|
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437篇
大模型相关(72篇)
【1】Squeezing Capacity from Multimodal Large Language Models for Subject-driven Generation
标题：从多模式大型语言模型中挤压容量以实现主题驱动生成
链接：https://arxiv.org/abs/2605.26111
备注：33 pages, 18 figures, Project Page: https://zsh2000.github.io/squeeze-mllm-subject-gen/
摘要：Subject-driven image generation aims to synthesize new images that preserve the identity of the given subject while following textual instructions. Existing approaches often encode text and reference images separately. This limits cross-modal reasoning abilities and causes copy-paste artifacts. Recent frameworks that connect multimodal models and diffusion models improve instruction following, but largely overlook identity preservation. To address these limitations, we condition diffusion models on Multimodal Large Language Models (MLLMs) that jointly encode text and reference images, and augment it with VAE-based identity conditioning. A novel Dual Layer Aggregation (DLA) module is designed to aggregate multi-level MLLM features for optimal conditioning, and a multi-stage denoising strategy is applied to progressively balance the semantic information from MLLM and fine-detail identity from VAE during inference. Extensive experiments demonstrate that our approach harmonizes multimodal understanding with identity preservation, mitigates copy-paste issues, and achieves superior performance regarding human preference on subject-driven image generation. Our project website is available at https://zsh2000.github.io/squeeze-mllm-subject-gen/.
【2】Looped Diffusion Language Models
标题：循环扩散语言模型
链接：https://arxiv.org/abs/2605.26106
备注：23 pages
摘要：Masked diffusion models (MDMs) have emerged as a promising alternative to autoregressive models for language modeling, yet the effective design of transformer architectures for MDMs remains underexplored. In this paper, we show that selectively looping the early-middle transformer layers significantly improves both training efficiency and model performance in MDMs. We call this approach LoopMDM(Looped Masked Diffusion Model), which brings two key benefits: looping layers at training-time yields a depth-scaling effect without adding parameters, while varying the number of loops at inference-time enables flexible compute scaling. Despite the simplicity, the results are striking: across multiple pre-training corpora, LoopMDM matches the performance of same-size MDMs with up to 3.3 fewer training FLOPs, while its final performance outperforms them on various reasoning benchmarks, including up to 8.5 points on GSM8K. It even surpasses deeper non-looped MDMs trained with comparable per-step compute, indicating that selective looping is more effective than naive depth scaling. Furthermore, LoopMDM can scale inference-time compute by increasing the number of loops. Adaptively adjusting the number of loops throughout the sampling process further yields additional gains in compute efficiency while maintaining performance. Lastly, with attention analysis, we provide evidence that looping is effective in MDMs by promoting interactions among masked positions. Our code and weights will be publicly released.
【3】Forgetting in Language Models: Capacity, Optimization, and Self-Generated Replay
标题：语言模型中的遗忘：容量、优化和自我生成的回放
链接：https://arxiv.org/abs/2605.26097
摘要：Models trained on a new task typically degrade on prior tasks, a phenomenon known as forgetting. Traditionally, mitigating forgetting has required replaying stored exemplars from prior tasks, which is often impractical. By contrast, language models can sample from their own training distribution, and we show that these self-generated samples serve as effective replay data, nearly eliminating forgetting. We find that forgetting nonetheless persists when the model has little remaining capacity: models pretrained close to saturation cannot absorb new information without overwriting prior knowledge. When capacity is not the limiting factor, low learning rates reduce forgetting but require substantially more training steps. Replay breaks this tradeoff, enabling fast, high-learning-rate finetuning without forgetting.
【4】When Gradients Collide: Failure Modes of Multi-Objective Prompt Optimization for LLM Judges
标题：当学生碰撞时：LLM评委多目标即时优化的失败模式
链接：https://arxiv.org/abs/2605.26046
备注：Accepted at ACL 2026 CustomNLP4U Workshop. Code, prompts and data available at https://github.com/ARDivekar/PromptMOO
摘要 ：Customizing an LLM judge to a specific task or domain often involves optimizing its prompt across multiple evaluation criteria simultaneously. Textual gradient methods automate this for a single judge criterion, however they produce natural-language critiques, not numerical vectors. Thus, the conflict-resolution toolkit of multi-task learning (PCGrad, MGDA) doesn't apply to the multi-objective textual gradient setting. We test five decomposition modes of textual gradient optimizers by varying how much cross-task information the loss, gradient and optimizer LLMs share. In 6 of 10 configurations, we observe that optimization never improves over the initial prompt. Gradient specificity drops by 59% (from 9.0 to 3.7) when the gradient LLM processes multiple criteria jointly. Separately, we observe that naively combining per-task instructions into a single prompt degrades Spearman's rho by -5.3%. These results identify two separable failure modes: optimization-time gradient dilution and inference-time instruction interference, which together constrain the design space for multi-objective judge customization using textual feedback.
【5】Causal methods for LLM development and evaluation
标题：LLM开发和评估的因果方法
链接：https://arxiv.org/abs/2605.25998
备注：Published in KDD 2026
摘要：Large language model (LLM) development is currently driven by large-scale empirical iteration over data mixtures, reward models, routing strategies, and evaluation pipelines. Here, we argue that many central questions in LLM development and evaluation are inherently causal: What is the effect of adding a data domain during pretraining? How do annotator preferences change when LLMs generate text in a different style? Should a prompt be routed to a larger or smaller model given inference cost constraints? In general, causal methods are well-suited to such settings where interventions change outcomes but, surprisingly, are underrepresented in LLM development. Our contribution is threefold: (1) We explain how causal methods can help develop modern LLM development and evaluation: LLM development relies heavily on logged data, which are often subject to confounding and distribution shifts; evaluation uses learned but potentially biased judges; and deployment environments are non-stationary. These conditions make purely predictive approaches fragile and create opportunities for principled identification and estimation methods from causal inference. (2) We further map opportunities for causal methods in the entire LLM development pipeline, including pretraining, alignment, routing, agentic workflows, and evaluation. (3) We discuss new research opportunities around leveraging causal methods for LLM development and evaluation. Overall, we argue that causal methods are potentially underutilized for the LLM development and evaluation pipeline, despite the fact that such methods can ensure a reliable and scientifically grounded design.
【6】QUIET: A Multi-Blank Cascaded Story Cloze Benchmark for LLM Creative Generation Capability
标题：QUIET：LLM创意生成能力的多空白级联故事完形基准
链接：https://arxiv.org/abs/2605.25955
摘要：Large language models (LLMs) face a dual challenge in creative capability evaluation: existing benchmarks (e.g., Story Cloze Test, HellaSwag) measure models' discriminative ability over narrative continuation using multiple-choice recognition paradigms, rather than directly measuring creative generation capability; rubric-based scoring and LLM-as-Judge methods rely on subjective dimension assessment or natural language model outputs, and cannot provide objective, automated scoring mechanisms. This paper proposes QUIET (Quality Understanding via Interlocked Evaluation Testing), a diagnostic benchmark for LLM creative capability based on multi-blank cascaded story cloze. QUIET sets N blanks (10-20) in a story with complete structure, with each blank accompanied by an explicit content constraint, and cascade dependency relationships between blanks -- the content filled into earlier blanks constrains the feasible solution space for later blanks. The evaluated model (or human participants) fills all blanks in open-ended generation mode; the results are scored by an information-theoretic automated scoring protocol without human grading. The scoring protocol directly operationalizes the "calibrated surprise" theoretical framework (Zou & Xu, 2026a). For each blank k, a composite score is computed: score = satisfy * (1 + lambda * surprise), where lambda = 1.0. Here, "satisfy" measures how well the blank filling satisfies the content constraint (objective logical reasoning judgment, not subjective aesthetic scoring), and "surprise" measures the degree of surprise given that the constraint is satisfied. Creative answers that do not satisfy the constraint score zero; answers that satisfy the constraint but are mediocre score low; answers that satisfy the constraint and are surprising score high.
【7】Step-TP: A Grounded, Step-Level Dataset with Chain-of-Thought Reasoning for LLM-Guided Tensor Program Optimization
标题：Step-TP：具有思想链推理的接地、分步数据集，用于LLM引导的张量程序优化
链接：https://arxiv.org/abs/2605.25954
摘要：Despite the strong reasoning capabilities of large language models (LLMs), optimizing the execution efficiency of tensor programs remains challenging due to the need for precise, composable transformation decisions. Recent LLM-guided approaches frame tensor program optimization as an iterative decision process, but existing datasets provide only end-to-end optimized program pairs using token-inefficient representations, lacking verifiable step-level supervision and interpretability. As a result, LLMs struggle to make reliable single-step decisions in large combinatorial optimization spaces. We introduce Step-TP, a post-training dataset for tensor program optimization that provides grounded, atomic, step-level supervision with structured chain-of-thought (CoT) reasoning. Step-TP forms a closed reasoning loop over intermediate program states, enabling reliable multi-step optimization rather than outcome imitation. Its design is guided by four principles: (i) a token-efficient, verifiable intermediate representation (IR) that deterministically lowers to TVM TIR; (ii) atomic and composable optimization strategies that decompose complex trajectories into interpretable single-step decisions; (iii) structured CoT supervision coupled with explicit IR-to-IR state transitions; and (iv) strategy filtering to balance coverage while preventing shortcut exploitation. The dataset and implementation are available at a GitHub link, https://github.com/LIUMENGFAN-gif/StepTP.
【8】Merge-Bench: Resolve Merge Conflicts with Large Language Models
标题：合并长凳：解决与大型语言模型的合并冲突
链接：https://arxiv.org/abs/2605.25890
备注：14 pages, 7 figures
摘要 ：This paper applies machine learning to the difficult and important task of version control merging. (1) We constructed a dataset, Merge-Bench, of 7938 real-world merge conflict hunks from 1439 GitHub repositories. The ground truth is the merge resolution that developers committed to the repository. Our dataset construction methodology is scalable to arbitrary amounts of data since no manual labeling is required. (2) We trained a model, LLMergeJ, to resolve merge conflicts in Java programs. Our approach uses Group Relative Policy Optimization (GRPO), an online reinforcement learning method, to train a Large Language Model (LLM). (3) We performed two evaluations of the performance of LLMs on resolving merge conflicts. On Java programs, LLMergeJ with 14B parameters outperforms 3 commercial LLMs, trailing only Gemini 2.5 Pro. Across 11 programming languages, commercial LLM performance is largely stable from language to language. The best models correctly resolve less than 60% of merge conflicts.
