines, reducing the global LLM accuracy by up to 26% and degrading the average accuracy of local LLM agents by up to 22%. Meanwhile, AugMP maintains high statistical and geometric consistency with benign updates, enabling it to evade conventional distance- and similarity-based defense methods.
【7】MatryoshkaLoRA: Learning Accurate Hierarchical Low-Rank Representations for LLM Fine-Tuning
标题：MatryoshkaLoRA：学习准确的分层低等级表示以进行LLM微调
链接：https://arxiv.org/abs/2605.07850
摘要：With the rise in scale for deep learning models to billions of parameters, the computational cost of fine-tuning remains a significant barrier to deployment. While Low-Rank Adaptation (LoRA) has become the standard for parameter-efficient fine-tuning, the need to set a predefined, static rank $r$ requires exhaustive grid searches to balance efficiency and performance. Existing rank-adaptive solutions such as DyLoRA mitigate this by sampling ranks during the training from a predefined distribution. However, they often yield sub-optimal results at higher ranks due to lack of consistent gradient signals across the full hierarchy of ranks, thus making these methods data-inefficient. In this paper, we propose MatryoshkaLoRA, a general, Matryoshka-inspired training framework for LoRA that learns accurate hierarchical low-rank representations by inserting a fixed, carefully crafted diagonal matrix $P$ between the existing LoRA adapters to scale their sub-ranks accordingly. By introducing this simple modification, our general framework recovers LoRA and DyLoRA only by changing $P$ and ensures all sub-ranks embed the available gradient information efficiently. Our MatryoshkaLoRA supports dynamic rank selection with minimal degradation in accuracy. We further propose Area Under the Rank Accuracy Curve (AURAC), a metric that consistently evaluates the performance of hierarchical low-rank adapters. Our results demonstrate that MatryoshkaLoRA learns more accurate hierarchical low-rank representations than prior rank-adaptive approaches and achieves superior accuracy-performance trade-offs across ranks on the evaluated datasets. Our code is available at https://github.com/IST-DASLab/MatryoshkaLoRA.
【8】RelAgent: LLM Agents as Data Scientists for Relational Learning
标题：RelAgent：LLM代理作为关系学习的数据科学家
链接：https://arxiv.org/abs/2605.07840
摘要：Relational learning is a challenging problem that has motivated a wide range of approaches, including graph-based models (e.g., graph neural networks, graph transformers), tabular methods (e.g., tabular foundation models), and sequence-based approaches (e.g., large language models), each with its own advantages and limitations. We propose RelAgent, an LLM-based autonomous data scientist for relational learning, which operates in two phases. In the search phase, an LLM agent uses database, validation, and evaluation workspace tools to construct SQL feature programs and select a predictive model. In the inference phase, the resulting program is executed without further LLM calls. The final predictor consists of SQL queries and a classical model, enabling fast, deterministic, and intrinsically interpretable predictions: features are human-readable queries, and predictions depend only on the resulting query-defined feature map, enabling scalable deployment using standard database systems.
【9】Beyond Confidence: Rethinking Self-Assessments for Performance Prediction in LLMs
标题：超越信心：重新思考LLM绩效预测的自我评估
链接：https://arxiv.org/abs/2605.07806
摘要：Large Language Models (LLMs) are increasingly used in settings where reliable self-assessment is critical. Assessing model reliability has evolved from using probabilistic correctness estimates to, more recently, eliciting verbalized confidence. Confidence, however, has been shown to be an inconsistent and overoptimistic predictor of model correctness. Drawing on cognitive appraisal theory, a framework from human psychology that decomposes self-evaluation into multiple components, we propose a multidimensional perspective on model self-assessment. We elicit six appraisal-based dimensions of self-assessment, alongside confidence, and evaluate their utility for predicting model failure across 12 LLMs and 38 tasks spanning eight domains. We find that competence-related appraisal dimensions, particularly effort and ability, consistently match or outperform confidence across most settings. Effort additionally yields less overoptimistic estimates that remain stable across model sizes. In contrast, affective dimensions provide marginally predictive signals. Furthermore, the most informative dimension varies systematically with task characteristics: effort is most predictive for reasoning-intensive tasks, while ability and confidence dominate on retrieval-oriented tasks. Broadly, our findings indicate that structured multidimensional self-assessment is a promising approach to improving the reliability and safety of language model deployment across diverse real-world settings.
【10】Tracing Uncertainty in Language Model "Reasoning"
链接：https://arxiv.org/abs/2605.07776
摘要：Language model (LM) "reasoning", commonly described as Chain-of-Thought or test-time scaling, often improves benchmark performance, but the dynamics underlying this process remain poorly understood. We study these dynamics through the lens of uncertainty quantification by treating the "reasoning" traces, the intermediate token sequences generated by LMs, as evolving model states. We summarize each trace by an uncertainty trace profile: a small set of features describing the shape of the uncertainty signal over its trace, such as its slope and linearity. We find that across five LMs evaluated on GSM8K and ProntoQA, these profiles predict whether a trace yields a correct final answer with AUROC up to 0.807, improving markedly on recent related work. We reach AUROC 0.801 using only the first few hundred tokens of full traces, suggesting that errors can be detected early in the generation. A detailed comparison of correct and incorrect traces further reveals qualitatively distinct uncertainty profiles, with correct traces showing a steeper and less linear decline in uncertainty. Together, the results suggest that our method, grounded in decision-making under uncertainty, provides a principled lens for studying the generative process underlying LM "reasoning".
【11】POETS: Uncertainty-Aware LLM Optimization via Compute-Efficient Policy Ensembles
标题：POCTS：通过计算机高效的政策集成实现具有不确定性的LLM优化
链接：https://arxiv.org/abs/2605.07775
备注：preprint
摘要：Balancing exploration and exploitation is a core challenge in sequential decision-making and black-box optimization. We introduce POETS ($\textbf{Po}$licy $\textbf{E}$nsembles for $\textbf{T}$hompson $\textbf{S}$ampling), a novel framework that bridges uncertainty quantification and policy optimization. Our approach is grounded in the insight that policies trained with Kullback-Leibler (KL) regularization implicitly encode an underlying reward function. Building on this, POETS bypasses the complex, nested process of training an uncertainty-aware reward model and separately fitting a policy to this model. Instead, we directly train a policy ensemble to capture epistemic uncertainty by matching implicitly encoded reward functions to online, bootstrapped data. To overcome the prohibitive compute and memory constraints of ensembling Large Language Models (LLMs), POETS utilizes an efficient architecture: the ensemble shares a pre-trained backbone while maintaining diversity through independent Low-Rank Adaptation (LoRA) branches. Theoretically, we prove that POETS implicitly conducts KL-regularized Thompson sampling and thus inherits strong cumulative regret bounds of ${\mathcal O}(\sqrt{T γ_T})$. Empirically, we demonstrate that POETS achieves state-of-the-art sample efficiency across diverse scientific discovery domains, including protein search and quantum circuit design. Furthermore, it improves the optimization trajectories of reinforcement learning, proving particularly robust in off-policy settings with experience replay or in small dataset regimes.
【12】Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language Models
标题：内存高效的循环Transformer：在循环语言模型中将计算与内存脱钩
链接：https://arxiv.org/abs/2605.07721
备注：22 pages, 5 figures, 11 tables
摘要：Recurrent LLM architectures have emerged as a promising approach for improving reasoning, as they enable multi-step computation in the embedding space without generating intermediate tokens. Models such as Ouro perform reasoning by iteratively updating internal representations while retaining a standard Key-Value (KV) cache across iterations, causing memory consumption to grow linearly with reasoning depth. Consequently, increasing the number of reasoning iterations can lead to prohibitive memory usage, limiting the practical scalability of such architectures. In this work, we propose Memory-Efficient Looped Transformer (MELT), a novel architecture that decouples reasoning depth from memory consumption. Instead of using a standard KV cache per layer and loop, MELT maintains a single KV cache per layer that is shared across reasoning loops. This cache is updated over time via a learnable gating mechanism. To enable stable and efficient training under this architecture, we propose to train MELT using chunk-wise training in a two phase procedure: interpolated transition, followed by attention-aligned distillation, both from the LoopLM starting model to MELT. Empirically, we show that MELT models fine-tuned from pretrained Ouro parameters outperform standard LLMs of comparable size, while maintaining a memory footprint comparable to those models and dramatically smaller than Ouro's. Overall, MELT achieves constant-memory iterative reasoning without sacrificing LoopLM performance, using only a lightweight post-training procedure.
【13】Post-training makes large language models less human-like
标题：后训练使大型语言模型不太像人
链接：https://arxiv.org/abs/2605.07632
摘要：Large language models (LLMs) are increasingly used as surrogates for human participants, but it remains unclear which models best capture human behavior and why. To address this, we introduce Psych-201, a novel dataset that enables us to measure behavioral alignment at scale. We find that post-training -- the stage that turns base models into useful assistants -- consistently reduces alignment with human behavior across model families, sizes, and objectives. Moreover, this misalignment widens in newer model generations even as base models continue to improve. Finally, we find that persona-induction -- a popular technique for eliciting human-like behavior by conditioning models on participant-specific information -- does not improve predictions at the level of individuals. Taken together, our results suggest that the very processes that are currently employed to turn LLMs into useful assistants also make them less accurate models of human behavior.
【14】Mathematical Reasoning via Intervention-Based Time-Series Causal Discovery Using LLMs as Concept Mastery Simulators
标题：使用LLM作为概念掌握模拟器，通过基于干预的时间序列因果发现进行数学推理
链接：https://arxiv.org/abs/2605.07600
备注：17 pages, 0 figures
摘要：Recent methods for improving LLM mathematical reasoning, whether through MCTS-based test-time search or causal graph-guided knowledge injection, cannot identify which concepts causally contribute to a correct answer, as the observed association may be spurious, driven by confounders such as problem difficulty. We propose CIKA (Causal Intervention for Knowledge Activation), a framework that uses the LLM itself as an interventional simulator: a prompt sets the concept state to ``mastered'' and the correctness change estimates the causal effect. We formalize this quantity as an Interventional Capability Probe (ICP), which diagnoses whether the LLM can use a given concept -- distinct from merely possessing knowledge. Because the intervention exogenously sets the concept state independently of problem difficulty, ICP separates confounding that observational methods cannot. On 67 screened problems, the ICP of the top-ranked concept (+0.219) is significantly larger than that of the negative control (+0.039; paired $t$-test, $p < 10^{-6}$, Cohen's $d = 0.86$), confirming that the probe discriminates causally relevant concepts from irrelevant ones. Analysis of 601 Omni-MATH problems further shows that solved problems have 6.1$\times$ higher ATE than unsolved ones (0.338 vs. 0.055), confirming that ICP is predictive of problem-solving success. With a 7B-parameter LLM whose weights are entirely frozen, CIKA achieves 69.7\% on the contamination-free Omni-MATH-Rule benchmark and 64.0\% overall, compared to 60.5\% for o1-mini, and 97.2\% on GSM8K, 46--50\% on AIME 2024--2026, and 46.2\% on MathArena. The Causal Knowledge Activation component contributes 33.8\% of correct answers on problems where the base model alone fails, demonstrating that the LLM already possessed but had not activated the requisite knowledge.
【15】Your Language Model is Its Own Critic: Reinforcement Learning with Value Estimation from Actor's Internal States
标题：你的语言模型是它自己的批评者：强化学习和来自行动者内部状态的价值估计
链接：https://arxiv.org/abs/2605.07579
摘要 ：Reinforcement learning with verifiable rewards (RLVR) for Large Reasoning Models hinges on baseline estimation for variance reduction, but existing approaches pay a heavy price: PPO requires a policy-model scale critic, while GRPO needs multiple rollouts per prompt to keep its empirical group mean stable. We introduce Policy Optimization with Internal State Value Estimation), which obtains a baseline at negligible cost by using the policy model's internal signals already computed during the policy forward pass. A lightweight probe predicts the expected verifiable reward from the hidden states of the prompt and generated trajectory, as well as token-entropy statistics, and is trained online alongside the policy. To preserve gradient unbiasedness despite using trajectory-conditioned features, we introduce a cross-rollout construction that predicts each rollout's value from an independent rollout's internal states. Because POISE estimates prompt value using only a single rollout, it enables higher prompt diversity for a fixed compute budget during training. This reduces gradient variance for more stable learning and also eliminates the compute overhead of sampling costs for detecting zero-advantage prompts. On Qwen3-4B and DeepSeek-R1-Distill-Qwen-1.5B across math reasoning benchmarks, POISE matches DAPO while requiring less compute. Moreover, its value estimator shows similar performance to a separate LLM-scale value model and generalizes to various verifiable tasks. By leveraging the model's own internal representations, POISE enables more stable and efficient policy optimization.
【16】ProteinJEPA: Latent prediction complements protein language models
标题：ProteinJEPA：潜在预测补充了蛋白质语言模型
链接：https://arxiv.org/abs/2605.07554
摘要：Protein language models are trained primarily with masked language modeling (MLM), which predicts amino-acid identities at masked positions. We ask whether latent-space prediction can complement these token-level objectives under matched wall-clock budget. Across pretrained and random-init protein sequence encoders at 35--150M parameters, we find that the best protein-JEPA design is not all-position latent prediction but a variant: predicting latent targets only at masked positions, and retaining the MLM cross-entropy. We call this recipe masked-position MLM+JEPA. On a 16-task downstream suite (15 frozen linear probes plus SCOPe-40 zero-shot fold retrieval), under matched wall-clock budgets, this recipe wins more tasks than it loses against MLM-only continuation: 10 wins / 3 losses / 3 ties (hereafter W/L/T) on pretrained ESM2-35M, 11/2/3 on ESM2-150M while results in pretraining from scratch are mixed (6/8/2). Gains are seen for multiple models on 11 of 16 tasks, including stability, \b{eta}β\b{eta}-lactamase fitness, variant effect, intrinsic disorder, remote homology, enzyme classification, and SCOPe-40 fold retrieval. Tasks with more losses than wins are Fluorescence (TAPE) and Peptide-HLA Binding. All-position MLM+JEPA matches MLM-only overall but does not reproduce the masked-position gains. JEPA-only (no MLM) collapses in nearly every experiment. We conclude that JEPA, when combined with MLM, is competitive and can outperform pure MLM in pretraining and continued training, even under matched wall-clock budgets.
【17】GameGen-Verifier: Parallel Keypoint-Based Verification for LLM-Generated Games via Runtime State Injection
标题：GameGen-Verification：通过状态注入对LLM生成的游戏进行基于关键点的并行验证
链接：https://arxiv.org/abs/2605.07442
摘要：LLM-based game generation promises to turn natural-language specifications into executable games, but progress is limited by the lack of reliable automated verification. Unlike conventional code generation, game correctness is defined over long-horizon interaction: a game may appear correct while violating core mechanics such as state updates, interaction rules, and phase transitions. Existing Agent-as-a-Verifier approaches collapse verification into open-ended gameplay, making verdicts reachability-bound, time-consuming, coverage-limited, and sensitive to the agent's gameplay ability. We present GameGen-Verifier, an automated verification paradigm for LLM-generated games that decomposes a specification into verifiable keypoints and grounds them into independent verification units. Each unit patches the game runtime into a concrete target state, executes a bounded interaction, and judges the outcome against the keypoint assertion. We implement GGV-Harness, a scalable agentic harness providing concurrency management, runtime isolation, and fault recovery. On VeriGame, our dataset of 100 games across seven genres, GameGen-Verifier achieves up to 92.2% accuracy against human judgments versus 58.8% for the coverage-enforced Agent-as-a-Verifier baseline, while reducing wall-clock time by up to 16.6x.
【18】Unsolvability Ceiling in Multi-LLM Routing: An Empirical Study of Evaluation Artifacts
标题：多LLM路由中的不可解性上限：评估伪影的实证研究
链接：https://arxiv.org/abs/2605.07395
备注：12 pages, 14 tables
摘要：Efficient routing across multiple LLMs enables cost-quality tradeoffs by directing queries to the cheapest capable model. Prior work attributes routing headroom to an "unsolvability ceiling", queries no model in the pool can solve. We present a large-scale study of multi-tier LLM routing with 206,000 query-model pairs across six benchmarks (MMLU, MedQA, HumanEval, MBPP, Alpaca, ShareGPT) using the Gemma 4 and Llama 3.1 families. Evaluating with both LLM-as-a-judge and exact-match metrics, we show that a substantial portion of reported unsolvability stems from evaluation artifacts: (i) systematic judge biases favoring verbosity over correctness, (ii) truncation under fixed generation budgets, and (iii) output format mismatches. Through dual-judge validation and exact-match grounding, we reduce measured unsolvability across tasks. We introduce a decomposition framework attributing failures to these artifacts, revealing consistent patterns across domains and model families. These artifacts also distort router training signals: standard routers collapse to majority-class prediction (~79% smallest-tier optimal), confirmed via random-feature and shuffled-label controls, incurring a 13-17 percentage point opportunity cost. We provide actionable recommendations including dual-judge validation, exact-match anchoring, and cost-sensitive objectives. Our findings suggest existing routing headroom estimates are substantially inflated, underscoring the need for reliable evaluation protocols in multi-LLM systems.
【19】MISA: Mixture of Indexer Sparse Attention for Long-Context LLM Inference
标题：MISA：索引员对长上下文LLM推理的分散注意力的混合
链接：https://arxiv.org/abs/2605.07363
备注：https://github.com/MuLabPKU/TransArch
摘要 ：DeepSeek Sparse Attention (DSA) sets the state of the art for fine-grained inference-time sparse attention by introducing a learned token-wise indexer that scores every prefix token and selects the most relevant ones for the main attention. To remain expressive, the indexer uses many query heads (for example, 64 on DeepSeek-V3.2) that share the same selected token set; this multi-head design is precisely what makes the indexer the dominant cost on long contexts. We propose MISA (Mixture of Indexer Sparse Attention), a drop-in replacement for the DSA indexer that treats its indexer heads as a pool of mixture-of-experts. A lightweight router uses cheap block-level statistics to pick a query-dependent subset of only a few active heads, and only those heads run the heavy token-level scoring. This preserves the diversity of the original indexer pool while reducing the per-query cost from scoring every prefix token with every head to scoring it with only a handful of routed heads, plus a negligible router term computed on a small set of pooled keys. We further introduce a hierarchical variant of MISA that uses the routed pass to keep an enlarged candidate set and then re-ranks it with the original DSA indexer to recover the final selected tokens almost exactly. With only eight active heads and no additional training, MISA matches the dense DSA indexer on LongBench across DeepSeek-V3.2 and GLM-5 while running with eight and four times fewer indexer heads respectively, and outperforms HISA on average. It also preserves fully green Needle-in-a-Haystack heatmaps up to a 128K-token context and recovers more than 92% of the tokens selected by the DSA indexer per layer. Our TileLang kernel delivers roughly a 3.82 times speedup over DSA's original indexer kernel on a single NVIDIA H200 GPU.
【20】Mage: Multi-Axis Evaluation of LLM-Generated Executable Game Scenes Beyond Compile-Pass Rate
标题：Mage：LLM生成的可执行游戏场景的多轴评估超出通知通过率
链接：https://arxiv.org/abs/2605.07342
备注：Main Content: 10 pages, 1 figure. In total 22 pages
【21】Rethinking Importance Sampling in LLM Policy Optimization: A Cumulative Token Perspective
标题：重新思考LLM政策优化中的重要性抽样：累积代币视角
链接：https://arxiv.org/abs/2605.07331
【22】Discovering Ordinary Differential Equations with LLM-Based Qualitative and Quantitative Evaluation
标题：利用基于LLM的定性和定量评估发现常微方程
链接：https://arxiv.org/abs/2605.07323
备注：Accepted at ICML 2026
【23】The Convergence Gap: Instruction-Tuned Language Models Stabilize Later in the Forward Pass
标题：收敛差距：指令调整的语言模型在向前传递中稳定下来
链接：https://arxiv.org/abs/2605.07282
【24】When Are Experts Misrouted? Counterfactual Routing Analysis in Mixture-of-Experts Language Models
标题：专家何时误入歧途？专家混合语言模型中的反事实路由分析
链接：https://arxiv.org/abs/2605.07260
【25】Experience Sharing in Mutual Reinforcement Learning for Heterogeneous Language Models
标题：异类语言模型相互强化学习的经验分享
链接：https://arxiv.org/abs/2605.07244
备注：50 pages, 10 figures, 14 tables
【26】PSK@EEUCA 2026: Fine-Tuning Large Language Models with Synthetic Data Augmentation for Multi-Class Toxicity Detection in Gaming Chat
标题：PKC @EEUCA 2026：通过合成数据增强微调大型语言模型，用于游戏聊天中的多类毒性检测
链接：https://arxiv.org/abs/2605.07201
备注：Accepted to the EEUCA workshop at ACL 2026
【27】Star Elastic: Many-in-One Reasoning LLMs with Efficient Budget Control
标题：Star Elastic：具有高效预算控制的多合一推理LLM
链接：https://arxiv.org/abs/2605.07182
【28】Adaptive Negative Reinforcement for LLM Reasoning:Dynamically Balancing Correction and Diversity in RLVR
标题：LLM推理的自适应负强化：动态平衡WLVR中的纠正和多样性
链接：https://arxiv.org/abs/2605.07137
【29】RRCM: Ranking-Driven Retrieval over Collaborative and Meta Memories for LLM Recommendation
标题：WRCM：基于协作和Meta的排名驱动检索LLM推荐
链接：https://arxiv.org/abs/2605.07129
【30】The Position Curse: LLMs Struggle to Locate the Last Few Items in a List
标题：职位诅咒：LLM难以找到列表中的最后几个项目
链接：https://arxiv.org/abs/2605.07127
【31】Theoretical Limits of Language Model Alignment
标题：语言模型对齐的理论局限
链接：https://arxiv.org/abs/2605.07105
【32】Self-Consolidating Language Models: Continual Knowledge Incorporation from Context
标题：自我巩固的语言模型：从上下文中不断融入知识
链接：https://arxiv.org/abs/2605.07076
备注：9 pages
【33】Dr. Post-Training: A Data Regularization Perspective on LLM Post-Training
标题：博士后训练：LLM后训练的数据正规化视角
链接：https://arxiv.org/abs/2605.07063
【34】A Systematic Investigation of The RL-Jailbreaker in LLMs
标题：对LLC中RL越狱者的系统调查
链接：https://arxiv.org/abs/2605.07032
备注：Warning: To demonstrate vulnerabilities, this paper contains unfiltered and potentially offensive jailbreaking examples. Reader discretion advised
【35】Bias and Uncertainty in LLM-as-a-Judge Estimation
标题：LLM-as-a-Judge估计中的偏差和不确定性
链接：https://arxiv.org/abs/2605.06939
【36】A Reproducible Optimisation Protocol for Calibrating Prompt-Based Large Language Model Workflows in Evidence Synthesis
标题：证据合成中校准基于预算的大型语言模型工作流的可重复优化协议
链接：https://arxiv.org/abs/2605.06937
【37】LLMs are not (consistently) Bayesian: Quantifying internal (in)consistencies of LLMs' probabilistic beliefs