【9】Capability and Robustness Cannot Both Be Free: An Information-Theoretic Bound for Vision-Language-Action Models
标题：能力和稳健性不能同时自由：视觉-语言-动作模型的信息理论界限
链接：https://arxiv.org/abs/2605.25889
摘要：Vision-Language-Action (VLA) models are increasingly deployed on real robots, where each predicted action is executed and each failure carries a safety cost. They reach high success rates on clean inputs but collapse under small adversarial perturbations. A $16/255$ PGD attack on OpenVLA-7B drops LIBERO success from above $95\%$ to under $5\%$. Empirical defenses recover some robustness at a cost in clean accuracy, but the literature does not say whether the trade-off has a theoretical floor. We prove that it does. For any VLA policy with discrete actions, the sum of capability (mutual information between policy action and oracle action) and robustness (mutual information preserved under adversarial perturbation, net of trivial channel leakage) is upper-bounded by a policy-independent budget: task entropy plus adversarial channel capacity. The proof is two applications of the Data Processing Inequality plus MI non-negativity. The pixel-level bound is loose on current models ($\sim 10^3$ nats), but an encoder-specific corollary restricts the channel to the policy-relevant subspace, reducing the budget from $\sim 5{,}000$ to $\sim 31$ nats on OpenVLA; the policy already consumes $\sim 24\%$ of this tighter budget, leaving limited room for simultaneous robustness improvement. We validate the bound across $252$ closed-form Gaussian-VLA cells and $48$ OpenVLA-7B $\times$ LIBERO $\times$ PGD cells (zero violations). We propose encoder-specific slack as a normalized comparison axis for defense papers, and release all code, manifests, and results.
【10】TIAR: Trajectory-Informed Advantage Reweighting for LLM Abstention Learning
标题：TIAR：LLM弃权学习的轨迹知情优势重新加权
链接：https://arxiv.org/abs/2605.25850
备注：10 pages, 1 figure, 4 tables
摘要：This paper investigates large language model (LLM) abstention learning, specifically using ternary reward, which incentivize truthfulness in large language models. This paper extends that idea by moving from a ternary reward to a Trajectory-Informed advantage reweighting, dynamically re-weights the abstention reward during Group Relative Policy Optimization (GRPO) training. The objective of this work focuses on abstention learning instead of improving truthfulness, serving as an exploration into hallucination reduction. The novelty of this paper lies in methodological innovation, advantage re-weighting, and benchmark selection. Leveraging GRPO's multiple trajectories as a natural abstention signal, this method uses a reward signal to explore knowledge boundaries and encourage consistency. By demonstrating that trajectories can be used as a confidence indicator of the policy relative to the query, they are then used to dynamically calculate the abstention advantage. AbstentionBench is used as the evaluation benchmark, as this work aims to contribute to the field of abstention learning. All datasets on the benchmark were tested against this method and various baselines. Empirical results demonstrate that TIAR achieves state-of-the-art abstention F1 scores across five of six evaluation categories, outperforming the static ternary baseline on 17 of 31 benchmark datasets while fully preserving baseline accuracy.
【11】Visual-Redundancy-Controlled Parallel Decoding for Diffusion-Based Multimodal Large Language Models
标题：基于扩散的多模式大型语言模型的视觉冗余控制并行解码
链接：https://arxiv.org/abs/2605.25820
备注：18 pages, 5 figures
摘要：Diffusion-based multimodal large language models (dMLLMs) decode by iteratively predicting tokens at multiple masked positions in parallel. This turns each decoding step into a position-selection problem: the model must choose not only which predictions are reliable in isolation, but also which positions should be committed together as context for later decoding steps. Existing confidence-based decoding ranks masked positions independently and commits the top-K positions, largely ignoring whether the committed tokens provide complementary visual grounding. We identify a step-level limitation of this strategy in multimodal settings: high-confidence tokens selected in the same step can rely on overlapping visual grounding, introducing visual redundancy among the committed tokens and leaving less complementary visual grounding available for later decoding. To quantify this effect, we introduce the Visual Redundancy Index (VRI), which measures visual grounding overlap among tokens committed in parallel. To control this redundancy during decoding, we propose Visual-Redundancy-Controlled Decoding (VRCD), a training-free inference-time decoding method that uses token-to-image attention to prioritize visually complementary positions. Across diverse multimodal benchmarks, VRCD reduces visual redundancy and remaining-position entropy with modest runtime overhead. In longer decoding experiments, it also achieves relative accuracy gains of up to 18.8% on M^3CoT and 6.9% on MMBench over confidence-based decoding. Code will be released at https://github.com/infiniteYuanyl/VRCD.
【12】PowLU: An Activation Function for Stable Pre-Training of LLMs
标题：PowLU：LLM稳定预训练的激活功能
链接：https://arxiv.org/abs/2605.25704
备注：17 pages, 7 figures, techreport
摘要 ：In contemporary large language models (LLMs), the swish-gated linear unit (SwiGLU) activation function is widely adopted to regulate the information flow and introduce non-linearity. For large positive inputs, SwiGLU approximates the quadratic function $x^2$, providing strong nonlinearity and expressive capacity. However, this property also causes numerical instability as the input or model scale increases, particularly in low-precision LLM training. The main reason is its approximate quadratic amplification, which enlarges the output range and exacerbates outliers. To address this issue, we propose a stable activation function, Power Linear Unit (PowLU), for large-scale LLM pre-training. Specifically, PowLU employs a rational power function to achieve adaptive nonlinearity, thereby improving representation ability and enabling stable training in spike regions. Moreover, we provide theoretical justification for several key properties of PowLU. Scaling law experiments confirm that the performance is consistent across model sizes, and further experimental results with the Ling architecture (7.9B and 124B total parameters) demonstrate that PowLU achieves competitive results against SwiGLU and SwiGLU-Clip in large-scale training of LLMs. In addition, the experimental results also show that PowLU effectively improves the scalability of the large-scale training of LLMs.
【13】How Should LLMs Consume High-Quality Data? Optimal Data Scheduling via Quality-Aware Functional Scaling Laws
标题：LLM应如何保存高质量数据？通过质量感知功能缩放定律进行最佳数据调度
链接：https://arxiv.org/abs/2605.25698
摘要：High-quality data is scarce in large language model (LLM) training, yet how to schedule its use jointly with training dynamics lacks theoretical guidance. We extend functional scaling laws by incorporating a data-quality dimension, and solve the joint data-quality and batch-size scheduling problem in asymptotic closed form. The solution reveals two regimes and a dual role of high-quality data. In the noise-limited regime, high-quality data should be used as a signal amplifier: lowering the batch size converts cleaner data into more signal without amplifying noise. In the signal-limited regime, it should be used as a noise suppressor: late placement reduces terminal noise without sacrificing signal accumulation. Existing curriculum-style pipelines primarily exploit the second role by placing cleaner data late, but miss the first role because conventional decay schedules reduce update intensity exactly when high-quality data becomes available. Guided by this, we propose Drop-Stable-Rampup for LLM midtraining: upon the quality transition, drop the batch size, hold it stable to accumulate signal, then ramp up to suppress terminal noise. On a 15B Mixture-of-Experts model midtrained on 108B tokens, Drop-Stable-Rampup improves average accuracy over Warmup-Stable-Decay (WSD) by +1.70 and over Cosine-decay by +2.98, with particularly large gains on mathematical reasoning benchmarks such as GSM8K (+4.23) and MATH (+2.80).
【14】SAE-FD: Sparse Autoencoder Feature Distillation for Continual Learning of Large Language Models
标题：SAE-FD：用于大型语言模型连续学习的稀疏自动编码器特征提取
链接：https://arxiv.org/abs/2605.25525
摘要：Continual learning enables large language models to adapt to evolving tasks without retraining from scratch, yet catastrophic forgetting remains a central obstacle. Among continual learning methods, regularization-based approaches are widely used to constrain model updates and reduce forgetting, operating in weight space, gradient space, or output space. However, these dense representation spaces suffer from feature superposition, where multiple concepts are encoded in overlapping dimensions, making it difficult to selectively protect previously learned knowledge without impeding new-task learning. To address this issue, we propose \method (Sparse Autoencoder Feature Distillation), which anchors model representations in the sparse feature space of a pre-trained Sparse Autoencoder, where dense activations are decomposed into a sparse overcomplete basis that reduces representational entanglement, enabling more targeted regularization with less interference to new-task learning. Experiments on two continual learning benchmarks across three model architectures show that \method consistently outperforms existing regularization-based methods, achieving up to 52.70% average accuracy with only -0.46 backward transfer.
【15】From Simulation to Enaction: Post-trained language models recognize and react to their own generations
标题：从模拟到制定：后训练的语言模型识别自己的世代并对其做出反应
链接：https://arxiv.org/abs/2605.25459
备注：Anthropic fellows project mentored by Jack Lindsey
摘要：Language models are pretrained as passive predictors with no incentive to model the consequences of their own outputs. Post-training changes this: a model producing its own responses can benefit from recognizing that it is on-policy. We present evidence that post-trained models recognize their on-policy generations, and this recognition is implicitly encoded in their output distributions. In particular, on-policy output distribution entropy is 3--4$\times$ lower than off-policy entropy, across model families and size classes. We trace part of this effect to an internal representation of input surprise, tracking the unlikeliness of the most recent input token according to the model's prior predictions, that causally modulates output entropy. One example of these phenomena can be observed in response to open-ended prompts; post-trained models (unlike pretrained models) collapse their uncertainty over the topic of their upcoming response before the first output token; violating this cached intention with a different-topic prefill results in higher output entropy. We also tested whether models can distinguish on-policy contexts from prefills via explicit verbal report. We find that they can, but that interestingly, this explicit recognition routes through a different mechanism than implicit recognition.