标题：LLM不是（一致）Bayesian：量化LLM概率信念的内部（在）状态
链接：https://arxiv.org/abs/2605.06915
【38】Same Signal, Opposite Meaning: Direction-Informed Adaptive Learning for LLM Agents
标题：相同的信号，相反的含义：LLM代理的方向知情自适应学习
链接：https://arxiv.org/abs/2605.06908
【39】Dataset Watermarking for Closed LLMs with Provable Detection
标题：具有可证明检测的封闭LLM的数据集水印
链接：https://arxiv.org/abs/2605.06865
【40】Semantic State Abstraction Interfaces for LLM-Augmented Portfolio Decisions: Multi-Axis News Decomposition and RL Diagnostics
标题：LLM增强投资组合决策的语义状态抽象接口：多轴新闻分解和RL诊断
链接：https://arxiv.org/abs/2605.06730
备注：18 pages, 3 figures. NeurIPS 2024 manuscript style (preprint)
【41】When Does a Language Model Commit? A Finite-Answer Theory of Pre-Verbalization Commitment
标题：语言模型何时提交？言语化前承诺的一个问答理论
链接：https://arxiv.org/abs/2605.06723
【42】CASCADE: Case-Based Continual Adaptation for Large Language Models During Deployment
标题：CASCADE：部署期间大型语言模型的基于案例的连续适应
链接：https://arxiv.org/abs/2605.06702
【43】LKV: End-to-End Learning of Head-wise Budgets and Token Selection for LLM KV Cache Eviction
标题：LKV：LLM KV缓存驱逐的头到端预算和令牌选择的端到端学习
链接：https://arxiv.org/abs/2605.06676
【44】Domain-level metacognitive monitoring in frontier LLMs: A 33-model atlas
标题：前沿LLM中的领域级元认知监控：33个模型地图集
链接：https://arxiv.org/abs/2605.06673
备注：25 pages, 7 figures, 1 supplementary table. Code and data: https://github.com/synthiumjp/metacognitive-profile-atlas
【45】Evaluating Prompt Injection Defenses for Educational LLM Tutors: Security-Usability-Latency Trade-offs
标题：评估教育LLM导师的即时注入防御：安全性-可用性-延迟权衡
链接：https://arxiv.org/abs/2605.06669
备注：18 pages, 4 figures, 9 tables
【46】MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems
标题：MASPO：基于LLM的多智能体系统的联合提示优化
链接：https://arxiv.org/abs/2605.06623
备注：Accepted at ICML 2026
【47】CommFuse: Hiding Tail Latency via Communication Decomposition and Fusion for Distributed LLM Training
标题：启动：通过分布式LLM训练的通信分解和融合隐藏尾部延迟
链接：https://arxiv.org/abs/2604.24013
备注：Slightly modified the title, and corresponding minor wording change in the content
【48】An Interpretable and Scalable Framework for Evaluating Large Language Models
标题：用于评估大型语言模型的可解释和可扩展框架
链接：https://arxiv.org/abs/2605.07046
Graph相关(图学习|图神经网络|图优化等)(14篇)
【1】GRAPHLCP: Structure-Aware Localized Conformal Prediction on Graphs
标题：GRAPHLCP：图形上的结构感知局部共形预测
链接：https://arxiv.org/abs/2605.08074
备注：20 pages, 9 Figures, 8 Tables
【2】Graph-Structured Hyperdimensional Computing for Data-Efficient and Explainable Process-Structure-Property Prediction
标题：用于数据高效和可解释的流程结构属性预测的图结构多维计算
链接：https://arxiv.org/abs/2605.07999
备注：19 pages, 18 figures
【3】GRASP -- Graph-Based Anomaly Detection Through Self-Supervised Classification
标题：GRASP --通过自我监督分类的基于图的异常检测
链接：https://arxiv.org/abs/2605.07812
备注：17 pages
【4】Bilevel Graph Structure Learning, Revisited: Inner-Channel Origins of the Reported Gain
标题：双层图结构学习，重温：报告收益的内渠道起源
链接：https://arxiv.org/abs/2605.07577
【5】GESR: Graph-Based Edge Semantic Reconstruction for Stealthy Communication Detection with Benign-Only Training
标题：GESR：基于图的边缘语义重建，用于纯良性训练的隐形通信检测
链接：https://arxiv.org/abs/2605.07536
【6】Have Graph -- Will Lift? The Case for Higher-Order Benchmarks
标题：有图表-- Will Lift吗？更高级基准的案例
链接：https://arxiv.org/abs/2605.07397
【7】PerCaM-Health: Personalized Dynamic Causal Graphs for Healthcare Reasoning
标题：PerCaM-Health：用于医疗保健推理的个性化动态因果图
链接：https://arxiv.org/abs/2605.07267
【8】Learning Multi-Relational Graph Representations for DNA Methylation-Based Biological Age Estimation
标题：学习基于DNA甲基化的生物年龄估计的多关系图表示
链接：https://arxiv.org/abs/2605.07175
【9】GAD in the Wild: Benchmarking Graph Anomaly Detection under Realistic Deployment Challenges
标题：野外GAD：现实部署挑战下的基准图异常检测
链接
：https://arxiv.org/abs/2605.07133
【10】AdaTKG: Adaptive Memory for Temporal Knowledge Graph Reasoning
标题：AdaTKG：用于时态知识图推理的自适应记忆
链接：https://arxiv.org/abs/2605.07121
【11】Solving Max-Cut to Global Optimality via Feasibility-Preserving Graph Neural Networks
标题：通过保留可能性的图神经网络解决全局最优的最大切
链接：https://arxiv.org/abs/2605.07113
【12】Unlocking High-Fidelity Molecular Generation from Mass Spectra via Dual-Stream Line Graph Diffusion
标题：基于双流线图扩散的质谱高保真分子生成
链接：https://arxiv.org/abs/2605.07048
【13】Inductive Power Grid Cascading Failure Analysis with GRU-Gated Graph Attention
标题：具有GRU门控图关注的感应电网级联故障分析
链接：https://arxiv.org/abs/2605.07010
备注：10 pages, 10 figures, IEEE format
【14】From Model to Data (M2D): Shifting Complexity from GNNs to Graphs for Transparent Graph Learning
标题：从模型到数据（M2 D）：将复杂性从GNN转移到图形以实现透明图形学习
链接：https://arxiv.org/abs/2605.06814
Transformer(11篇)
【1】Fast Byte Latent Transformer
标题：快速字节潜伏Transformer
链接：https://arxiv.org/abs/2605.08044
【2】Training-Induced Escape from Token Clustering in a Mean-Field Formulation of Transformers
标题：Transformer平均场公式中训练诱导的代币聚集逃避
链接：https://arxiv.org/abs/2605.07772
备注：48 pages, 6 figures, comments are wellcome!
【3】Revisiting Transformer Layer Parameterization Through Causal Energy Minimization
标题：基于因果能量最小化的Transformer层参数化方法研究
链接：https://arxiv.org/abs/2605.07588
【4】Approximation Error Upper and Lower Bounds for Hölder Class with Transformers
标题：带变形器的Hölder类的逼近误差上界和下界
链接：https://arxiv.org/abs/2605.07463
备注：31 pages, 2 figures. Accepted by ICML2026
【5】Beyond Linear Attention: Softmax Transformers Implement In-Context Reinforcement Learning
标题：超越线性注意力：Softmax Transformers实现上下文强化学习
链接：https://arxiv.org/abs/2605.07333
【6】Attention Transfer Is Not Universally Effective for Vision Transformers
标题：注意力转移对视觉Transformer来说并不普遍有效
链接：https://arxiv.org/abs/2605.07191
【7】ProtSent: Protein Sentence Transformers
标题：ProtSent：蛋白质句子Transformer
链接：https://arxiv.org/abs/2605.06830
备注：9 figures, appendix, 2 figures, open code and models
【8】The E$Δ$-MHC-Geo Transformer: Adaptive Geodesic Operations with Guaranteed Orthogonality
标题：E$Δ$-MHC-Geo Transformer：具有保证同色性的自适应测地操作
链接：https://arxiv.org/abs/2605.06729
备注：21 pages, 8 figures; code will be available at https://github.com/arash-shahmansoori/edelta
【9】Transformer-Based Wildlife Species Classification from Daily Movement Trajectories
标题：基于变形者的每日运动轨迹的野生动物物种分类
链接：https://arxiv.org/abs/2605.06726
备注：8 pages
【10】Mean Mode Screaming: Mean--Variance Split Residuals for 1000-Layer Diffusion Transformers
标题：平均模式尖叫：1000层扩散Transformer的平均-方差分裂残留
链接：https://arxiv.org/abs/2605.06169
备注：43 pages (9-page main paper + appendix)
【11】Spectrum-Adaptive Generalization Bounds for Trained Deep Transformers
标题：训练有素的深度Transformer的光谱自适应概括界限
链接：https://arxiv.org/abs/2605.07297
GAN|对抗|攻击|生成相关(16篇)
【1】DVD: Discrete Voxel Diffusion for 3D Generation and Editing
标题：DVD：用于3D生成和编辑的离散体素扩散
链接：https://arxiv.org/abs/2605.07971
【2】Slowly Annealed Langevin Dynamics: Theory and Applications to Training-Free Guided Generation
标题：缓慢软化的Langevin动力学：理论和在免训练引导生成中的应用
链接：https://arxiv.org/abs/2605.07950
【3】Disagreement-Regularized Importance Sampling for Adversarial Label Corruption
标题：对抗性标签腐败的分歧正规化重要性抽样
链接：https://arxiv.org/abs/2605.07551
【4】Sparse Autoencoders as Plug-and-Play Firewalls for Adversarial Attack Detection in VLMs
标题：稀疏自动编码器作为即插即用防火墙，用于VLM中的对抗性攻击检测
链接：https://arxiv.org/abs/2605.07447
【5】Generating training datasets for legal chatbots in Korean
标题：为韩语法律聊天机器人生成训练数据集
链接：https://arxiv.org/abs/2605.07432
【6】MIPIAD: Multilingual Indirect Prompt Injection Attack Defense with Qwen -- TF-IDF Hybrid and Meta-Ensemble Learning
标题：MIPIAD：使用Qwen -- TF-IDF混合和元集合学习的多语言间接提示注入攻击防御
链接：https://arxiv.org/abs/2605.07269
【7】PaT: Planning-after-Trial for Efficient Test-Time Code Generation
标题：PaT：有效的测试时代码生成的试验后规划
链接：https://arxiv.org/abs/2605.07248
备注：Accepted to ACL 2026 main conference
【8】Coupling Models for One-Step Discrete Generation
标题：一步离散发电的耦合模型
链接：https://arxiv.org/abs/2605.07193
备注：Code is available at https://github.com/pengzhangzhi/Coupling-Models
【9】Can You Break RLVER? Probing Adversarial Robustness of RL-Trained Empathetic Agents
标题：你能打破RLVER吗？探索RL训练的同理心代理人的对抗稳健性
链接：https://arxiv.org/abs/2605.07138
【10】FlashMol: High-Quality Molecule Generation in as Few as Four Steps
标题：Flash Mol：只需四个步骤即可产生高质量分子
链接：https://arxiv.org/abs/2605.07020
【11】Dual-Agent Co-Training for Health Coaching via Implicit Adversarial Preference Optimization
标题：通过隐式对抗偏好优化的健康教练双代理联合训练
链接：https://arxiv.org/abs/2605.07011
【12】Rollback-Free Stable Brick Structures Generation
标题：无回滚稳定砖结构生成
链接：https://arxiv.org/abs/2605.06947
【13】Streaming Adversarial Robustness in Fuzzy ARTMAP: Mechanism-Aligned Evaluation, Progressive Training, and Interpretable Diagnostics
标题：模糊ARTMAP中的流媒体对抗鲁棒性：机制一致的评估、渐进式训练和可解释诊断
链接：https://arxiv.org/abs/2605.06902
备注 ：35 pages, 3 figures, 11 tables. Preprint submitted to Neural Networks
【14】Conditional generation of antibody sequences with classifier-guided germline-absorbing discrete diffusion
标题：基于分类器引导的种系吸收离散扩散的抗体序列条件生成
链接：https://arxiv.org/abs/2605.06720
备注：9 pages, 2 figures, 2 tables
【15】A Wasserstein GAN-based climate scenario generator for risk management and insurance: the case of soil subsidence
标题：基于Wasserstein GAN的风险管理和保险气候情景生成器：土壤沉降案例
链接：https://arxiv.org/abs/2605.06678
【16】Debiased Counterfactual Generation via Flow Matching from Observations
标题：通过观测数据流匹配去偏反事实生成
链接：https://arxiv.org/abs/2605.07665
半/弱/无/有监督|不确定性|主动学习(9篇)
【1】TAVIS: A Benchmark for Egocentric Active Vision and Anticipatory Gaze in Imitation Learning
标题：TARIS：模仿学习中自我中心主动视觉和预期凝视的基准
链接：https://arxiv.org/abs/2605.07943
【2】Flexible Routing via Uncertainty Decomposition
标题：通过不确定性分解灵活路由
链接：https://arxiv.org/abs/2605.07805
【3】SMT-Based Active Learning of Weighted Automata
标题：基于MSG的加权自动机主动学习
链接：https://arxiv.org/abs/2605.07758
备注：Appearing in CAV 2026
【4】Robust and Reliable AI for Predictive Quality in Semiconductor Materials Manufacturing with MLOps and Uncertainty Quantification
标题：通过MLOps和不确定性量化，实现半导体材料制造中预测质量的稳健可靠的人工智能
链接：https://arxiv.org/abs/2605.07752
【5】Beyond Distribution Estimation: Simplex Anchored Structural Inference Towards Universal Semi-Supervised Learning
标题：超越分布估计：面向通用半监督学习的单纯形锚定结构推理
链接：https://arxiv.org/abs/2605.07557
备注：The paper is accepted by ICML 2026
【6】SR$^2$-LoRA: Self-Rectifying Inter-layer Relations in Low-Rank Adaptation for Class-Incremental Learning