【16】BigMac: Breaking the Pareto Frontier of Compute and Memory in Multimodal LLM Training
标题：BigMac：在多模式LLM训练中打破计算和记忆的帕累托前沿
链接：https://arxiv.org/abs/2605.25451
摘要 ：Training multimodal large language models (MLLMs) is challenged by both model and data heterogeneity. Existing systems redesign the training pipeline to address these challenges, but remain bound by a Pareto frontier between compute and memory efficiency, improving one only at the expense of the other. We present BigMac, a new training pipeline for multimodal LLMs. The core idea of BigMac is to elegantly nest the encoder and generator computation into the original LLM pipeline, forming a dependency-safe nested pipeline structure. With this design, BigMac reduces the activation memory complexity of the encoder and generator to O(1) while keeping the activation memory complexity of the LLM unchanged. At the same time, it achieves the same computational efficiency as the idealized setting with unlimited memory. As a result, BigMac breaks the Pareto frontier between computational efficiency and memory usage, enabling simultaneous optimization of both computation and memory in MLLM training. We evaluate BigMac on multiple MLLMs and training workloads. Experimental results show that BigMac achieves a 1.08$\times$-1.9$\times$ training speedup over baseline systems while maintaining stable memory usage as batch size increases.
【17】A Signal-Language Foundation Model for Broad-Spectrum Cardiovascular Assessment from Routine Electrocardiography
标题：根据常规心电图进行广谱心血管评估的信号语言基础模型
链接：https://arxiv.org/abs/2605.25446
摘要：Electrocardiography (ECG) is central to cardiovascular care, but conventional AI models are often restricted to common arrhythmias and may generalize poorly across populations or clinically subtle diseases. We developed ECG Contrastive Language-Image Pre-training (ECGCLIP), a signal-language contrastive learning framework that aligns ECG waveforms with expert diagnostic reports. ECGCLIP was pre-trained on 2,837,962 ECG studies from 1,324,856 patients and evaluated on a held-out internal test set plus nine independent external cohorts comprising about 1.5 million ECGs. Evaluation covered 89 downstream tasks, including 45 ECG diagnoses, 39 echocardiographic targets, and 5 rare cardiac diseases, using PRAUC as the primary metric. ECGCLIP consistently improved performance over random initialization and Merl-R18 baselines. On the internal test set, ECGCLIP-R34 achieved strong performance for atrial fibrillation (PRAUC 0.900) and ST-segment elevation myocardial infarction (PRAUC 0.383), with robust generalization across all external cohorts. It also improved low-prevalence and diagnostically elusive diseases, including Ebstein anomaly, constrictive pericarditis, dextrocardia, and cardiac amyloidosis, with internal PRAUC values of 0.253, 0.175, 0.121, and 0.201, respectively. ECGCLIP was data efficient, matching or exceeding full-dataset baseline performance with only 10% of training data. Feature visualization and saliency analysis suggested clinically meaningful representations aligned with established electrocardiographic criteria. These findings indicate that large-scale ECG-report contrastive pre-training can expand routine ECG interpretation beyond common arrhythmias toward broad cardiovascular assessment and opportunistic screening of echocardiographic and rare conditions.
【18】SeqRoute: Global Budget-Aware Sequential LLM Routing via Offline Reinforcement Learning
标题：SeqRoute：通过离线强化学习的全球预算感知顺序LLM路由
链接：https://arxiv.org/abs/2605.25424
摘要：Existing LLM routing frameworks treat queries as independent events, neglecting the sequential nature of real-world user sessions constrained by global computational budgets. This mismatch inevitably leads to budget bankruptcy: myopic routing policies exhaust resources on early interactions, forcing subsequent and often more complex queries onto inadequate models. We introduce SeqRoute, a framework that formulates multi-turn routing as a finite-horizon Markov Decision Process and solves it via offline reinforcement learning. By incorporating the remaining budget into the state space and training with Conservative Q-Learning (CQL), SeqRoute learns delayed gratification to strategically preserve resources for high-stakes turns later in the session. To overcome data starvation, we propose Hindsight Budget Relabeling (HBR). This technique retrospectively simulates historical trajectories under diverse hypothetical budgets, expanding 10,000 raw sessions into 2.38 million transitions enriched with critical bankruptcy signals. At deployment, a dynamic $λ$-sweep mechanism enables zero-shot navigation of the cost-quality Pareto frontier without retraining. Extensive evaluations demonstrate that SeqRoute reduces operational costs by 6.0-73.5% while maintaining or improving quality, and suppresses bankruptcy rates to under 1%, strictly dominating behavior cloning, budget-aware heuristics, and static baselines across the entire Pareto frontier.
【19】A general tensor-structured compression scheme for efficient large language models
标题：用于高效大型语言模型的通用张量结构压缩方案
链接：https://arxiv.org/abs/2605.25344
备注：12 pages, 4 figures
摘要：Large language models (LLMs) are dominated by dense linear transformations, whose storage, memory and computational overheads hinder efficient adaptation and deployment while masking the functional impacts of structural simplification. Here we present Tensor Mixture (MixT), a general tensor-structured compression scheme that replaces targeted dense linear layers with natively executable mixtures of tensor operators. Operating directly on generic linear projections instead of model-specific components, MixT is potentially applicable across Transformer-based LLMs and other dense neural mappings. We evaluate MixT on Qwen3-8B and LLaMA2-7B under a unified recovery protocol, identifying a broad compressible regime in which MMLU accuracy is largely preserved before an abrupt transition at model-specific boundaries. This transition coincides with coordinated shifts in output entropy, prediction entropy and inter-layer geometry. At the LLaMA2-7B transition boundary, MixT reduces full-model parameters by 47.5\%, inference FLOPs by 37.1\%, training FLOPs by 52.1\% and peak inference memory by 60.4\%, demonstrating its practical potential for lower-cost LLM compression.
【20】CausalFlow: Causal Attribution and Counterfactual Repair for LLM Agent Failures
标题：因果流：LLM代理故障的因果归因和反事实修复
链接：https://arxiv.org/abs/2605.25338
【21】Influence-Inspired Spectral Rotations for Extreme Low-Bit LLM Quantization
标题：极低比特LLM量化的影响激励谱旋转
链接：https://arxiv.org/abs/2605.25203
备注：14 pages, no figures. Companion application paper to arXiv:2605.01637 (theory). Code and pinned eval stack: https://github.com/gogipav14/spectral-llm
【22】Directional Alignment Mitigates Reward Hacking in Reinforcement Learning for Language Models
标题：方向对齐减轻语言模型强化学习中的奖励黑客
链接：https://arxiv.org/abs/2605.25189
【23】Polynomial Context-Truncation Sensitivity in Autoregressive Language Models: Sequential Wyner-Ziv Bounds for KV Cache Compression
标题：自回归语言模型中的多项上下文截断敏感性：KV缓存压缩的顺序Wyner-Ziv界限
链接：https://arxiv.org/abs/2605.25085
【24】Security in the Fine-Tuning Lifecycle of Large Language Models: Threats, Defenses,Evaluation, and Future Directions
标题：大型语言模型微调收件箱中的安全性：威胁、防御、评估和未来方向
链接：https://arxiv.org/abs/2605.25073
备注：39 pages, 7 figures, 22 tables
【25】Large Language Model Selection with Limited Annotations
标题：带有有限注释的大型语言模型选择
链接：https://arxiv.org/abs/2605.24981
备注：33 pages, 5 figures, 4 tables
【26】Mitigating Object Hallucinations in Vision-Language Models through Region-Aware Attention Recalibration
标题：通过区域感知注意力重新校准缓解视觉语言模型中的对象幻觉
链接：https://arxiv.org/abs/2605.24957
【27】Memory-Induced Tool-Drift in LLM Agents
标题：LLM代理中的记忆引起的工具漂移
链接：https://arxiv.org/abs/2605.24941
【28】When Reasoning Hurts: Source-Aware Evaluation of Frontier LLMs for Clinical SOAP Note Generation
标题：当推理受到伤害时：临床肥皂笔记生成的前沿LLM的源感知评估
链接：https://arxiv.org/abs/2605.24902
【29】Efficient DP-SGD for LLMs with Randomized Clipping
标题：用于具有随机裁剪的LLM的高效DP-SGD
链接：https://arxiv.org/abs/2605.24879
备注：Accepted at ICML 2026
【30】Tiny Brains, Giant Impact: Uncovering the Keystone Neurons of LLM with Just a Few Prompts
标题：微小的大脑，巨大的影响：只需几个预言就能揭开法学硕士的Keystone神经元
链接：https://arxiv.org/abs/2605.24846
【31】RouteScan: A Non-Intrusive Approach to Auditing MoE LLMs Safety via Expert Routing Telemetry
标题：RouteScan：一种通过专家路由远程通信审计MoE LLM安全的非侵入性方法
链接：https://arxiv.org/abs/2605.24817
备注：20 pages. Under submission
【32】CONF-KV: Confidence-Aware KV Cache Eviction with Mixed-Precision Storage for Long-Horizon LLM
标题：CONF-KV：具有信任意识的KV缓存驱逐，采用混合精度存储，用于长地平线LLM
链接：https://arxiv.org/abs/2605.24786
【33】From Theory to Decision Rule: Calibrating the Noisy-Label Crossover for Vision-Language Model Weak Supervision Across Three Medical-Imaging Benchmarks
标题：从理论到决策规则：校准视觉语言模型的噪音标签交叉三个医学成像基准的弱监督
链接：https://arxiv.org/abs/2605.24771
备注：5 pages, 2 figures, 4 tables
【34】CyberMaskQA: A Privacy-Aware Benchmark for Evaluating Large Language Models in Cybersecurity Question Answering
标题：CyberMaskQA：用于评估网络安全问题解答中大型语言模型的隐私意识基准
链接：https://arxiv.org/abs/2605.24765
【35】Bilevel Optimization of Synthetic Trajectories for Multi-Turn LLM Fine-Tuning
标题：多圈LLM微调综合轨迹的二层优化
链接：https://arxiv.org/abs/2605.24743
【36】When Mean CE Fails: Median CE Can Better Track Language Model Quality
标题：当平均CE失败时：平均CE可以更好地跟踪语言模型质量
链接：https://arxiv.org/abs/2605.24667
备注：20 pages
【37】How Many Tools Should an LLM Agent See? A Chance-Corrected Answer
标题：LLM代理应该查看多少个工具？机会纠正的答案
链接：https://arxiv.org/abs/2605.24660
备注：13 pages, 2 figures
【38】IterInject: Indirect Prompt Injection Against LLM Agents via Feedback-Guided Iterative Optimization
标题：IterInjects：通过反馈引导迭代优化针对LLM代理的间接即时注入
链接：https://arxiv.org/abs/2605.24659
备注：Submitted to EMNLP 2026
【39】Measuring the Depth of LLM Unlearning via Activation Patching
标题：通过激活修补来衡量LLM忘记学习的深度
链接：https://arxiv.org/abs/2605.24614
备注：18 pages
【40】Summoning the Oracle to Slay It: Mitigating Look-Ahead Bias in Financial Backtesting with Large Language Models
标题：召唤Oracle来消灭它：用大型语言模型减轻财务回溯测试中的前瞻偏见
链接：https://arxiv.org/abs/2605.24564
【41】Jailbreak to Protect: Buffering and Reinforcing via Temporary Jailbreaking for Safe Fine-Tuning in Large Language Models
标题：越狱保护：通过临时越狱缓冲和强化，以在大型语言模型中进行安全微调
链接：https://arxiv.org/abs/2605.24550
备注：ICML 2026 Spotlight
【42】SemanticZip: A Pilot Framework for Lossy Text Compression with LLMs as Semantic Decompressors
标题：SemanticZip：一个使用LLM作为语义解压缩器的有损文本压缩试验框架
链接：https://arxiv.org/abs/2605.24541
备注：13 pages, 1 figure, 2 tables. Pilot framework paper; code and supplementary artifacts available in ancillary files
【43】Hypothesis Generation and Inductive Inference in Children and Language Models
标题：儿童和语言模型中的假设生成和归纳推理
链接：https://arxiv.org/abs/2605.24528
【44】Market Regime Council for Dynamic Credit Assignment in Multi-Agent LLM Decision Systems
标题：多代理LLM决策系统中动态信用分配市场制度委员会
链接：https://arxiv.org/abs/2605.24490
备注：35 pages, 13 figures, preprint
【45】Poisoning the Watchtower: Prompt Injection Attacks Against LLM-Augmented Security Operations Through Adversarial Log Content
标题：毒害瞭望塔：通过对抗性日志内容对LLM增强安全操作进行即时注入攻击
链接：https://arxiv.org/abs/2605.24421
备注：10 pages
【46】LLMTabBench: Evaluating LLMs on Binary Tabular Classification From Zero to Few Shots
标题：LLMTabBench：从零到几个镜头评估二进制表格分类的LLM
链接：https://arxiv.org/abs/2605.24417
【47】CurveRL: Principled Distribution-Aware Context Reweighting for LLM Reasoning
标题：CurveRL：LLM推理的原则性分布感知上下文重新加权
链接：https://arxiv.org/abs/2605.24331
【48】Omissive Bias in Religious Representation: Benchmarking LLM Answers to Everyday Ethical Decision-making
标题：宗教代表中的宽松偏见：LLM日常道德决策的答案基准
链接：https://arxiv.org/abs/2605.24319
【49】Discovering Lexical Gaps Using Embeddings from Multilingual LLMs
标题：使用多语言LLM的嵌入发现词汇差距
链接：https://arxiv.org/abs/2605.24310
备注：CoNLL 2026
【50】ChaosBench-Logic v2: Evaluating LLM Logical Reasoning over Dynamical Systems at Scale
标题：ChaosBench-Logic v2：大规模评估动态系统上的LLM逻辑推理
链接：https://arxiv.org/abs/2605.24305
备注：14 pages, 8 figures. Published at the ICLR 2026 Workshop on LLM Reasoning
【51】Enhancing Reliability in LLM-Based Secure Code Generation
标题：增强基于LLM的安全代码生成的可靠性
链接：https://arxiv.org/abs/2605.24300
备注：15 pages; 7 tables; 3 figures
【52】LLMs Show No Signs Of Individuated Metacognition
标题：法学硕士没有表现出个性化元认知的迹象
链接：https://arxiv.org/abs/2605.24299
【53】An Empirical Evaluation of LLM-Generated Code Security Across Prompting Methods
标题：不同预算方法对LLM生成的代码安全性的实证评估
链接：https://arxiv.org/abs/2605.24298
备注：40 pages, 22 tables, 8 figures
【54】TUBE: Tangent Upper Bound on Evidence for Discrete Diffusion Language Models
标题：TUBE：离散扩散语言模型证据的切向上界
链接：https://arxiv.org/abs/2605.24292
备注：Preprint. 9 pages main text, 5 figures, plus appendix
【55】Agent-ToM: Learning to Monitor Autonomous LLM Agents via Theory-of-Mind Reasoning
标题：Agent-ToM：学习通过心理理论推理监控自主LLM代理
链接：https://arxiv.org/abs/2605.24216
备注：23 pages, 9 figures
【56】When Does Multi-Agent RL Improve LLM Workflows? Workflow, Scale, and Policy-Sharing Tradeoffs
标题：多代理RL何时改善LLM工作流程？工作流程、规模和政策共享权衡
链接：https://arxiv.org/abs/2605.24202
【57】Extracting Training Data from Diffusion Language Models via Infilling
标题：通过填充从扩散语言模型中提取训练数据
链接：https://arxiv.org/abs/2605.24173
【58】PromptAudit: Auditing Prompt Sensitivity in LLM-Based Vulnerability Detection
标题：EntAudit：审计基于LLM的漏洞检测中的提示敏感性
链接：https://arxiv.org/abs/2605.24171
【59】EVA: Accelerating LLM Decoding via an Efficient Vector Quantization Architecture
标题：伊娃：通过高效的载体量化架构加速LLM解码
链接：https://arxiv.org/abs/2605.24144
备注：17 pages. Accepted to ISCA 2026
【60】Breaking the Chains of Probability: Neutrosophic Logic as a New Framework for Epistemic Uncertainty in Large Language Models
标题：打破概率的链条：中性逻辑作为大型语言模型中认知不确定性的新框架
链接：https://arxiv.org/abs/2605.24053
备注：Published in Neutrosophic Sets and Systems, Vol. 99 (2026). Author's preprint version. Open code and data available at: github.com/mleyvaz/neutrosophic-llm-logic
【61】Truthful Online Preference Aggregation for LLM Fine-Tuning in Mobile Crowdsourcing
标题：移动众包中LLM微调的真实在线偏好聚合
链接：https://arxiv.org/abs/2605.24052
【62】Mixture of Complementary Agents for Robust LLM Ensemble
标题：强大的LLM注册的补充剂混合物
链接：https://arxiv.org/abs/2605.24048
【63】LLM-AutoSciLab: Closed-Loop Scientific Discovery via Active Experimentation with LLMs
标题：LLM-AutoSciLab：通过LLM主动实验实现闭环科学发现
链接：https://arxiv.org/abs/2605.24043
【64】A Tabular Schedule Abstraction for Communication-Aware Evaluation of Pipeline-Parallel LLM Training
标题：流水线并行LLM训练的沟通意识评估的表格时间表抽象
链接：https://arxiv.org/abs/2605.24006
备注：Accepted at the 25th IEEE International Symposium on Parallel and Distributed Computing (ISPDC 2026)
【65】Reason--Imagine--Act: Closed-Loop LLM Decision Making with World Models for Autonomous Driving
标题：理由--想象--行动：采用自动驾驶世界模型的闭环LLM决策
链接：https://arxiv.org/abs/2605.24004
备注：Accepted by the 2026 IEEE International Conference on Intelligent Transportation Systems (ITSC 2026). 8 pages, 2 figures
【66】LGMT: Logic-Grounded Metamorphic Testing for Evaluating the Reasoning Reliability of LLMs
标题：LGMT：用于评估LLM推理可靠性的逻辑基础变形测试
链接：https://arxiv.org/abs/2605.23965
备注：Zheng Zheng is the corresponding author
【67】Authority Inversion in LLM-Mediated Ubiquitous Systems: When Models Trust Users Over Sensors
标题：LLM介导的泛在系统中的权威倒置：当模型信任用户而不是传感器时
链接：https://arxiv.org/abs/2605.23938
【68】When Correct Beliefs Collapse: Epistemic Resilience of LLMs under Clinical Pressure
标题：当正确信念崩溃时：临床压力下LLM的认识弹性
链接：https://arxiv.org/abs/2605.23932
备注：ACL 2026
【69】How Much Thinking is Enough? Quantifying and Understanding Redundancy in LLM Reasoning
标题：思考多少才足够？量化和理解LLM推理中的冗余
链接：https://arxiv.org/abs/2605.23926
【70】Confidence Calibration in Large Language Models
标题：大型语言模型中的置信度校准
链接：https://arxiv.org/abs/2605.23909