标题：SR $#2 $-LoRA：类增量学习的低等级适应中的自纠正层间关系
链接：https://arxiv.org/abs/2605.07420
【7】Physics-based Digital Twins for Integrated Thermal Energy Systems Using Active Learning
标题：使用主动学习的集成热能系统基于物理的数字双胞胎
链接：https://arxiv.org/abs/2605.06756
备注：23 pages, 12 figures, and 2 tables
【8】Enabling Unsupervised Training of Deep EEG Denoisers With Intelligent Partitioning
标题：通过智能分区实现深度EEG去噪器的无监督训练
链接：https://arxiv.org/abs/2605.06724
【9】A Refined Generalization Analysis for Extreme Multi-class Supervised Contrastive Representation Learning
标题：极端多类监督对比表示学习的精细概括分析
链接：https://arxiv.org/abs/2605.07596
备注：Accepted at ICML 2026
迁移|Zero/Few/One-Shot|自适应(15篇)
【1】Zero-Shot Imagined Speech Decoding via Imagined-to-Listened MEG Mapping
标题：通过想象到收听的MEG映射的Zero-Shot想象语音解码
链接：https://arxiv.org/abs/2605.08075
【2】PET-Adapter: Test-Time Domain Adaptation for Full and Limited-Angle PET Image Reconstruction
标题：PET-适配器：全角和有限角PET图像重建的测试时间域自适应
链接：https://arxiv.org/abs/2605.08030
【3】Adaptive Domain Decomposition Physics-Informed Neural Networks for Traffic State Estimation with Sparse Sensor Data
标题：自适应区域分解物理信息神经网络用于稀疏传感器数据的交通状态估计
链接：https://arxiv.org/abs/2605.08028
备注：56 pages, 5 figures, 12 tables. Submitted to Transportation Research Part C
【4】STEPS: A Temporal Smooth Error Propagation Solver on the Manifolds for Test-Time Adaptation in Time Series Forecasting
标题：STEPS：时间序列预测中测试时间自适应的Manifest上的时间平滑误差传播求解器
链接：https://arxiv.org/abs/2605.08005
备注：9 pages main text, appendix included. 7 figures. Submitted to NeurIPS 2026
【5】Adaptive Regularization for Sparsity Control in Bregman-Based Optimizers
标题：基于Bregman的优化器中稀疏性控制的自适应正规化
链接：https://arxiv.org/abs/2605.07892
备注：21 pages, 15 figures
【6】ExpThink: Experience-Guided Reinforcement Learning for Adaptive Chain-of-Thought Compression
标题：ExpThink：用于自适应思维链压缩的经验引导强化学习
链接：https://arxiv.org/abs/2605.07501
备注：39 pages, 18 figures. Code and model checkpoints will be released upon publication
【7】Transfer Learning Across Fast- and Full-Simulation Domains in High-Energy Physics
标题：高能物理中快速和全模拟领域的迁移学习
链接：https://arxiv.org/abs/2605.07471
备注 ：16 pages, 8 figures
【8】A Flexible Adaptive Stable Clustering Algorithm for Archive-Scale Online Mass Spectrometry
标题：档案规模在线MS的灵活自适应稳定聚集算法
链接：https://arxiv.org/abs/2605.07424
【9】Zero-Shot Neural Network Evaluation with Sample-Wise Activation Patterns
标题：采用样本激活模式的零激发神经网络评估
链接：https://arxiv.org/abs/2605.07378
备注：Accepted by IEEE Transactions on Pattern Analysis and Machine Intelligence. arXiv admin note: text overlap with arXiv:2403.04161
【10】Pretraining Induces a Reusable Spectral Basis for Downstream Task Adaptation
标题：预训练为下游任务适应引入可重复使用的频谱基础
链接：https://arxiv.org/abs/2605.07302
【11】Adaptive Memory Decay for Log-Linear Attention
标题：日志线性注意力的自适应记忆衰退
链接：https://arxiv.org/abs/2605.06946
备注：19 pages, 13 figures. Preprint
【12】Learned Lyapunov Shielding for Adaptive Control
标题：用于自适应控制的学习李亚普诺夫屏蔽
链接：https://arxiv.org/abs/2605.06934
【13】Temporal Attention for Adaptive Control of Euler-Lagrange Systems with Unobservable Memory
标题：具有不可观察记忆的欧拉-拉格朗日系统自适应控制的时间注意力
链接：https://arxiv.org/abs/2605.06877
【14】Knowledge Transfer Scaling Laws for 3D Medical Imaging
标题：3D医学成像的知识转移缩放定律
链接：https://arxiv.org/abs/2605.06859
备注：20 Pages
【15】STDA-Net: Spectrogram-Based Domain Adaptation for cross-dataset Sleep Stage Classification
标题：STDA-Net：跨数据集睡眠阶段分类的基于谱图的领域自适应
链接：https://arxiv.org/abs/2605.06736
备注：submitted to IEEE SMC conference
强化学习(11篇)
【1】Reinforcement Learning for Exponential Utility: Algorithms and Convergence in Discounted MDPs
标题：指数效用的强化学习：折扣MDP中的算法和收敛
链接：https://arxiv.org/abs/2605.08053
【2】Interpreting Reinforcement Learning Agents with Susceptibilities
标题：解释具有敏感性的强化学习代理
链接：https://arxiv.org/abs/2605.08007
备注：55 pages, comments welcome
【3】LiteGUI: Distilling Compact GUI Agents with Reinforcement Learning
标题：LiteGUI：利用强化学习提炼紧凑的GUI代理
链接：https://arxiv.org/abs/2605.07505
【4】Improved Model-based Reinforcement Learning with Smooth Kernels
标题：具有光滑核的改进的基于模型的强化学习
链接：https://arxiv.org/abs/2605.07218
备注：38 pages, 5 figures
【5】HyperEyes: Dual-Grained Efficiency-Aware Reinforcement Learning for Parallel Multimodal Search Agents
标题：HyperEyes：并行多模式搜索代理的双粒度高效感知强化学习
链接：https://arxiv.org/abs/2605.07177
备注：Code & Data: https://github.com/Guankai-Li/HyperEyes
【6】Convergence and Emergence of In-Context Reinforcement Learning with Chain of Thought
标题：与思想链的上下文强化学习的融合和出现
链接：https://arxiv.org/abs/2605.07123
【7】Stabilized neural Hamilton--Jacobi--Bellman solvers: Error analysis and applications in model-based reinforcement learning
标题：稳定神经Hamilton-Jacobi-Bellman求解器：基于模型的强化学习中的误差分析和应用
链接：https://arxiv.org/abs/2605.07116
【8】Almost Sure Convergence Rates of Stochastic Approximation and Reinforcement Learning via a Poisson-Moreau Drift
标题：Poisson-Moreau漂移的随机逼近和强化学习的几乎肯定收敛率
链接：https://arxiv.org/abs/2605.07104
【9】Towards Differentially Private Reinforcement Learning with General Function Approximation
标题：利用一般函数逼近实现差异私密强化学习
链接：https://arxiv.org/abs/2605.07049
【10】Multi-Objective Constraint Inference using Inverse reinforcement learning
标题：使用反向强化学习的多目标约束推理
链接：https://arxiv.org/abs/2605.06951
【11】Revisiting Adam for Streaming Reinforcement Learning
标题：重温Adam的流强化学习
链接：https://arxiv.org/abs/2605.06764
符号|符号学习(1篇)
【1】Emergent Symbolic Structure in Health Foundation Models: Extraction, Alignment, and Cross-Modal Transfer
标题：健康基金会模型中的紧急符号结构：提取、对齐和跨模式转移
链接：https://arxiv.org/abs/2605.07407
备注：8 pages ICML workshop, 4 main figures
分层学习(1篇)
【1】Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoders
标题：树型AE：在稀疏自动编码器中学习分层特征结构
链接：https://arxiv.org/abs/2605.07922
备注：21 pages
医学相关(6篇)
【1】Spectral Surgery: Class-Targeted Post-Hoc Rebalancing via Hessian Spike Perturbation
标题：光谱手术：通过Hessian Spike扰动的针对类别的事后再平衡
链接：https://arxiv.org/abs/2605.07790
【2】Self Driving Datasets: From 20 Million Papers to Nuanced Biomedical Knowledge at Scale
标题：自动驾驶数据集：从2000万篇论文到大规模细致入微的生物医学知识
链接：https://arxiv.org/abs/2605.07022
【3】Medical Imaging Classification with Cold-Atom Reservoir Computing using Auto-Encoders and Surrogate-Driven Training
标题：使用自动编码器和代理人驱动训练的冷原子库计算的医学成像分类
链接：https://arxiv.org/abs/2605.06727
备注：8 pages, 6 figures. Accepted to the 2025 IEEE International Conference on Quantum AI (IEEE QAI). Supported by FCT and the Open Quantum Institute (OQI)
【4】XDecomposer: Learning Prior-Free Set Decomposition for Multiphase X-ray Diffraction
标题：XDecomoser：学习多相X射线散射的无优先集分解
链接：https://arxiv.org/abs/2605.05866
备注：28pages, 8figures, 6tables
【5】PPI-Net connects molecular protein interactions to functional processes in disease
标题：PPI-Net将分子蛋白质相互作用与疾病功能过程联系起来
链接：https://arxiv.org/abs/2605.07838
备注：17 pages, 3 figures, 2 tables
【6】Multimodal synthesis of MRI and tabular data with diffusion in a joint latent space via cross-attention
标题：MRI和表格数据的多模式合成，通过交叉注意在关节潜在空间中扩散
链接：https://arxiv.org/abs/2605.06699
蒸馏|知识提取(9篇)
【1】Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation
标题：轨迹作为老师：通过能量导航蒸馏的少步离散流匹配
链接：https://arxiv.org/abs/2605.07924
【2】KL for a KL: On-Policy Distillation with Control Variate Baseline
标题：KL的KL：含控制变量基线的按政策蒸馏
链接：https://arxiv.org/abs/2605.07865
【3】Prune-OPD: Efficient and Reliable On-Policy Distillation for Long-Horizon Reasoning
标题：Prune-OPD：高效可靠的按策略蒸馏，用于长期推理
链接：https://arxiv.org/abs/2605.07804
备注：17 pages, 8 figures
【4】Stochastic Transition-Map Distillation for Fast Probabilistic Inference
标题：快速概率推理的随机转移图蒸馏
链接：https://arxiv.org/abs/2605.07661
【5】SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion
标题：SHRED：通过Logit Demotion的自蒸馏进行无保留消除学习
链接：https://arxiv.org/abs/2605.07482
【6】Rubric-based On-policy Distillation
标题：基于条目的按政策蒸馏
链接：https://arxiv.org/abs/2605.07396
备注：Preprint. Code is available at https://github.com/Peregrine123/ROPD_official
【7】Closed-Form Linear-Probe Dataset Distillation for Pre-trained Vision Models
标题：预训练视觉模型的封闭形式线性探针数据集蒸馏
链接：https://arxiv.org/abs/2605.07194
【8】Structural Rationale Distillation via Reasoning Space Compression
标题：通过推理空间压缩进行结构原理蒸馏
链接：https://arxiv.org/abs/2605.07139
【9】How to Compress KV Cache in RL Post-Training? Shadow Mask Distillation for Memory-Efficient Alignment
标题：如何在RL后训练中压缩KV缓存？用于存储器高效对齐的阴影掩模蒸馏
链接：https://arxiv.org/abs/2605.06850
聚类(3篇)
【1】Simple KNN-Based Outlier Detection Achieves Robust Clustering
标题：简单的基于KNN的离群点检测实现鲁棒性集群
链接：https://arxiv.org/abs/2605.07130
备注：Code: https://github.com/MasterZhou1/Robust-Clustering
【2】Classification Fields: Arbitrarily Fine Recursive Hierarchical Clustering From Few Examples
标题：分类字段：来自少数示例的初步精细的递阶分层集群
链接：https://arxiv.org/abs/2605.07119
【3】Drawing Lines in Psychological Space: What K-means Clustering Reveals in Simulated and Real Psychometric Data
标题：心理空间中的界限：K均值聚集在模拟和真实心理测量数据中揭示了什么
链接：https://arxiv.org/abs/2605.06989
备注：Methodological study on K-means clustering in psychometric data using simulated and empirical datasets
超分辨率|去噪|去模糊|去雾(1篇)
【1】Kurtosis-Guided Denoising Score Matching for Tabular Anomaly Detection
标题：用于表格异常检测的突变引导去噪分数匹配
链接：https://arxiv.org/abs/2605.06955
备注：39 pages, 10 figures, 14 tables
自动驾驶|车辆|车道检测等(1篇)
【1】From Canopy to Collision: A Hybrid Predictive Framework for Identifying Risk Factors in Tree-Involved Traffic Crashes
标题：从树冠到碰撞：识别树木交通事故风险因素的混合预测框架
链接：https://arxiv.org/abs/2605.06684
备注：30 pages, 10 figures
联邦学习|隐私保护|加密(4篇)
【1】FLAM: Evaluating Model Performance with Aggregatable Measures in Federated Learning
标题：FLAM：在联邦学习中使用可聚合指标评估模型性能
链接：https://arxiv.org/abs/2605.07962
备注：Accepted for publication in 2nd IEEE International Conference on Federated Learning and Intelligent Computing Systems(FLICS2026)
【2】Enhancing Federated Quadruplet Learning: Stochastic Client Selection and Embedding Stability Analysis
标题：增强联邦四重学习：随机客户选择和嵌入稳定性分析
链接：https://arxiv.org/abs/2605.07888
备注：arXiv admin note: substantial text overlap with arXiv:2509.04107
【3】HARMONY: Bridging the Personalization-Generalization Gap by Mitigating Representation Skew in Heterogeneous Split Federated Learning
标题：HARMONY：通过缓解异类分离联邦学习中的表示偏差来弥合个性化与概括化的差距
链接：https://arxiv.org/abs/2605.07211
备注：7 pages (except references), 5 figures