【71】Check Your LLM's Secret Dictionary! Five Lines of Code Reveal What Your LLM Learned (Including What It Shouldn't Have)
标题：查看您的LLM秘密词典！五行代码揭示了您的LLM学到了什么（包括它不应该学到的东西）
链接：https://arxiv.org/abs/2605.22005
【72】DiscoverPhysics: Benchmarking LLMs for Out-of-the-Box Scientific Thinking
标题：Discovery Physics：开箱即用的科学思维的法学硕士基准
链接：https://arxiv.org/abs/2605.26087
Graph相关(图学习|图神经网络|图优化等)(20篇)
【1】MDGMIX: Boundary-Aware Subgraph Mixing for Multi-Domain Graph Pre-Training
标题：MDGMIX：用于多域图预训练的边界感知子图混合
链接：https://arxiv.org/abs/2605.25771
备注：Accepted by ICML2026
【2】Closed-Form Node Classification with Exact Graph Unlearning
标题：具有精确图取消学习的闭合形式节点分类
链接：https://arxiv.org/abs/2605.25662
备注：19 pages, 5 figures, 12 tables (7 main + 5 appendix)
【3】'Si'multaneous 'S'patial-'T'emporal Message Passing for Dynamic Graph Representation Learning
标题：动态图表示学习的“Si”多次S '空间-' T '商业消息传递
链接：https://arxiv.org/abs/2605.25548
【4】Rethinking Feature Alignment in Generalist Graph Anomaly Detection: A Relational Fingerprint-based Approach
标题：重新思考通用图异常检测中的特征对齐：一种基于关系指纹的方法
链接：https://arxiv.org/abs/2605.25429
备注：9 pages, 7 figures. Accepted by ICML 2026
【5】Capture-Calibrate-Coach: A Graph-Based Framework for Knowledge Monitoring Estimation and Adaptive Feedback
标题：捕获-校准-教练：基于图形的知识监控估计和自适应反馈框架
链接：https://arxiv.org/abs/2605.25419
备注：To be published in Proceedings of the 27th International Conference on Artificial Intelligence in Education (AIED 2026)
【6】GL-LFGNN:A Global-Local Dual-branch Causal Graph Neural Network Based on Liang-Kleeman Information Flow for EEG Emotion Recognition
标题：GL-LFGNN：基于Liang-Kleeman信息流的全局-局部双分支因果图神经网络用于脑电情绪识别
链接：https://arxiv.org/abs/2605.25061
备注：10 pages, 3 figures
【7】TGFormer: Towards Temporal Graph Transformer with Auto-Correlation Mechanism
标题：TG Former：迈向具有自相关机制的时间图Transformer
链接：https://arxiv.org/abs/2605.24971
【8】MedMamba: Multi-View State Space Models with Adaptive Graph Learning for Medical Time Series Classification
标题：MedMamba：用于医学时间序列分类的具有自适应图学习的多视图状态空间模型
链接：https://arxiv.org/abs/2605.24961
备注：Accepted to 2026 ICML
【9】Aligning Molecular Graph Explanations with Chemical Identity via InChIfied Invariants
标题：通过InChItified不变量将分子图解释与化学同一性对齐
链接：https://arxiv.org/abs/2605.24742
【10】Beyond the Aggregation Dilemma: Prior-Retaining Decoupled Learning for Multimodal Graphs
标题：超越聚合困境：多模式图的预先保留的脱钩学习
链接：https://arxiv.org/abs/2605.24684
【11】Representation-Guided Discrete Molecular Graph Retrosynthesis
标题：表示引导的离散分子图逆合成
链接：https://arxiv.org/abs/2605.24428
【12】Gaussian Rank-Based Neighborhood Degree for Graph Neural Networks in Image Classification
标题：图像分类中基于高斯排序的图神经网络邻居度
链接：https://arxiv.org/abs/2605.24367
【13】Treatment Effect Estimation with Differentiated Networked Effect on Graph Data
标题：图数据上区分网络效应的治疗效果估计
链接：https://arxiv.org/abs/2605.24358
备注：Accepted by the research track of the KDD 2026 conference
【14】Knowledge Graph Modulated Deep Learning for Limited-Sample Clinical Data Analysis
标题：知识图调制的深度学习用于有限样本临床数据分析
链接：https://arxiv.org/abs/2605.24162
备注：17 pages, 4 figures, 12 supplementary figures
【15】Generative Representation Learning on Hyper-relational Knowledge Graphs via Masked Discrete Diffusion
标题：通过掩蔽离散扩散进行超关系知识图的生成性表示学习
链接：https://arxiv.org/abs/2605.24064
备注：28 pages, 16 figures, 18 tables, 43rd International Conference on Machine Learning (ICML 2026)
【16】Fuzzy, Neutrosophic, and Uncertain Graph Theory: Properties and Applications
标题：模糊、中性与不确定图论：性质与应用
链接：https://arxiv.org/abs/2605.23936
备注：326 pages. Publisher: Neutrosophic Science International Association (NSIA) Publishing House. ISBN: 978-197250204-4
【17】Different Statistical Perspectives for Understanding Generalisation in Graph Neural Networks
标题：理解图神经网络中推广的不同统计视角
链接：https://arxiv.org/abs/2605.25452
备注：15 pages, 4 figures, submission for Special Issue in AStA Advances in Statistical Analysis
【18】Learning manifold diffusion semigroups from graph transition matrices
标题：从图转移矩阵学习多维扩散半群
链接：https://arxiv.org/abs/2605.25383
【19】Parameter-Efficient CT Reconstruction via Deep Graph Laplacian Regularization
标题：通过深度图拉普拉斯正规化进行参数高效的CT重建
链接：https://arxiv.org/abs/2605.25348
备注：7 pages, 3 figures, conference
【20】Affinity Graph Connectivity in Convex Clustering
标题：凸聚类中的仿射图连通性
链接：https://arxiv.org/abs/2605.24673
备注：28 pages, 6 figures
Transformer(16篇)
【1】OrpQuant: Geometric Orthogonal Residual Projection for Multiplier-Free Power-of-Two Transformer Quantization
标题：OrpQuant：用于无乘子二次乘势Transformer量化的几何垂直剩余投影
链接：https://arxiv.org/abs/2605.26092
【2】The Quantization Benefits of Residual-Free Transformers
标题：无残留Transformer的量化好处
链接：https://arxiv.org/abs/2605.25880
备注：Under review
【3】Geometric Evolution Maps: Extracting Stable Concept Probes from Transformer Residual Streams
标题：几何进化图：从Transformer剩余流中提取稳定的概念探针
链接：https://arxiv.org/abs/2605.25848
备注：24 pages, 3 figures. Reference implementation: rosetta_tools v1.3.1 (doi:10.5281/zenodo.20361433)
【4】Analogies between Transformer Layers and Power Method
标题：Transformer层与幂法的类比
链接：https://arxiv.org/abs/2605.25619
【5】Continuous-Depth Field Theory for Transformer Patching and Mechanistic Interpretability
标题：Transformer修补和机械解释的连续深度场理论
链接：https://arxiv.org/abs/2605.25225
【6】Blocked Gibbs meets Diffusion Transformers: Unsupervised Learning for Constraint Optimization
标题：受阻的吉布斯会见扩散变形者：约束优化的无监督学习
链接：https://arxiv.org/abs/2605.25129
【7】PQDT: Pseudo-Query Dual Transformer for Robust Point Cloud Restoration
标题：PQDT：用于点云鲁棒恢复的伪查询对偶Transformer
链接：https://arxiv.org/abs/2605.25127
备注：To be published in The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026
【8】The Concept Allocation Zone: Tracking How Concepts Form Across Transformer Depth
标题：概念分配区：跟踪概念如何在Transformer深度中形成
链接：https://arxiv.org/abs/2605.24856
备注：34 models, 8 architectural families, 7 concepts. Companion papers: GEM (arXiv forthcoming), CAZ Validation (arXiv forthcoming), PRH Validation (arXiv forthcoming). Code: https://github.com/jamesrahenry/Rosetta_Tools
【9】Muon in Vision Transformers: Optimizer-Recipe Interactions and Gradient Spectra
标题：Vision Transformers中的μ子：优化器与配方的相互作用和梯度光谱
链接：https://arxiv.org/abs/2605.24770
备注：25 pages, 15 figures
【10】CSP-Atlas: Concept-Specific Neural Circuits in a Sparse Python Transformer
标题：CSP-Atlas：稀疏Python Transformer中的特定概念神经电路
链接：https://arxiv.org/abs/2605.24603
备注：Code: https://github.com/piotrwilam/AtlasCSP
【11】Polymorphism Is Rotation: Operational Mechanistic Interpretability from a Two-Layer Transformer to Pythia-70m
标题：多形性就是旋转：从两层Transformer到Pythia-70 m的操作机制解释
链接：https://arxiv.org/abs/2605.24577
备注：26 pages, 4 figures, 40 references. Pre-registered four-bar framework; all numerical claims reproducible
【12】Momentum Streams for Optimizer-Inspired Transformers
标题：受优化者启发的Transformer的势头源源不断
链接：https://arxiv.org/abs/2605.24425
【13】Spectral Probe-Circuits: A Three-Step Recipe for Identifying Attention-Head Circuits in Pretrained Transformers
标题：频谱探测电路：识别预训练Transformer中注意力头电路的三步方法
链接：https://arxiv.org/abs/2605.24059
备注：35 pages, 4 figures
【14】Towards Verifiable Transformers: Solver-Checkable Circuit Explanations
标题：迈向可验证的Transformer：可求解器检查的电路简化
链接：https://arxiv.org/abs/2605.24033
【15】StrTransformer: Source-Wise Structured Transformers for Unsupervised Blind Source Recovery
标题：StrTransformer：用于无监督盲源恢复的源级结构化Transformer
链接：https://arxiv.org/abs/2605.25648
【16】From Index to Equity: Pre-Training Transformers for Stock Return Prediction
标题：从指数到股票：股票回报预测的预训练Transformer
链接：https://arxiv.org/abs/2605.23962
GAN|对抗|攻击|生成相关(17篇)
【1】Paris 2.0: A Decentralized Diffusion Model for Video Generation
标题：巴黎2.0：视频生成的去中心扩散模型
链接：https://arxiv.org/abs/2605.26064
备注：6 pages, 5 figures
【2】Building an Adversarial Malware Dataset by Family and Type: Generation, Evasion, and Poisoning Evaluation
标题：按系列和类型构建对抗性恶意软件数据集：生成、规避和中毒评估
链接：https://arxiv.org/abs/2605.25937
【3】Context-Instrumental Data Distillation for Kubernetes Manifest Generation: Method and Experimental Evaluation