【4】Resource-Element Energy Difference for Noncoherent Over-the-Air Federated Learning
标题：非连贯空中联邦学习的资源元素能量差异
链接：https://arxiv.org/abs/2605.07263
备注：Preprint; Under-review; Codes to replicate the results is available at: https://github.com/zavareh1/REED
推理|分析|理解|解释(23篇)
【1】Bayesian Sensitivity of Causal Inference Estimators under Evidence-Based Priors
标题：基于证据的先验条件下因果推理估计器的Bayesian敏感性
链接：https://arxiv.org/abs/2605.07993
备注：TMLR 2026
【2】It Just Takes Two: Scaling Amortized Inference to Large Sets
标题：只需两个：将摊销推理扩展到大集
链接：https://arxiv.org/abs/2605.07972
【3】Convergent Stochastic Training of Attention and Understanding LoRA
标题：注意力和理解LoRA的收敛随机训练
链接：https://arxiv.org/abs/2605.07959
【4】An Efficient Hybrid Sparse Attention with CPU-GPU Parallelism for Long-Context Inference
标题：一种具有CPU-GPU-并行性的高效混合稀疏注意力用于长上下文推理
链接：https://arxiv.org/abs/2605.07719
【5】FactoryBench: Evaluating Industrial Machine Understanding
标题：FactoryBench：评估工业机器理解
链接：https://arxiv.org/abs/2605.07675
备注：9 pages, 4 figures, 14 tables; appendix with 24 pages
【6】SGD for Variational Inference: Tackling Unbounded Variance via Preconditioning and Dynamic Batching
标题：变分推理的SGD：通过预处理和动态批处理无界方差
链接：https://arxiv.org/abs/2605.07531
【7】Why Self-Inconsistency Arises in GNN Explanations and How to Exploit It
标题：GNN解释中为何会出现自我不一致以及如何利用它
链接：https://arxiv.org/abs/2605.07527
【8】Inference-Time Attribute Distribution Alignment for Unconditional Diffusion
标题：无条件扩散的推理时属性分布对齐
链接：https://arxiv.org/abs/2605.07456
备注：Preprint. 35 pages, 13 figures
【9】SSP-based construction of evaluation-annotated data for fine-grained aspect-based sentiment analysis
标题：基于DSP的评估注释数据构建，用于细粒度的基于方面的情感分析
链接：https://arxiv.org/abs/2605.07446
【10】StreamPhy: Streaming Inference of High-Dimensional Physical Dynamics via State Space Models
标题：StreamPhy：基于状态空间模型的高维物理动力学流推理
链接：https://arxiv.org/abs/2605.07384
【11】Structured Role-Aware Policy Optimization for Multimodal Reasoning
标题：多模式推理的结构化角色感知策略优化
链接：https://arxiv.org/abs/2605.07274
备注：32 pages
【12】Three-in-One World Model: Energy-Based Consistency, Prediction, and Counterfactual Inference for Marketing Intervention
标题：三位一体世界模型：基于能源的一致性、预测和营销干预的反事实推理
链接：https://arxiv.org/abs/2605.07199
【13】IntentGrasp: A Comprehensive Benchmark for Intent Understanding
标题：IntentGrasp：意图理解的全面基准
链接：https://arxiv.org/abs/2605.06832
备注：IntentGrasp data is available on [Hugging Face](https://huggingface.co/datasets/yuweiyin/IntentGrasp), and the code is released on [GitHub](https://github.com/YuweiYin/IntentGrasp)
【14】Why DDIM Hallucinates More than DDPM: A Theoretical Analysis of Reverse Dynamics
标题：为什么DDIM比DDPM更容易产生幻觉：反向动力学的理论分析
链接：https://arxiv.org/abs/2605.06831
【15】A Theory of Online Learning with Autoregressive Chain-of-Thought Reasoning
标题：具有自回归思维链推理的在线学习理论
链接：https://arxiv.org/abs/2605.06819
【16】Sparse Attention as a Range Searching Problem: Towards an Inference-Efficient Index for KV Cache
标题：作为范围搜索问题的稀疏注意力：为KV缓存寻找推理效率索引
链接：https://arxiv.org/abs/2605.06763
【17】R$^3$L: Reasoning 3D Layouts from Relative Spatial Relations
标题：R$^3$L：从相对空间关系推理3D布局
链接：https://arxiv.org/abs/2605.06758
备注：ICML 2026
【18】State Representation and Termination for Recursive Reasoning Systems
标题：回归推理系统的状态表示与终止
链接：https://arxiv.org/abs/2605.06690
【19】More Thinking, More Bias: Length-Driven Position Bias in Reasoning Models
标题：更多思考，更多偏见：推理模型中长度驱动的位置偏见
链接：https://arxiv.org/abs/2605.06672
【20】Statistical inference with belief functions: A survey
标题：使用信念函数的统计推断：一项调查
链接：https://arxiv.org/abs/2605.07908
备注：9 pages, 0 figures
【21】Inference of Qualitative Models from Steady-State Data via Weighted MaxSMT
标题：通过加权MaxMT从稳态数据推断定性模型
链接：https://arxiv.org/abs/2605.07433
【22】BGM-IV: an AI-powered Bayesian generative modeling approach for instrumental variable analysis
标题：BGM-IV：一种用于工具变量分析的人工智能支持的Bayesian生成式建模方法
链接：https://arxiv.org/abs/2605.07029
【23】A Differentiable Bayesian Relaxation for Latent Partial-Order Inference
标题：潜在偏序推理的可微Bayesian松弛
链接：https://arxiv.org/abs/2605.06976
检测相关(8篇)
【1】Fortifying Time Series: DTW-Certified Robust Anomaly Detection
标题：强化时间序列：DTW认证的稳健异常检测
链接：https://arxiv.org/abs/2605.07690
【2】Probabilistic Object Detection with Conformal Prediction
标题：基于保形预测的概率对象检测
链接：https://arxiv.org/abs/2605.07549
备注：Code is available at https://github.com/mos-ks/OD-CP
【3】Hallucination Detection via Activations of Open-Weight Proxy Analyzers
标题：通过激活开重代理分析仪进行幻觉检测
链接：https://arxiv.org/abs/2605.07209
备注：12 pages, 4 figures. Code available at https://github.com/hallu-detect/llm_hallucination_detection
【4】Delulu: A Verified Multi-Lingual Benchmark for Code Hallucination Detection in Fill-in-the-Middle Tasks
标题：Delulu：中间填充任务中代码幻觉检测的经过验证的多语言基准
链接：https://arxiv.org/abs/2605.07024
【5】McNdroid: A Longitudinal Multimodal Benchmark for Robust Drift Detection in Android Malware
标题：McNdroid：Android恶意软件中稳健漂移检测的纵向多模式基准
链接：https://arxiv.org/abs/2605.06894
备注：28 pages, 14 figures, 14 tables
【6】TUANDROMD-X: Advanced Entropy and Visual Analytics Dataset for Enhanced Malware Detection and Classification
标题：TUANDROMD-X：用于增强恶意软件检测和分类的高级熵和视觉分析数据集
链接：https://arxiv.org/abs/2605.06718
【7】A Hierarchical Ensemble Pipeline for Anomaly Detection in ESA Satellite Telemetry
标题：ESA卫星遥测中异常检测的分级包围流水线
链接：https://arxiv.org/abs/2605.06681
备注：15 pages, 3 figures, 1 table. Submitted to the ML4ITS workshop at the ECML PKDD 2025 conference. Awarded 2nd place in the final round of the Spacecraft Anomaly Challenge on ESA dataset. (Ranked 1st on the Kaggle public leaderboard and 3rd on the private leaderboard)
【8】You Only Stack Once (YOSO): A Motion-Filtered, Deep-Learning Framework for Detecting Faint Moving Sources
标题：You Only Stack Once（YOSO）：用于检测微弱移动源的运动过滤深度学习框架
链接：https://arxiv.org/abs/2605.06913
备注：Accepted to The Astronomical Journal; 13 pages, 9 figures
分类|识别(6篇)
【1】Position: Mechanistic Interpretability Must Disclose Identification Assumptions for Causal Claims
标题：立场：机械解释性必须披露因果关系主张的识别假设
链接：https://arxiv.org/abs/2605.08012
备注：10 pages, 2 figures. Submitted to NeurIPS 2026 (Position Track)
【2】Aggregation in conformal e-classification
标题：保形e分类中的聚集
链接：https://arxiv.org/abs/2605.07963
备注：23 pages, 10 figures
【3】Black-box model classification under the discriminative factorization
标题：区分因子分解下的黑匣子模型分类
链接：https://arxiv.org/abs/2605.07878
【4】When Symbol Names Should Not Matter: A Logistic Theory of Fresh-Symbol Classification
标题：何时符号名称不重要：新鲜符号分类的逻辑理论
链接：https://arxiv.org/abs/2605.07120
【5】Beyond the Wrapper: Identifying Artifact Reliance in Static Malware Classifiers using TRUSTEE
标题：超越包装：使用TRUSEE识别静态恶意软件分类器中的收件箱依赖性
链接：https://arxiv.org/abs/2605.07034
【6】LookWhen? Fast Video Recognition by Learning When, Where, and What to Compute
标题：看什么时候？通过学习何时、何地和计算内容来快速视频识别
链接：https://arxiv.org/abs/2605.06809
表征(8篇)
【1】Prototype Guided Post-pretraining for Single-Cell Representation Learning
标题：原型引导的单细胞表示学习后预训练
链接：https://arxiv.org/abs/2605.07938
【2】Toward Better Geometric Representations for Molecule Generative Models
标题：分子生成模型的更好的几何表示
链接：https://arxiv.org/abs/2605.07693
【3】Direction-Preserving Number Representations
标题：方向保持数字表示
链接：https://arxiv.org/abs/2605.07662
备注：9 pages excluding appendices and references, 18 in total. 5 figures
【4】TRAJGANR: Trajectory-Centric Urban Multimodal Learning via Geospatially Aligned Neural Representations
标题：TRAJGANR：通过地理空间对齐神经表示的以轨迹为中心的城市多模式学习
链接：https://arxiv.org/abs/2605.06990
【5】Don't Retrain, Align: Adapting Autoregressive LMs to Diffusion LMs via Representation Alignment
标题：不要重新训练，对齐：通过表示对齐将自回归LM适应扩散LM
链接：https://arxiv.org/abs/2605.06885
备注：Code available at https://github.com/pengzhangzhi/Open-dLLM
【6】Beyond Factor Aggregation: Gauge-Aware Low-Rank Server Representations for Federated LoRA
标题：超越因素聚合：联邦LoRA的衡量感知低级别服务器表示
链接：https://arxiv.org/abs/2605.06733
【7】Hidden Coalitions in Multi-Agent AI: A Spectral Diagnostic from Internal Representations
标题：多智能体人工智能中的隐藏联盟：来自内部表示的光谱诊断
链接：https://arxiv.org/abs/2605.06696
备注：18 pages
【8】Learning Cross-Atlas Consistent Brain Disorder Representations via Disentangled Multi-Atlas Functional Connectivity Learning
标题：通过解开的多阿特拉斯功能连接性学习学习跨阿特拉斯一致的脑部疾病表示
链接：https://arxiv.org/abs/2605.07026
3D|3D重建等相关(1篇)
【1】PropSplat: Map-Free RF Field Reconstruction via 3D Gaussian Propagation Splatting
标题：PropSplat：通过3D高斯传播飞溅进行无地图射频场重建
链接：https://arxiv.org/abs/2605.08035
备注：Accepted for presentation at IEEE DySPAN 2026
编码器(1篇)
【1】Estimation of Motor Unit Parameters from Surface Electromyograms using an Informed Autoencoder
标题：使用知情自动编码器从表面肌电信号中估计运动单位参数
链接：https://arxiv.org/abs/2605.07458
优化|敛散性(22篇)
【1】Globally Optimal Training of Spiking Neural Networks via Parameter Reconstruction
标题：通过参数重建实现尖峰神经网络的全局最优训练
链接：https://arxiv.org/abs/2605.08022
【2】Exploring the non-convexity in machine learning using quantum-inspired optimization
标题：使用量子启发的优化探索机器学习中的非凸性
链接：https://arxiv.org/abs/2605.07947
【3】ADKO: Agentic Decentralized Knowledge Optimization
标题：ADKO：庞大的去中心化知识优化
链接：https://arxiv.org/abs/2605.07863
备注：31 pages
【4】NSPOD: acceleratingthe convergence ofKrylov-based iterative linearsolvers via approximated PODs
标题：NSPOD：通过逼近Pod加速基于Krylov的迭代线性求解器的收敛
链接：https://arxiv.org/abs/2605.07828
备注：17 pages, 9 figures, 3 tables
【5】OrScale: Orthogonalised Optimization with Layer-Wise Trust-Ratio Scaling
标题：OrScale：采用分层信任比扩展的个性化优化
链接：https://arxiv.org/abs/2605.07815
【6】Optimal Recourse Summaries via Bi-Objective Decision Tree Learning
标题：通过双目标决策树学习的最佳追索摘要
链接：https://arxiv.org/abs/2605.07598
【7】Efficient Data Selection for Multimodal Models via Incremental Optimization Utility
标题：通过增量优化实用程序为多峰模型进行高效数据选择
链接：https://arxiv.org/abs/2605.07488
【8】Convex Optimization with Nested Evolving Feasible Sets
标题：具有嵌套进化可行集的凸优化
链接：https://arxiv.org/abs/2605.07386
【9】Sample Complexity of Stochastic Optimization with Integer Variables
标题：具有预设变量的随机优化的样本复杂性
链接：https://arxiv.org/abs/2605.07239
【10】Where to Spend Rollouts: Hit-Utility Optimal Rollout Allocation for Group-Based RLVR
标题：在哪里花费铺开：基于组的RLVR的命中效用最优铺开分配
链接：https://arxiv.org/abs/2605.07114
【11】Less Random, More Private: What is the Optimal Subsampling Scheme for DP-SGD?