标题：用于Kubernetes清单生成的上下文工具数据蒸馏：方法和实验评估
链接：https://arxiv.org/abs/2605.25835
备注：15 pages, 4 figures, 2 tables
【4】Clarify, Abstain or Answer? Strategising in Conversation with Belief-Augmented Generation
标题：澄清、弃权还是回答？与信仰增强一代的对话制定策略
链接：https://arxiv.org/abs/2605.25831
【5】Opportunistic Target Selection: Early Directional Commitment for Query-Efficient Black-Box Adversarial Attacks
标题：启发式目标选择：查询高效黑匣子对抗攻击的早期定向承诺
链接：https://arxiv.org/abs/2605.25663
备注：13 pages, 10 figures, 3 tables; code available at https://github.com/Tariolle/opportunistic-target-selection
【6】Geometric Flow Matching for Molecular Conformation Generation via Manifold Decomposition
标题：通过总管分解生成分子形态的几何流匹配
链接：https://arxiv.org/abs/2605.25577
【7】Generating 3D models from sketches of human faces using a combined approach of Convolutional Neural Networks, Procedural Modeling, and Contour Mapping
标题：使用卷积神经网络、过程建模和轮廓映射的组合方法从人脸草图生成3D模型
链接：https://arxiv.org/abs/2605.25418
备注：A thesis submitted in conformity with the requirements for the degree of Master of Science in Computer Science Graduate Department of Computer Science University of Toronto
【8】When Interpretability Becomes a Liability: Adversarial Attacks on CBM Concept Layers
标题：当可解释性成为一种负担：对煤层气概念层的对抗性攻击
链接：https://arxiv.org/abs/2605.25304
备注：Accepted to CVPR 2026 (Findings). 9 pages, 6 figures
【9】Localization then Neutralization: Gradient-guided Token Suppression against Visual Prompt Injection Attack
标题：本地化后中和：针对视觉提示注入攻击的攻击者引导令牌抑制
链接：https://arxiv.org/abs/2605.25194
【10】TRACE: A taxonomy-grounded synthetic dataset for teaching-program generation and session interpretation in Applied Behavior Analysis
标题：TRACE：基于分类学的合成数据集，用于应用行为分析中的教学程序生成和会话解释
链接：https://arxiv.org/abs/2605.25038
备注：11 pages, 3 tables. Dataset: https://huggingface.co/datasets/PomboLabs/TRACE ; code: https://github.com/Pombo-Labs/TRACE
【11】Self-Balancing Gradient Allocation for Heterogeneity-Aware Feature Generation in Click-Through Rate Prediction
标题：点入率预测中具有异类意识的特征生成的自平衡梯度分配
链接：https://arxiv.org/abs/2605.24986
备注：12 pages, 5 figures, 4 tables
【12】DriftingMol: Decoder-Coupled Drift for One-Pass Property-Conditional Molecular Generation
标题：DriftingMol：用于单程性质条件分子生成的解码器耦合漂移
链接：https://arxiv.org/abs/2605.24841
备注：9 pages, 5 figures
【13】Cross-Domain Energy-Guided Diffusion Generation for Off-Dynamics Reinforcement Learning
标题：用于非动态强化学习的跨域能量引导扩散生成
链接：https://arxiv.org/abs/2605.24810
备注：29 pages, 3 figures, and 14 tables
【14】Steering Beyond the Support: Adversarial Training on Unsupervised Jailbroken Activation Simulation
标题：超越支持：无监督越狱激活模拟的对抗训练
链接：https://arxiv.org/abs/2605.24535
备注：accepted by ICML 2026
【15】Balancing Fairness, Privacy, and Accuracy: A Multitask Adversarial Framework for Centralized Data-Driven Systems
标题：平衡公平性、隐私和准确性：集中式数据驱动系统的多任务对抗框架
链接：https://arxiv.org/abs/2605.24458
备注：13 Pages, 6 figures, IEEE TKDE
【16】Structure-Aware RAG: Structured Retrieval Augmented Generation from Noisy Data for Conversational Agents
标题：结构感知RAG：对话代理的有噪数据的结构化检索增强生成
链接：https://arxiv.org/abs/2605.24366
【17】Multimodal Alignment and Preference Optimization for Zero-Shot Conditional RNA Generation
标题：Zero-Shot条件RNA生成的多模式比对和偏好优化
链接：https://arxiv.org/abs/2605.23961
半/弱/无/有监督|不确定性|主动学习(22篇)
【1】Goal-driven Bayesian Optimal Experimental Design for Robust Decision-Making Under Model Uncertainty
标题：模型不确定性下稳健决策的目标驱动的Bayesian最优实验设计
链接：https://arxiv.org/abs/2605.26093
【2】Active Query Synthesis for Preference Learning
标题：用于偏好学习的主动查询合成
链接：https://arxiv.org/abs/2605.26072
备注：27 pages, 12 figures
【3】Rethinking Weak Supervision in Anomaly Detection: A Comprehensive Benchmark
标题：重新思考异常检测中的薄弱监管：综合基准
链接：https://arxiv.org/abs/2605.26068
备注：Accepted at KDD 2026 Datasets and Benchmarks Track (Cycle 2)
【4】UNATE: UNsupervised ATomic Embedding for crystal structures property prediction
标题：UNATE：用于晶体结构性质预测的无监督原子嵌入
链接：https://arxiv.org/abs/2605.25866
【5】When Self-Belief Misleads: Active Label Acquisition for Reinforcement Learning with Verifiable Rewards
标题：当自信误导时：具有可验证奖励的强化学习的主动标签获取
链接：https://arxiv.org/abs/2605.25864
【6】NPSolver: Neural Poisson Solver with Iterative Physics Supervision
标题：NPSolver：具有迭代物理监督的神经Poisson求解器
链接：https://arxiv.org/abs/2605.25786
备注：kdd 2026
【7】DeGRe: Dense-supervised Generative Reranking for Recommendation
标题：DeGR：密集监督的生成式推荐重新排名
链接：https://arxiv.org/abs/2605.25749
备注：Accepted to KDD 2026 (ADS Track)
【8】Courtroom Analogy: New Perspective on Uncertainty-Aware Classification
标题：法庭类比：不确定性意识分类的新视角
链接：https://arxiv.org/abs/2605.25616
备注：ICML 2026
【9】Learning Permutation from Structure Without Supervision
标题：从没有监督的结构中学习排列
链接：https://arxiv.org/abs/2605.25551
【10】Quantifying Empirical Compute-Supervision Tradeoffs in RLVR
标题：量化WLVR中的经验计算机监管权衡
链接：https://arxiv.org/abs/2605.25252
备注：Workshop on Combining Theory and Benchmarks @ ICML 2026
【11】On the Epistemic Uncertainty of Overparametrized Neural Networks
标题：过度参数化神经网络的认识不确定性
链接：https://arxiv.org/abs/2605.25234
备注：Accepted at ICML 2026 (Main Track)
【12】Multi-Objective Learning for Diffusion Models: A Statistical Theory under Semi-Supervised Learning
标题：扩散模型的多目标学习：半监督学习下的统计理论
链接：https://arxiv.org/abs/2605.25210
【13】Uncertainty-DTW for Sequences and Visual Tokens
标题：序列和视觉标记的不确定性DTW
链接：https://arxiv.org/abs/2605.25110
备注：Research report
【14】Selective Test-Time Compute Scaling for Click-Through Rate Prediction via Uncertainty-Triggered Feature Path Exploration
标题：通过不确定性触发特征路径探索进行点入率预测的选择性测试时计算缩放
链接：https://arxiv.org/abs/2605.24989
备注：12 pages, 4 Figures, 3 Tables
【15】SEED: Semi-supervised Continual MalwarE Detection for Tackling ConcEpt Drift on a BuDget
标题：SEED：半监督连续MalwarE检测，用于解决BuDget上的概念漂移
链接：https://arxiv.org/abs/2605.24903
【16】Active Learning for Stochastic Contextual Linear Bandits
标题：随机上下文线性盗贼的主动学习
链接：https://arxiv.org/abs/2605.24803
【17】Physen-Noise2Noise: Physics-Guided Self-Supervised Defocus Deblurring with Bias Correction under Low-Light Conditions
标题：Physen-Noise 2 Noise：弱光条件下物理引导的自监督散焦去模糊和偏差修正
链接：https://arxiv.org/abs/2605.24590
备注：14 pages
【18】AvAtar: Learning to Align via Active Optimal Transport
标题：AvAtar：学会通过主动最佳交通保持一致
链接：https://arxiv.org/abs/2605.24395
备注：Published as a conference paper at ICML 2026
【19】Concept Drift Adaptation Using Self-Supervised and Reinforcement Learning In Android Malware Detection
标题：Android恶意软件检测中使用自监督和强化学习的概念漂移适应
链接：https://arxiv.org/abs/2605.24294
备注：9 pages, 2 figures, 2 tables
【20】Music Transcription with (Almost) No Supervision
标题：（几乎）没有监督的音乐抄写
链接
：https://arxiv.org/abs/2605.24193
【21】Overcoming "Physics Shock" in Earth Observation A Heteroscedastic Uncertainty Framework for PINN-based Flood Inference
链接：https://arxiv.org/abs/2605.24106
备注：This article is accepted in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
【22】Adaptive Preference Optimization with Uncertainty-aware Utility Anchor
标题：具有不确定性感知效用锚的自适应偏好优化
链接：https://arxiv.org/abs/2509.10515
备注：Accepted by EMNLP 2025 Findings
迁移|Zero/Few/One-Shot|自适应(21篇)
【1】A Multimodal 3D Foundation Model for Light Sheet Fluorescence Microscopy Enables Few-Shot Segmentation, Classification, and Deblurring
标题：用于光片荧光显微镜的多模式3D基础模型实现Few-Shot分割、分类和去模糊
链接：https://arxiv.org/abs/2605.26026
备注：11 pages, 3 figures
【2】Creative Quality Alignment: Expert Tacit Knowledge Transfer via Chain-of-Thought Fine-Tuning
标题：创意质量一致：通过思想链微调进行专家隐性知识转移
链接：https://arxiv.org/abs/2605.25977
【3】Quantitative Evaluation of the Severity of Posttraumatic Stress Disorder through Transfer Learning from Specific Phobia Data
标题：通过特定恐惧症数据的迁移学习定量评估创伤后应激障碍的严重程度
链接：https://arxiv.org/abs/2605.25933
备注：Submitted to a peer-reviewed journal, comments welcome
【4】DVAO: Dynamic Variance-adaptive Advantage Optimization for Multi-reward Reinforcement Learning
标题：DVAO：多奖励强化学习的动态方差自适应优势优化
链接：https://arxiv.org/abs/2605.25604
【5】Grow-Prune-Freeze Networks: Adaptive & Continual Learning Technique for Olfactory Navigation