标题：更少随机，更私密：DP-Singapore的最佳二次抽样方案是什么？
链接：https://arxiv.org/abs/2605.07072
备注：17 pages, 1 table. Submitted to NeurIPS 2026
【12】PLOT: Progressive Localization via Optimal Transport in Neural Causal Abstraction
标题：PLOT：通过神经因果抽象中的最佳传输进行渐进定位
链接：https://arxiv.org/abs/2605.06979
【13】Causal-Aware Foundation-Model for Bilevel Optimization in Discrete Choice Settings
标题：离散选择设置中的两层优化的Cause-Aware基础模型
链接：https://arxiv.org/abs/2605.06941
【14】Multi-Objective Multi-Agent Bandits: From Learning Efficiency to Fairness Optimization
标题：多目标多智能体盗贼：从学习效率到公平优化
链接：https://arxiv.org/abs/2605.06864
【15】Christoffel-DPS: Optimal sensor placement in diffusion posterior sampling for arbitrary distributions
标题：Christoffel-DPS：任意分布的扩散后验抽样中的最佳传感器放置
链接：https://arxiv.org/abs/2605.06861
【16】Distributional Process Reward Models: Calibrated Prediction of Future Rewards via Conditional Optimal Transport
标题：分布过程报酬模型：基于条件最优运输的未来报酬校准预测
链接：https://arxiv.org/abs/2605.06785
【17】Gradient Extrapolation-Based Policy Optimization
标题：基于梯度外推的政策优化
链接：https://arxiv.org/abs/2605.06755
备注：26 pages, 9 figures
【18】RateQuant: Optimal Mixed-Precision KV Cache Quantization via Rate-Distortion Theory
标题：RateQuant：通过速率失真理论进行最佳混合精度KV缓存量化
链接：https://arxiv.org/abs/2605.06675
备注：18 pages, 7 figures, 5 tables
【19】Penalty-Based First-Order Methods for Bilevel Optimization with Minimax and Constrained Lower-Level Problems
标题：带约束下层问题的Minimax双层优化的基于罚函数的一阶方法
链接：https://arxiv.org/abs/2605.08006
【20】Asymptotically Log-Optimal Bayes-Assisted Confidence Sequences for Bounded Means
标题：有界均值的渐进log最优Bayes辅助置信序列
链接：https://arxiv.org/abs/2605.07964
备注：Valentin and Stefano are equal first author
【21】Breaking QAOA's Fixed Target Hamiltonian Barrier: A Fully Connected Quantum Boltzmann Machine via Bilevel Optimization
标题：打破QAOA的固定目标Hamilton障碍：通过二层优化的全连接量子Boltzmann机
链接：https://arxiv.org/abs/2605.07473
备注：34 pages, 8 figures, 3 tables, 1 algorithm
【22】Locally Near Optimal Piecewise Linear Regression in High Dimensions via Difference of Max-Affine Functions
标题：通过最大仿射函数差异的局部近优多维分段线性回归
链接：https://arxiv.org/abs/2605.06959
预测|估计(18篇)
【1】On the Tradeoffs of On-Device Generative Models in Federated Predictive Maintenance Systems
标题：联邦预测维护系统中设备上生成模型的权衡
链接：https://arxiv.org/abs/2605.07860
【2】Interactive Trajectory Planning with Learning-based Distributionally Robust Model Predictive Control and Markov Systems
标题：基于学习的分布鲁棒模型预测控制和马尔科夫系统的交互式轨迹规划
链接：https://arxiv.org/abs/2605.07768
【3】Pre-trained Tabular Foundation Models as Versatile Summary Networks for Neural Posterior Estimation
标题：基于预训练表基模型的神经后验估计通用汇总网络
链接：https://arxiv.org/abs/2605.07765
【4】Future Validity is the Missing Statistic: From Impossibility to $Φ$-Estimation for Grammar-Faithful Speculative Decoding
标题：未来有效性是缺失的统计量：从不可能到$Φ$-估计的语法忠实推测解码
链接：https://arxiv.org/abs/2605.07698
【5】NPMixer: Hierarchical Neighboring Patch Mixing for Time Series Forecasting
标题：NPMixer：用于时间序列预测的分层邻近补丁混合
链接：https://arxiv.org/abs/2605.07476
【6】FlightSense: An End-to-End MLOps Platform for Real-Time Flight Delay Prediction via Rotation-Chain Propagation Features and Agentic Conversational AI
标题：FlightSense：一个端到端MLOps平台，通过旋转链传播特征和统计对话人工智能进行实时航班延误预测
链接：https://arxiv.org/abs/2605.07364
备注：12 pages, 5 figures, 9 tables; machine learning, MLOps, aviation delay prediction
【7】Predictive but Not Plannable: RC-aux for Latent World Models
标题：预测性但不可规划：潜在世界模型的RC辅助
链接：https://arxiv.org/abs/2605.07278
【8】Don't Learn the Shape: Forecasting Periodic Time Series by Rank-1 Decomposition
标题：不要学习形状：通过Rank-1分解预测周期性时间序列
链接：https://arxiv.org/abs/2605.07222
备注：9 pages main text + appendix. Code: https://github.com/TakatoHonda/FLAIR
【9】Same Brain, Different Prediction: How Preprocessing Choices Undermine EEG Decoding Reliability
标题：相同的大脑，不同的预测：预处理选择如何削弱脑电解码的可靠性
链接：https://arxiv.org/abs/2605.07212
【10】FAME: Forecasting Academic Impact via Continuous-Time Manifold Evolution
标题：FAME：通过连续时间多元化演变预测学术影响
链接：https://arxiv.org/abs/2605.07208
【11】Target-Aware Data Augmentation for SAT Prediction
标题：SAT预测的目标感知数据增强
链接：https://arxiv.org/abs/2605.06931
【12】Generalising Travel Time Prediction To Varying Route Choices In Urban Networks
标题：将旅行时间预测推广到城市网络中的不同路线选择
链接：https://arxiv.org/abs/2605.06918
【13】Tyche: One Step Flow for Efficient Probabilistic Weather Forecasting
标题：Tyche：高效概率天气预报的一步流程
链接：https://arxiv.org/abs/2605.06916
【14】Dual-Scale Temporal Fusion Reveals Structured Predictability in Subseasonal-to-Seasonal Temperature Prediction
标题：双尺度时间融合揭示了亚季至季节气温预测的结构化可预测性
链接：https://arxiv.org/abs/2605.06911
备注：10 pages, 5 figures
【15】Better Protein Function Prediction by Modeling Survivorship Bias
标题：通过生存偏差建模更好地预测蛋白质功能
链接：https://arxiv.org/abs/2605.06879
备注：29 pages, 12 figures, 3 tables
【16】Information-theoretic Limits of Learning and Estimation
标题：学习和估计的信息论限制
链接：https://arxiv.org/abs/2605.06710
【17】TRACE: Transport Alignment Conformal Prediction via Diffusion and Flow Matching Models
标题：TRACE：通过扩散和流量匹配模型的运输对齐保形预测
链接：https://arxiv.org/abs/2605.07100
备注：22 pages, 5 figures and 5 tables
【18】Physics-Based Flow Matching for Full-Field Prediction of Silicon Photonic Devices
标题：基于物理的流匹配用于硅光电器件全场预测
链接：https://arxiv.org/abs/2605.06929
备注：11 pages, 4 figures
其他神经网络|深度学习|模型|建模(59篇)
【1】Normalizing Trajectory Models
标题：轨道模型标准化
链接：https://arxiv.org/abs/2605.08078
【2】Susceptibilities and Patterning: A Primer on Linear Response in Bayesian Learning
标题：敏感性和模式化：Bayesian学习中线性反应的入门
链接：https://arxiv.org/abs/2605.07980
备注：34 pages, 3 figures, comments welcome!
【3】When Diffusion Model Can Ignore Dimension: An Entropy-Based Theory
标题：扩散模型何时可以忽略维度：一个基于熵的理论
链接：https://arxiv.org/abs/2605.07969
【4】Flatness and Gradient Alignment Are Both Necessary: Spectral-Aware Gradient-Aligned Exploration for Multi-Distribution Learning
标题：平坦性和梯度对齐都是必要的：多分布学习的频谱感知一致性对齐探索
链接：https://arxiv.org/abs/2605.07914
备注：Preprint - Submitted to NeurIPS 2026
【5】Distributional simplicity bias and effective convexity in Energy Based Models
标题：基于能量的模型中的分布简单性偏差和有效凸性
链接：https://arxiv.org/abs/2605.07844
备注：13 pages, 2 figures
【6】\mathsf{VISTA}: Decentralized Machine Learning in Adversary Dominated Environments
链接：https://arxiv.org/abs/2605.07841
【7】Toward Privileged Foundation Models:LUPI for Accelerated and Improved Learning
标题：迈向特权基础模型：LUPI加速和改进的学习
链接：https://arxiv.org/abs/2605.07799
【8】Rethinking State Tracking in Recurrent Models Through Error Control Dynamics
标题：通过错误控制动力学重新思考回归模型中的状态跟踪
链接：https://arxiv.org/abs/2605.07755
【9】Learning Large-Scale Modular Addition with an Auxiliary Modulus
标题：用辅助模学习大规模模加法
链接：https://arxiv.org/abs/2605.07648
备注：10+11 pages, 5 figures
【10】MAVEN: Multi-Agent Verification-Elaboration Network with In-Step Epistemic Auditing
标题：MAVEN：具有分步认识审计的多代理验证-生成网络
链接：https://arxiv.org/abs/2605.07646
备注：24 pages, 2 figures
【11】Learning to Communicate Locally for Large-Scale Multi-Agent Pathfinding
标题：学习本地通信以进行大规模多智能体寻路
链接：https://arxiv.org/abs/2605.07637
【12】Does Your Neural Network Extrapolate? Feature Engineering as Identifiability Bias for OOD Generalization
标题：你的神经网络会推断吗？特征工程作为OOD概括的可识别性偏差
链接：https://arxiv.org/abs/2605.07483
【13】Uncovering Hidden Systematics in Neural Network Models for High Energy Physics
标题：揭示高能物理神经网络模型中隐藏的系统学
链接：https://arxiv.org/abs/2605.07470
备注：18 pages, 9 figures
【14】Learning Minimal-Deviation Corrections for Multi-Dimensional Mismodelling in HEP Simulations
标题：HEP模拟中多维错误建模的最小偏差修正
链接：https://arxiv.org/abs/2605.07460
备注：12 pages, 6 figures
【15】VNN-LIB 2.0: Rigorous Foundations for Neural Network Verification
标题：VNN-LIB 2.0：神经网络验证的严格基础
链接：https://arxiv.org/abs/2605.07451
【16】Effective and Memory-Efficient Alternatives to ECC for Reliable Large-Scale DNNs
标题：用于可靠大规模DNN的有效且内存效率高的椭圆曲线替代方案
链接：https://arxiv.org/abs/2605.07417
备注：7 pages, 7 figures, 3 tables. The paper is accepted at IEEE IOLTS'26
【17】Risk-Consistent Multiclass Learning from Random Label-Subset Membership Queries
标题：来自随机标签子集成员资格表的风险一致多类学习
链接：https://arxiv.org/abs/2605.07413
【18】Tracking Large-scale Shared Bikes with Inertial Motion Learning in GNSS Blocked Environments
标题：在GNSS受阻环境中利用惯性运动学习跟踪大规模共享自行车
链接：https://arxiv.org/abs/2605.07412
备注：It has been submitted to IEEE Transactions on Intelligent Transportation Systems (T-ITS). Journal article. 14 pages, 18 figures, 10 tables
【19】CellScientist: Dual-Space Hierarchical Orchestration for Closed-Loop Refinement of Virtual Cell Models
标题：CellScientist：用于虚拟细胞模型闭环细化的双空间分层规划
链接：https://arxiv.org/abs/2605.07335
【20】Generative Modeling with Flux Matching
标题：基于通量匹配的生成式建模
链接：https://arxiv.org/abs/2605.07319
【21】Sparse Random-Feature Neural Networks with Krylov-Based SVD for Singularly Perturbed ODE
标题：奇异扰动ODE的基于Krylov奇异值分解的稀疏随机特征神经网络
链接：https://arxiv.org/abs/2605.07286
【22】Bifurcation Models: Learning Set-Valued Solution Maps with Weight-Tied Dynamics
标题：分歧模型：使用权重绑定动态学习集值解图
链接：https://arxiv.org/abs/2605.07277
【23】How Big Should a Wireless Foundation Model Be?