标题：生长-修剪-冻结网络：嗅觉导航的自适应和持续学习技术
链接：https://arxiv.org/abs/2605.25170
【6】ASTRO: Adaptive Spatio-Temporal Reinforcement Optimization for GNN Powered Anomly Detection in Cyber Physical Systems
标题：ASTRO：网络物理系统中GNN支持的异常检测的自适应时空强化优化
链接：https://arxiv.org/abs/2605.25135
【7】Theoretical Analysis of Sparse Optimization with Reparameterization, Weight Decay, and Adaptive Learning Rate
标题：具有重新参数化、权重衰减和自适应学习率的稀疏优化理论分析
链接：https://arxiv.org/abs/2605.25134
备注：31 pages, 5 figures. Submitted to ICML 2026
【8】Trust-Aware Joint Feature-Prediction Discrepancy for Robust Domain Adaptation
标题：用于鲁棒领域自适应的信任感知联合竞争预测差异
链接：https://arxiv.org/abs/2605.25119
备注：Research report
【9】Courant: a State-Adaptive Perceiver-Based Neural Surrogate with Local Support and Interpretable Field Decomposition
标题：Courant：一种具有局部支持和可解释场分解的基于状态自适应感知器的神经代理
链接：https://arxiv.org/abs/2605.25115
【10】Scale When Needed: Adaptive Neuron-level Mixed Precision Quantization Aware Training
标题：需要时进行规模：自适应神经元级混合精度量化感知训练
链接：https://arxiv.org/abs/2605.25054
备注：Accepted at ICML - GlobalSouthML workshop, 2026
【11】HumanEgo: Zero-Shot Robot Learning from Minutes of Human Egocentric Videos
标题：HumanEgo：Zero-Shot机器人从几分钟的人类自我中心视频中学习
链接：https://arxiv.org/abs/2605.24934
备注：Project page: https://humanego-ai.github.io
【12】T2S-MPC: Time-Embedded Online Adaptive Model Predictive Control for Time-Varying Dynamics
标题：T2 S-MPC：时变动态的时间嵌入在线自适应模型预测控制
链接：https://arxiv.org/abs/2605.24852
【13】PILOT: Policy-Informed Learned Optimization for Adaptive Deep Network Training
标题：PILOT：基于策略的学习优化，用于自适应深度网络训练
链接：https://arxiv.org/abs/2605.24570
备注：16 pages, 5 figures
【14】AI-Driven Adaptive Adversaries and the Erosion of Cryptographic Trust in Public Key Systems
标题：人工智能驱动的自适应对手和公钥系统中加密信任的侵蚀
链接：https://arxiv.org/abs/2605.24542
【15】MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding
标题：MindAlign：桥接EEG，视觉和语言，用于零拍摄视觉解码
链接：https://arxiv.org/abs/2605.24523
备注：20 pages, 10 figures, 15 tables
【16】Asymmetric Adaptation-based Real-time Fault Diagnosis Under Transitional Operating Conditions
标题：过渡工作条件下基于非对称自适应的实时故障诊断
链接：https://arxiv.org/abs/2605.24457
备注：6 pages, 3 figures, Accepted by ICAIS & ISAS 2026
【17】A Reinforcement Learning Inspired Latent Yield Based Adaptive Algorithm Switching Mechanism
标题：基于强化学习的潜在收益率的自适应算法切换机制
链接：https://arxiv.org/abs/2605.24436
备注：Accepted and published in the Proceedings of the 29th European Conference on Applications of Evolutionary Computation (EvoApplications 2026), held as part of EvoStar 2026, Toulouse, France, April 8 to 10, 2026. Lecture Notes in Computer Science (LNCS), Springer Nature Switzerland
【18】Private Adaptive Covariance Estimation via Gaussian Graphical Models
标题：通过高斯图形模型的私人自适应协方差估计
链接：https://arxiv.org/abs/2605.24295
【19】Signs Beat Floats: Low-Rank Double-Binary Adaptation for On-Device Fine-Tuning
标题：Signs Beat Floats：用于设备上微调的低秩双二进制适配
链接：https://arxiv.org/abs/2605.24058
备注：34 pages, 3 figures
【20】Geometry Adaptive Counterfactual Distribution Learning with Diffusion-Guided Smoothing
标题：具有扩散引导平滑的几何自适应反事实分布学习
链接：https://arxiv.org/abs/2605.25811
【21】Distributionally Robust Transfer Learning with Structurally Missing Covariates, with Application to Cross-National Cardiac Arrest Prediction
标题：结构缺失协变量的分布鲁棒转移学习，应用于跨国心脏骤停预测
链接：https://arxiv.org/abs/2605.24212
强化学习(17篇)
【1】Global Convergence of Wasserstein Policy Gradient for Entropy-Regularized Reinforcement Learning
标题：Wasserstein政策梯度的全球收敛用于熵正规化强化学习
链接：https://arxiv.org/abs/2605.26078
【2】Learning in Low-Dimensional Subspaces: Orthogonal Bottlenecks for Reinforcement Learning
标题：低维度子空间中的学习：强化学习的垂直瓶颈
链接：https://arxiv.org/abs/2605.26012
【3】Joint Optimization of Training and Inference in Federated Edge Learning via Constrained Multi-Objective Deep Reinforcement Learning
标题：通过约束多目标深度强化学习联合优化联邦边缘学习中的训练和推理
链接：https://arxiv.org/abs/2605.25916
【4】Latent Representation Alignment for Offline Goal-Conditioned Reinforcement Learning
标题：离线目标条件强化学习的潜在表示对齐
链接：https://arxiv.org/abs/2605.25740
备注：Accepted in ICML 2026
【5】Reinforcement Learning from Denoising Feedback
标题：来自去噪反馈的强化学习
链接：https://arxiv.org/abs/2605.25638
【6】Latent Q-Barrier Shielding for Safe In-Context Reinforcement Learning
标题：用于安全上下文强化学习的潜在Q屏障屏蔽
链接：https://arxiv.org/abs/2605.25267
【7】Reinforcement Learning for Laser Additive Manufacturing Scan-Order Optimisation: A Bilevel Proxy--FEA Diagnostic Framework for Reward and World-Model Diagnosis
标题：激光增材制造扫描顺序优化的强化学习：双层代理--奖励和世界模型诊断的FEA诊断框架
链接：https://arxiv.org/abs/2605.25063
备注：31 pages, 7 figures, 3 tables
【8】A perspective on fluid mechanical environments for challenges in reinforcement learning
标题：从流体机械环境的角度应对强化学习挑战
链接：https://arxiv.org/abs/2605.25011
【9】Scaling up Energy-Aware Multi-Agent Reinforcement Learning for Mission-Oriented Drone Networks with Individual Reward
标题：为面向任务的无人机网络扩展能量感知多智能体强化学习，并具有个人奖励
链接：https://arxiv.org/abs/2605.24992
备注：IEEE Internet of Things Journal
【10】Unifying Value Alignment and Assignment in Cross-Domain Offline Reinforcement Learning with Heterogeneous Datasets
标题：统一跨域离线强化学习中的价值对齐和分配与异类数据集
链接：https://arxiv.org/abs/2605.24862
备注：Accepted at ICML 2026
【11】A Contractive Feedback Semantics for Reinforcement Learning
标题：强化学习的压缩反馈语义
链接：https://arxiv.org/abs/2605.24759
【12】Reinforcement Learning for Reachability: Guaranteeing Asymptotic Optimality
标题：可达性的强化学习：保证渐进最优性
链接：https://arxiv.org/abs/2605.24740
备注：Main text and appendix of work accepted in ICML 2026
【13】Streaming Reinforcement Learning under Partial Observability with Real-Time Recurrent Learning
标题：部分可观察性下的实时回归学习流强化学习
链接：https://arxiv.org/abs/2605.24709
备注：16 pages, 4 figures
【14】Vision-Guided Outdoor Flight and Obstacle Evasion via Reinforcement Learning
标题：通过强化学习的视觉引导户外飞行和障碍物躲避
链接：https://arxiv.org/abs/2605.24449
备注：Published in IEEE Robotics and Automation Letters, vol 11, no 2. Presented at the IEEE International Conference on Robotics and Automation 2026
【15】Evolving Robustness--Exploration Trade-off in Online Reinforcement Learning via Quantile Bayesian Risk MDPs
标题
：不断发展的鲁棒性--通过分位数Bayesian风险MDP进行在线强化学习的探索权衡
链接：https://arxiv.org/abs/2605.24345
【16】IVR-R1: Refining Trajectories through Iterative Visual-Grounded Reasoning in Reinforcement Learning
标题：IVR-R1：通过强化学习中的迭代视觉推理来精炼轨迹
链接：https://arxiv.org/abs/2605.23997
【17】Counterfactually Safe Reinforcement Learning
标题：反事实安全强化学习
链接：https://arxiv.org/abs/2605.25114
医学相关(2篇)
【1】Explainable Retinal Imaging for Prediction of Multi-Organ Dysfunction in Type 2 Diabetes
标题：可解释视网膜成像预测2型糖尿病多器官功能障碍
链接：https://arxiv.org/abs/2605.24912
备注：15 pages, 8 figures
【2】HeartBeatAI: An Interpretable and Robust Deep Learning Framework for Multi-Label ECG Arrhythmia Detection
标题：HeartBeatAI：一个用于多标签ECG心律失常检测的可解释且强大的深度学习框架
链接：https://arxiv.org/abs/2605.24588
蒸馏|知识提取(2篇)
【1】Extreme Region Policy Distillation
标题：极端地区政策提炼
链接：https://arxiv.org/abs/2605.25582
【2】MEDAL: Manifold Embedding Distillation via Autoencoder Learning
标题：MEDAL：通过自动编码器学习嵌入蒸馏的多路模块
链接：https://arxiv.org/abs/2605.24244
推荐(3篇)
【1】First, do no harm: Breaking suicidogenic echo chambers in media recommendation
标题：首先，不要伤害：在媒体推荐中打破自杀回音室
链接：https://arxiv.org/abs/2605.25258
备注：10 pages, 5 figures. Research on safety-aware recommender systems and algorithmic ethics
【2】An Interpretable CF-RL-TOPSIS Fusion Model for Skills-Aware Talent Recommendation
标题：技能意识人才推荐的可解释CF-RL-TOPSIS融合模型
链接：https://arxiv.org/abs/2605.24155
备注：Preprint submitted to Knowledge-Based Systems; 4 figures and 8 tables
【3】Choosing Online Experiment Designs under Interference in Ads, Recommendations, and Member-Experience Systems
标题：在广告、推荐和员工体验系统干扰下选择在线实验设计
链接：https://arxiv.org/abs/2605.25290
聚类(2篇)
【1】Cluster Frequency Conformal Prediction for Local Coverage
标题：局部覆盖的集群频率保形预测
链接：https://arxiv.org/abs/2605.24872
【2】Clustering based on Stochastic Dominance with application for risk averters and risk seekers
标题：基于随机优势的集群，应用于风险规避者和风险寻求者