标题：无线基金会模型应该有多大？
链接：https://arxiv.org/abs/2605.07266
【24】Modulated learning for private and distributed regression with just a single sample per client device
标题：针对私人和分布式回归的调制学习，每个客户端设备只需一个样本
链接：https://arxiv.org/abs/2605.07233
备注：30 pages
【25】Arrow: A Foundation Model for Causal Discovery
标题：阿罗：因果关系发现的基础模型
链接：https://arxiv.org/abs/2605.07204
【26】Neurosymbolic Imitation Learning with Human Guidance: A Privileged Information Approach
标题：人类指导下的神经符号模仿学习：一种特权信息方法
链接：https://arxiv.org/abs/2605.07166
备注：Under Review for ECML-PKDD 2026
【27】Learned Lagrangian Models of PDEs via Euler-Lagrange Residual Minimization
标题：通过Euler-Lagrange剩余最小化学习的偏头痛拉格朗日模型
链接：https://arxiv.org/abs/2605.07157
备注：9 pages, 8 figures, 2 tables, 7 pages of appendices
【28】Regret-Oracle Complexity Tradeoffs in Agnostic Online Learning
标题：不可知在线学习中的遗憾-先知复杂性权衡
链接：https://arxiv.org/abs/2605.07155
【29】Query-efficient model evaluation using cached responses
标题：使用缓存响应进行查询高效模型评估
链接：https://arxiv.org/abs/2605.07096
【30】Learning Visual Feature-Based World Models via Residual Latent Action
标题：通过剩余潜在动作学习基于视觉环境的世界模型
链接：https://arxiv.org/abs/2605.07079
【31】Test-Time Compositional Generalization in Diffusion Models via Concept Discovery
标题：通过概念发现在扩散模型中进行测试时组成概括
链接：https://arxiv.org/abs/2605.07078
备注：9 pages
【32】ModelLens: Finding the Best for Your Task from Myriads of Models
标题：Model Lens：从无数的模特中寻找最适合您任务的
链接：https://arxiv.org/abs/2605.07075
【33】A Behavioral Framework for Data-Driven Modeling of Nonlinear Systems in Vector-Valued Reproducing Kernel Hilbert Spaces
标题：向量值再生核Hilbert空间中非线性系统数据驱动建模的行为框架
链接：https://arxiv.org/abs/2605.07052
备注：12 pages
【34】PACEvolve++: Improving Test-time Learning for Evolutionary Search Agents
标题：PAC Evolve ++：改善进化搜索代理的测试时学习
链接：https://arxiv.org/abs/2605.07039
【35】Learning Material-Aware Hamiltonian Risk Fields for Safe Navigation
标题：安全导航的学习材料感知汉密尔顿风险场
链接：https://arxiv.org/abs/2605.07038
【36】Equivalence of Coarse and Fine-Grained Models for Learning with Distribution Shift
标题：具有分布转移的粗粒度学习模型和细粒度学习模型的等效性
链接：https://arxiv.org/abs/2605.07005
备注：26 pages, Accepted to COLT 2026
【37】FastOmniTMAE: Parallel Clause Learning for Scalable and Hardware-Efficient Tsetlin Embeddings
标题：FastOmniTMAE：可扩展且硬件高效的Tsetlin嵌入的并行句子学习
链接：https://arxiv.org/abs/2605.06982
【38】ProtoSSL: Interpretable Prototype Learning from Unlabeled Time-Series Data
标题：ProtoSSL：从未标记的时间序列数据中进行可解释原型学习
链接：https://arxiv.org/abs/2605.06943
【39】A Generalized Singular Value Theory for Neural Networks
标题：神经网络的广义奇异值理论
链接：https://arxiv.org/abs/2605.06938
【40】Conservative Flows: A New Paradigm of Generative Models
标题：保守流：生成模型的新范式
链接：https://arxiv.org/abs/2605.06905
【41】On the Divergence of Differential Temporal Difference Learning without Local Clocks
标题：无本地时钟的差异时间差异学习的分歧
链接：https://arxiv.org/abs/2605.06874
【42】When Descent Is Too Stable: Event-Triggered Hamiltonian Learning to Optimize
标题：当下降太稳定时：事件触发的汉密尔顿学习优化
链接：https://arxiv.org/abs/2605.06868
【43】A Finite-Iteration Theory for Asynchronous Categorical Distributional Temporal-Difference Learning
标题：非同步类别分布时间差异学习的随机迭代理论
链接：https://arxiv.org/abs/2605.06866
备注：53 pages
【44】AGWM: Affordance-Grounded World Models for Environments with Compositional Prerequisites
标题：AGWM：具有成分优势的环境的以负担为基础的世界模型
链接：https://arxiv.org/abs/2605.06841
备注：16 pages, 3 figures, 4 tables. Appendix on pages 11-16 (main text is self-contained)
【45】On Privacy Leakage in Tabular Diffusion Models: Influential Factors, Attacker Knowledge, and Metrics
标题：表格扩散模型中的隐私泄露：影响因素、攻击者知识和预设
链接：https://arxiv.org/abs/2605.06835
备注：23 pages, 11 Figures, 12 Tables
【46】Attribution-Based Neuron Utility for Plasticity Restoration in Deep Networks
标题：基于属性的神经元实用程序用于深度网络中的可塑性恢复
链接：https://arxiv.org/abs/2605.06834
【47】A Unified Measure-Theoretic View of Diffusion, Score-Based, and Flow Matching Generative Models
标题：扩散、基于分数和流匹配生成模型的统一测量理论观点
链接：https://arxiv.org/abs/2605.06829
备注：62 pages, 1 figure, jmlr preprint
【48】A Rod Flow Model for Adam at the Edge of Stability
标题：Adam处于稳定边缘的杆流模型
链接：https://arxiv.org/abs/2605.06821
【49】MIND: Monge Inception Distance for Generative Models Evaluation
标题：MIND：生成模型评估的Monge初始距离
链接：https://arxiv.org/abs/2605.06797
【50】A Closed-Form Upper Bound for Admissible Learning-Rate Steps in Belief-Space Dynamics
标题：信念空间动力学中可允许学习率步骤的封闭形式上界
链接：https://arxiv.org/abs/2605.06741
【51】Geometric Kolmogorov--Arnold Network (GeoKAN)
标题：几何Kolmogorov--Arnold Network（GeoKAN）
链接：https://arxiv.org/abs/2605.06740
备注：46 pages, 24 figures, 13 tables
【52】Gated QKAN-FWP: Scalable Quantum-inspired Sequence Learning
标题：门控QKAN-FWP：可扩展的量子启发序列学习
链接：https://arxiv.org/abs/2605.06734
备注：46 pages, 13 figures, 10 tables
【53】Toeplitz MLP Mixers are Low Complexity, Information-Rich Sequence Models
标题：Toeplitz MLP混合器是低复杂性、信息丰富的序列模型
链接：https://arxiv.org/abs/2605.06683
【54】Linear Response Estimators for Singular Statistical Models
标题：奇异统计模型的线性响应估计器
链接：https://arxiv.org/abs/2605.07970
备注：24 pages, comments welcome!
【55】Characterizing and Correcting Effective Target Shift in Online Learning
标题：描述和纠正在线学习中的有效目标转变
链接：https://arxiv.org/abs/2605.07886
备注：22 pages; 6 figures
【56】Physics-Informed Reduced-Order Operator Learning for Hyperelasticity in Continuum Micromechanics
标题：连续体微力学中超弹性的物理信息降阶操作员学习
链接：https://arxiv.org/abs/2605.07738
备注：22 pages, 12 figures
【57】Every Feedforward Neural Network Definable in an o-Minimal Structure Has Finite Sample Complexity
标题：每个可定义在o-极小结构中的前向神经网络具有有限的样本复杂性
链接：https://arxiv.org/abs/2605.07097
【58】Causal EpiNets: Precision-corrected Bounds on Individual Treatment Effects using Epistemic Neural Networks
标题：因果EpiNet：使用认知神经网络精确修正个体治疗效果的界限
链接：https://arxiv.org/abs/2605.07065
【59】Kernel Selection is Model Selection: A Unified Complexity-Penalized Approach for MMD Two-Sample Tests
标题：核选择就是模型选择：MMD两样本测试的统一复杂性惩罚方法
链接：https://arxiv.org/abs/2605.06883
其他(76篇)
【1】Don't Get Your Kroneckers in a Twist: Gaussian Processes on High-Dimensional Incomplete Grids
标题：不要让你的克罗内克斯扭曲：多维不完整网格上的高斯过程
链接：https://arxiv.org/abs/2605.08036
备注：51 pages, 8 figures
【2】INO-SGD: Addressing Utility Imbalance under Individualized Differential Privacy
标题：INO-Singapore：解决个性化差异隐私下的公用事业失衡
链接：https://arxiv.org/abs/2605.07930
备注：Accepted to the 14th International Conference on Learning Representations (ICLR-26)
【3】Curvature Beyond Positivity: Greedy Guarantees for Arbitrary Submodular Functions
标题：超越正性的弯曲：任意次模函数的贪婪保证
链接：https://arxiv.org/abs/2605.07902
备注：44 pages, 11 figures
【4】Actor-Critic Algorithm for Dynamic Expectile and CVaR
标题：动态预期和CVaR的行为者批评算法
链接：https://arxiv.org/abs/2605.07857
【5】Approximation-Free Differentiable Oblique Decision Trees
标题：无逼近可微斜决策树
链接：https://arxiv.org/abs/2605.07837
备注：Accepted for publication in JMLR, Vol. 27, 2026
【6】Scaling Categorical Flow Maps
标题：缩放类别流程图
链接：https://arxiv.org/abs/2605.07820
【7】The Minimax Rate of Second-Order Calibration
标题：二阶校准的极小极大率
链接：https://arxiv.org/abs/2605.07808
【8】Text-to-CAD Evaluation with CADTests
标题：使用CADTests进行文本到CAD评估
链接：https://arxiv.org/abs/2605.07807
【9】Neural Operators as Efficient Function Interpolators
标题：神经算子作为有效的函数插值器
链接：https://arxiv.org/abs/2605.07792
备注：12 pages, 9 figures
【10】Efficient Verification of Neural Control Barrier Functions with Smooth Nonlinear Activations
标题：具有光滑非线性激活的神经控制屏障函数的有效验证
链接：https://arxiv.org/abs/2605.07757
备注：9 pages, 4 figures
【11】When Losses Align: Gradient-Based Composite Loss Weighting for Efficient Pretraining
标题：当损失对齐时：基于对象的复合损失加权以实现有效的预训练
链接：https://arxiv.org/abs/2605.07756
【12】Intelligent Truck Matching in Full Truckload Shipments using Ping2Hex approach
标题：使用Ping 2 Hex方法在满载货物中进行智能卡车匹配
链接：https://arxiv.org/abs/2605.07733
备注：12 pages, 10 figures, 8 tables. Accepted at iSCSi 2026 (International Conference on Industry Sciences and Computer Sciences Innovation). To appear in Procedia Computer Science (Elsevier)
【13】Drifting Field Policy: A One-Step Generative Policy via Wasserstein Gradient Flow
标题：漂田政策：通过沃瑟斯坦梯度流的一步生成政策
链接：https://arxiv.org/abs/2605.07727
【14】Curated Synthetic Data Doesn't Have to Collapse: A Theoretical Study of Generative Retraining with Pluralistic Preferences
标题：精心策划的合成数据不必崩溃：具有多元偏好的生成性再训练的理论研究
链接：https://arxiv.org/abs/2605.07724
备注：Proceedings of the 43rd International Conference on Machine Learning, Seoul, South Korea. PMLR 306, 2026
【15】Bayesian Fine-tuning in Projected Subspaces
标题：投影子空间中的贝叶斯微调
链接：https://arxiv.org/abs/2605.07706
【16】Gradient Starvation in Binary-Reward GRPO: Why Group-Mean Centering Fails and Why the Simplest Fix Works
标题：二进制奖励GRPO中的梯度饥饿：为什么群体平均中心失败以及为什么最简单的修复有效
链接：https://arxiv.org/abs/2605.07689
【17】The Coupling Tax: How Shared Token Budgets Undermine Visible Chain-of-Thought Under Fixed Output Limits
标题：耦合税：共享代币预算如何破坏固定产出限制下可见的思想链
链接：https://arxiv.org/abs/2605.07686
备注：40 pages, 6 figures
【18】Structured Coupling for Flow Matching
标题：用于流量匹配的结构化耦合
链接：https://arxiv.org/abs/2605.07676
【19】Differentially Private Auditing Under Strategic Response
标题：战略应对下的差异私人审计
链接：https://arxiv.org/abs/2605.07674
【20】Quotient Semivalues for False-Name-Resistant Data Attribution
标题：抗假名数据归因的商半值
链接：https://arxiv.org/abs/2605.07663
【21】Safe, or Simply Incapable? Rethinking Safety Evaluation for Phone-Use Agents
标题：安全，还是根本无能为力？重新思考电话使用代理的安全性评估
链接：https://arxiv.org/abs/2605.07630
备注：work in progress
【22】Ensemble Distributionally Robust Bayesian Optimisation
标题：引入分布稳健的Bayesian优化
链接：https://arxiv.org/abs/2605.07565
【23】On the Invariance and Generality of Neural Scaling Laws
标题：神经尺度定律的不变性和普遍性
链接：https://arxiv.org/abs/2605.07546
备注：23 pages, 6 figures, 11 tables
【24】Tessellations of Semi-Discrete Flow Matching
标题：半离散流匹配的网格化
链接：https://arxiv.org/abs/2605.07513
【25】Excluding the Target Domain Improves Extrapolation: Deconfounded Hierarchical Physics Constraints
标题：排除目标域改进了外推：揭穿分层物理约束
链接：https://arxiv.org/abs/2605.07485
备注：16 pages, 2 figures
【26】Physical Simulators as Do-Operators: Causal Discovery under Latent Confounders for AI-for-Science
标题：物理模拟器作为操作员：人工智能科学潜在混淆下的因果发现
链接：https://arxiv.org/abs/2605.07467
备注：17 pages, 1 figure
【27】The Proxy Presumption: From Semantic Embeddings to Valid Social Measures
标题：代理推定：从语义嵌入到有效的社会措施
链接：https://arxiv.org/abs/2605.07409
备注：ACL 2026
【28】Exploring CoCo Challenges in ML Engineering Teams: Insights From the Semiconductor Industry
标题：探索机器学习工程团队中的CoCo挑战：来自半导体行业的见解
链接：https://arxiv.org/abs/2605.07389
【29】QuadNorm: Resolution-Robust Normalization for Neural Operators
标题：QuadNorm：神经运算符的分辨率稳健规范化
链接：https://arxiv.org/abs/2605.07375
备注：42 pages, 8 figures
【30】Mean-Pooled Cosine Similarity is Not Length-Invariant: Theory and Cross-Domain Evidence for a Length-Invariant Alternative
标题：均值合并的Cosine相似性不是长度不变的：长度不变替代方案的理论和跨领域证据
链接：https://arxiv.org/abs/2605.07345
备注：9 pages, 6 figures. Submitted to the Mechanistic Interpretability Workshop at ICML 2026
【31】SparseRL-Sync: Lossless Weight Synchronization with ~100x Less Communication
标题：SparseRL-同步：无损重量同步，通信量减少约100倍
链接：https://arxiv.org/abs/2605.07330
备注：Code will be released at https://github.com/scitix/helix
【32】Activation Differences Reveal Backdoors: A Comparison of SAE Architectures
标题：激活差异揭示后门：SAP架构的比较
链接：https://arxiv.org/abs/2605.07324
备注：Accepted at IJCNN 2026 (IEEE WCCI). ©2026 IEEE
【33】Latent Order Bandits
标题：潜秩序盗贼
链接：https://arxiv.org/abs/2605.07304
备注：arXiv admin note: text overlap with arXiv:2508.05367
【34】Instruction Tuning Changes How Upstream State Conditions Late Readout: A Cross-Patching Diagnostic
标题：指令调优改变上游状态条件延迟读出的方式：交叉修补诊断
链接：https://arxiv.org/abs/2605.07284
【35】Mask2Cause: Causal Discovery via Adjacency Constrained Causal Attention
标题：Mask2 Cause：通过邻近约束因果注意力发现因果
链接：https://arxiv.org/abs/2605.07280
【36】bispectrum: Selective $G$-Bispectra Made Practical
标题：双谱：选择性$G$-双谱变得实用
链接：https://arxiv.org/abs/2605.07270
【37】On the Robustness of Distribution Support under Diffusion Guidance
标题：扩散引导下分配支持的稳健性
链接：https://arxiv.org/abs/2605.07220
【38】Cost-Ordered Feasibility for Multi-Armed Bandits with Cost Subsidy
标题：多臂强盗的成本订购可行性，并提供成本补贴
链接：https://arxiv.org/abs/2605.07171
【39】Topic Is Not Agenda: A Citation-Community Audit of Text Embeddings
标题：主题不是议程：文本嵌入的引用社区审计
链接：https://arxiv.org/abs/2605.07158
备注：16 pages, 4 figures, 4 tables
【40】MathlibPR: Pull Request Merge-Readiness Benchmark for Formal Mathematical Libraries
标题：MathlibPR：正式数学库的拉式请求合并准备基准
链接：https://arxiv.org/abs/2605.07147
【41】Conformal-Style Quantile Analyses for Stochastic Bandits
标题：随机盗贼的保形分位数分析
链接：https://arxiv.org/abs/2605.07115
【42】CarCrashNet: A Large-Scale Dataset and Hierarchical Neural Solver for Data-Driven Structural Crash Simulation
标题：CarCrashNet：用于数据驱动结构碰撞模拟的大规模数据集和分层神经解算器
链接：https://arxiv.org/abs/2605.07098
【43】Actor-Critic with Active Importance Sampling
标题：积极重要性抽样的演员评论家
链接：https://arxiv.org/abs/2605.07094
【44】The Translation Tax Is Not a Scalar: A Counterfactual Audit of English-Source Cue Inheritance in Chinese Multilingual Benchmarks
标题：翻译税不是一个纯量：对中文多语言基准中英文源线索继承的反事实审计
链接：https://arxiv.org/abs/2605.07093
备注：13 pages, 3 figures. Submitted to NeurIPS 2026
【45】Task Relevance Is Not Local Replaceability: A Two-Axis View of Channel Information
标题：任务相关性不是本地可替换性：渠道信息的两轴视图
链接：https://arxiv.org/abs/2605.07086
【46】PolarAdamW: Disentangling Spectral Control and Schur Gauge-Equivariance in Matrix Optimisation
标题：PolarAdamW：解开矩阵优化中的谱控制和舒尔规范等方差
链接：https://arxiv.org/abs/2605.07067
【47】Integrating Causal DAGs in Deep RL: Activating Minimal Markovian States with Multi-Order Exposure
标题：在深度RL中集成因果DAB：用多阶暴露激活最小马尔科夫状态
链接：https://arxiv.org/abs/2605.07057
【48】The Context Gathering Decision Process: A POMDP Framework for Agentic Search
标题：上下文收集决策过程：用于统计搜索的POMDP框架
链接：https://arxiv.org/abs/2605.07042
备注：25 pages
【49】Echo: KV-Cache-Free Associative Recall with Spectral Koopman Operators
标题：Echo：带谱Koopman运算符的无NV缓存联想召回
链接：https://arxiv.org/abs/2605.06997
【50】Why Does Agentic Safety Fail to Generalize Across Tasks?
标题：为什么大型安全性未能在各个任务中普遍化？
链接：https://arxiv.org/abs/2605.06992
【51】Response Time Enhances Alignment with Heterogeneous Preferences
标题：响应时间增强与异类偏好的一致性
链接：https://arxiv.org/abs/2605.06987
【52】$f$-Divergence Regularized RLHF: Two Tales of Sampling and Unified Analyses
标题：$f$-分歧正规化的RL HF：抽样和统一分析的两个故事
链接：https://arxiv.org/abs/2605.06977
备注：ICML 2026
【53】MAGIQ: A Post-Quantum Multi-Agentic AI Governance System with Provable Security
标题：MAGIQ：具有可证明安全性的后量子多统计人工智能治理系统
链接：https://arxiv.org/abs/2605.06933
【54】In-Context Credit Assignment via the Core
标题：通过核心进行上下文信用分配
链接：https://arxiv.org/abs/2605.06920
【55】TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond
标题：TraXion：重新思考移动性及其他领域的预训练框架
链接：https://arxiv.org/abs/2605.06906
备注：31 pages, 2 figures
【56】Accelerated Relax-and-Round for Concave Coverage Problems
标题：凹凸覆盖问题的加速放松和循环
链接：https://arxiv.org/abs/2605.06900
备注：47 pages, 6 figures
【57】Towards Fairness under Label Bias in Image Segmentation: Impact, Measurement and Mitigation
标题：图像分割中标签偏差下的公平性：影响、测量和缓解
链接：https://arxiv.org/abs/2605.06891
【58】Continuous First, Discrete Later: VQ-VAEs Without Dimensional Collapse
标题：连续优先，离散后：VQ-VAE没有维度崩溃
链接：https://arxiv.org/abs/2605.06870
【59】Benchmarked Yet Not Measured -- Generative AI Should be Evaluated Against Real-World Utility
标题：已达到基准但未衡量--生成人工智能应根据现实世界的实用性进行评估
链接：https://arxiv.org/abs/2605.06856
备注：20 pages
【60】SHARP: A Self-Evolving Human-Auditable Rubric Policy for Financial Trading Agents
标题：SHARP：针对金融交易代理的自我进化的、可审核的主题政策
链接：https://arxiv.org/abs/2605.06822
【61】Conformal Agent Error Attribution
标题：保形代理错误归因
链接：https://arxiv.org/abs/2605.06788
备注：10 pages
【62】Weblica: Scalable and Reproducible Training Environments for Visual Web Agents
标题：Weblica：用于视觉Web代理的可扩展和可复制的训练环境
链接：https://arxiv.org/abs/2605.06761
备注：28 pages, 19 figures
【63】On Training in Imagination
标题：论想象力的训练
链接：https://arxiv.org/abs/2605.06732
【64】When Routine Chats Turn Toxic: Unintended Long-Term State Poisoning in Personalized Agents
标题：当例行聊天变得有毒时：个性化代理人意外的长期状态中毒
链接：https://arxiv.org/abs/2605.06731
备注：23 pages
【65】Robustness of Refugee-Matching Gains to Off-Policy Evaluation Choices
标题：难民匹配收益对政策外评估选择的稳健性
链接：https://arxiv.org/abs/2605.06686
备注：13 pages, 2 figures, 10 tables
【66】On the Role of Strain and Vorticity in Numerical Integration Error for Flow Matching
标题：应变和涡度在流量匹配数值积分误差中的作用
链接：https://arxiv.org/abs/2605.06680
备注：16 pages, 7 figures. Preliminary version. Includes qualitative CIFAR-10 comparison and supporting synthetic experiments
【67】Breaking the Illusion: When Positive Meets Negative in Multimodal Decoding
标题：打破幻觉：当多模式解码中积极遇到消极时
链接：https://arxiv.org/abs/2605.06679
备注：Accepted by CVPR 2026 (Conference on Computer Vision and Pattern Recognition). 11 pages, 5 figures. Code available at: https://github.com/JiangYubo4399/PND
【68】A Note on Non-Negative $L_1$-Approximating Polynomials
标题：关于非负$L_1$-逼近型多项的注记
链接：https://arxiv.org/abs/2605.08072
【69】Semiparametric Efficient Test for Interpretable Distributional Treatment Effects
标题：可解释分布治疗效果的半参数有效检验
链接：https://arxiv.org/abs/2605.08034
【70】Consistency Regularised Gradient Flows for Inverse Problems
标题：反问题的一致性正规化梯度流
链接：https://arxiv.org/abs/2605.07907
【71】Flow Matching for Count Data
标题：计数数据的流匹配
链接：https://arxiv.org/abs/2605.07746
【72】Reliable Chain-of-Thought via Prefix Consistency
标题：通过后缀一致性实现可靠的思想链
链接：https://arxiv.org/abs/2605.07654
备注：See our project page at https://naoto-iwase.github.io/prefix-consistency-page
【73】Robust stochastic first order methods in heavy-tailed noise via medoid mini-batch gradient sampling
标题：重尾噪音中的鲁棒随机一阶方法中小批量梯度采样
链接：https://arxiv.org/abs/2605.07634
【74】Functional-prior-based Bayesian PDE-constrained inversion using PINNs
标题：基于函数先验的贝叶斯偏微分方程约束反演
链接：https://arxiv.org/abs/2605.07060
【75】Muon with Nesterov Momentum: Heavy-Tailed Noise and (Randomized) Inexact Polar Decomposition
标题：具有Nesterov动量的μ子：重尾噪声和（随机）不精确极分解
链接：https://arxiv.org/abs/2605.06884
备注：33 pages, 4 figures, 1 table
【76】One Operator for Many Densities: Amortized Approximation of Conditioning by Neural Operators
标题：多种密度的一个操作员：神经操作员条件反射的摊销逼近
链接：https://arxiv.org/abs/2605.06873
机器翻译由腾讯交互翻译提供，仅供参考
点击“阅读原文”获取带摘要的学术速递
Python社区是高质量的Python/Django开发社区
本文地址：http://www.python88.com/topic/196164
本文地址：http://www.python88.com/topic/196164