链接：https://arxiv.org/abs/2605.24422
超分辨率|去噪|去模糊|去雾(1篇)
【1】Everything at Every Scale: Scale-Invariant Diffusion with Continuous Super-Resolution
标题：各种规模的一切：具有连续超分辨率的尺度不变扩散
链接：https://arxiv.org/abs/2605.26032
备注：29 pages, 17 figures
自动驾驶|车辆|车道检测等(3篇)
【1】RECTOR: Priority-Aware Rule-Based Reranking for Compliance-Aware Autonomous Driving Trajectory Selection
标题：Rector：基于规则的重新排序，以实现合规性意识的自动驾驶轨迹选择
链接：https://arxiv.org/abs/2605.25095
【2】Convex-Neural RRT*: Fast and Reliable Learning-Guided Sampling for High-Quality Robot Path Planning
标题：凸神经RTI *：快速可靠的学习引导采样，用于高质量机器人路径规划
链接：https://arxiv.org/abs/2605.25006
【3】Metropolis-Scale Resilient and Trustworthy Traffic Flow Inference Using Multi-Source Data
标题：利用多源数据进行大都市规模弹性且可信的交通流推断
链接：https://arxiv.org/abs/2605.25004
备注：The paper has been submitted to Elsevier for possible publication
点云|SLAM|雷达|激光|深度RGBD相关(4篇)
【1】Length Generalization with Log-Depth Recurrent Units
标题：具有对深回归单位的长度概括
链接：https://arxiv.org/abs/2605.26035
备注：39 pages, 11 figures
【2】OSDTW: Optimal Shared Depth and Task Weighting for Long-Tailed Recognition
标题：OSDTW：长尾识别的最佳共享深度和任务加权
链接：https://arxiv.org/abs/2605.24969
备注：ICIC 2026 Oral
【3】Learning Laplacian Eigenspace with Mass-Aware Neural Operators on Point Clouds
标题：利用点云上的质量感知神经运算学习拉普拉斯特征空间
链接：https://arxiv.org/abs/2605.24390
【4】Growing a Neural Network in Breadth, Depth, and Time
标题
：在广度、深度和时间上发展神经网络
链接：https://arxiv.org/abs/2605.25174
联邦学习|隐私保护|加密(4篇)
【1】Personalized Federated Learning by Energy-Efficient UAV Communications
标题：通过节能无人机通信进行个性化联邦学习
链接：https://arxiv.org/abs/2605.25212
【2】Hardware-Aware Federated Learning for Speech Emotion Recognition
标题：用于语音情感识别的硬件感知联邦学习
链接：https://arxiv.org/abs/2605.24712
备注：4 pages, 3 figures, 4 Tables
【3】ChainLearn: A Blockchain-Based Capacity-Aware Framework for Federated Ensemble Learning
标题：ChainLearn：基于区块链的联合参与学习能力感知框架
链接：https://arxiv.org/abs/2605.24418
备注：10 pages, 7 figures, 11 tables. IEEE conference format. Code: https://github.com/EdddTri/ChainLearn
【4】Federated Learning over Human-Body Communication for On-Body Edge Intelligence: A Survey, Taxonomy, and BODYFED-HBC Scheduling Vignette
标题：基于人体通信的联邦学习以实现身体边缘智能：调查、分类和BODYFED-HBC调度小插曲
链接：https://arxiv.org/abs/2605.24062
推理|分析|理解|解释(16篇)
【1】From Latent Space to Training Data: Explainable Specialization in Minimal MLPs
标题：从潜在空间到训练数据：最小MLP的可解释专业化
链接：https://arxiv.org/abs/2605.25939
【2】Conformalised imprecise inference for robust extrapolation under limited data
标题：有限数据下稳健外推的保形不精确推断
链接：https://arxiv.org/abs/2605.25882
备注：10 pages, 5 figures
【3】On Reliability of Efficient Membership Inference Vulnerability Evaluation
标题：高效隶属推理漏洞评估的可靠性研究
链接：https://arxiv.org/abs/2605.25819
备注：14 pages, 10 figures
【4】Electricity Consumption Forecasting: An Approach Using Cooperative Ensemble Learning with SHapley Additive exPlanations
标题：用电量预测：一种使用Shapley添加性exPlanations的合作集群学习的方法
链接：https://arxiv.org/abs/2605.25305
【5】Inference-Time Alignment of Diffusion Models via Trust-Region Iterative Twisted Sequential Monte Carlo
标题：通过信任区域迭代扭曲序列蒙特卡罗实现扩散模型的推理时间对齐
链接：https://arxiv.org/abs/2605.25123
备注：34 pages, 6 figures, and 7 tables
【6】Position: AI for Science Should Treat Measurement-to-Dataset Pipelines as Inference Components
标题：立场：科学人工智能应将测量到数据集管道视为推理组件
链接：https://arxiv.org/abs/2605.24558
备注：23 pages, 5 figures, Proceedings of the 43 rd International Conference on Machine Learning, Seoul, South Korea. PMLR 306, 2026
【7】Refined Analysis of Entropy-Regularized Actor-Critic
标题：对熵规范化演员评论家的细化分析
链接：https://arxiv.org/abs/2605.24357
【8】Faithfulness as Information Flow: Evaluating and Training Faithful Chain-of-Thought Reasoning
标题：作为信息流的忠诚度：评估和训练忠诚思想链推理
链接：https://arxiv.org/abs/2605.24286
【9】Inference Time Context Sparsity: Illusion or Opportunity?
标题：推理时间上下文稀疏性：幻觉还是机会？
链接：https://arxiv.org/abs/2605.24168
备注：19 pages, 8 figures
【10】Riemannian Archetypal Analysis: Interpretable non-linear data analysis on deformed star distributions
标题：Riemann原型分析：变形恒星分布的可解释非线性数据分析
链接：https://arxiv.org/abs/2605.24113
【11】DRIVE: Modeling Skills at the Reasoning and Interaction Levels for Web Agents under Continual Learning
标题：DRIVE：持续学习下Web代理推理和交互层面的建模技能
链接：https://arxiv.org/abs/2605.23939
备注：35 pages, 5 figures
【12】Statistical Inference for Stochastic Gradient Descent Beyond Finite Variance
标题：超出有限方差的随机梯度下降的统计推断
链接：https://arxiv.org/abs/2605.26000
【13】From DPPs to $k$-DPPs: identifiability analysis via spectral decomposition
标题：从DPP到$k$-DPP：通过谱分解进行可识别性分析
链接：https://arxiv.org/abs/2605.25526
备注：10 pages
【14】How Neural Reward Models Learn Features for Policy Optimization: A Single-Index Analysis
标题：神经奖励模型如何学习策略优化的特征：单指标分析
链接：https://arxiv.org/abs/2605.24749
备注：35 pages
【15】LWM-CDE: A Representation Space for Wireless Data Reasoning and Transferability
标题：LWM-ADE：无线数据推理和可传输性的表示空间
链接：https://arxiv.org/abs/2605.24077
备注：The model and relevant scripts are available on the WILab Hugging Face page: https://huggingface.co/wi-lab
【16】Physics-Guided Concentration Inference from Resistance Transients in a Mixed-Phase SnO-SnO$_2$ Carbon Monoxide Sensor with p-n Switching
标题：基于电阻瞬态的混合相SnO-SnO $2 $CO传感器p-n开关浓度推断
链接：https://arxiv.org/abs/2605.23971
备注：15 pages, 14 figures
检测相关(9篇)
【1】Retrieval-Augmented Detection of Potentially Abusive Clauses in Chilean Terms of Service
标题：智利服务条款中潜在虐待条款的检索增强检测
链接：https://arxiv.org/abs/2605.26019
备注：42 pages, 6 figures, 9 tables
【2】A Multimodal Framework for Dementia Detection via Linguistic and Acoustic Representation Learning
标题：通过语言和声学表示学习进行痴呆症检测的多模式框架
链接：https://arxiv.org/abs/2605.25540
【3】Cross-Domain Generalization Limits of Vision Foundation Models in Facial Deepfake Detection
标题：面部Deepfake检测中视觉基础模型的跨域概括限制
链接：https://arxiv.org/abs/2605.24965
【4】CALIBURN: A Regime-Sensitivity Study of Operationally Calibrated Streaming Intrusion Detection
标题：CalibURN：操作校准流入侵检测的机制敏感性研究
链接：https://arxiv.org/abs/2605.24696
备注：55 pages, 5 figures, 14 tables. Under review at Cyber Security and Applications. Code: https://github.com/MichelYsf/rcbsid-paper. Archived release: https://doi.org/10.5281/zenodo.20074590
【5】Lake Detection and Water Quality Estimation in Sentinel-2 Data
标题：Sentinel-2数据中的湖泊检测和水质估算
链接：https://arxiv.org/abs/2605.24515
【6】Rethinking Continual Anomaly Detection on the Edge: Benchmarking Under Realistic Industrial Conditions
标题：重新思考边缘连续异常检测：现实工业条件下的基准
链接：https://arxiv.org/abs/2605.24251
【7】CAFD: Concept-Aware DNN Fault Detection using VLMs
标题：CAFD：使用VLM的概念感知DNN故障检测
链接：https://arxiv.org/abs/2605.24008
【8】Parameter Efficient Multi-Class Intelligent Scheduling for Multimodal Online Distributed Industrial Anomaly Detection
标题：用于多模式在线分布式工业异常检测的参数高效多类智能调度
链接：https://arxiv.org/abs/2605.23984
【9】Detecting Metastable Basins in High Dimensions via Marginal Trajectory Distribution Discrimination
标题：通过边缘轨迹分布辨别检测高维度亚稳盆地
链接：https://arxiv.org/abs/2605.24136
分类|识别(1篇)
【1】Optimal Design for Multinomial Logit Model with Applications to Best Assortment Identification
标题：多项Logit模型的优化设计及其在最佳组合识别中的应用
链接：https://arxiv.org/abs/2605.25592
备注：Accepted at ICML 2026
表征(7篇)
【1】Neuronal Stochastic Attention Circuit (NSAC) for Probabilistic Representation Learning
标题：用于概率表示学习的神经元随机注意力电路（NSO）
链接：https://arxiv.org/abs/2605.26061
【2】CITYREP: A Unified Benchmark for Urban Representations Across Cities, Tasks, and Modalities
标题：CITYBEP：跨城市、任务和模式的城市代表的统一基准
链接：https://arxiv.org/abs/2605.26036
【3】TopoAlign: Topology-Aware Visual Representation Alignment
标题：TopoAlign：具有布局感知的视觉表示对齐
链接：https://arxiv.org/abs/2605.25541
【4】Abduction-Deduction Entanglement: Domain Generalization via Representation Transplants
标题：溯因-演绎纠缠：通过表征移植的领域泛化
链接：https://arxiv.org/abs/2605.25156
【5】Synheart Capacity: A Theory-Driven Physiological Representation of Cognitive Capacity Dynamics from Wearable Signals
标题：Synheart容量：来自可穿戴信号的认知能力动态的理论驱动生理表示
链接：https://arxiv.org/abs/2605.24416
【6】Learning Protein Structure-Function Relationships through Knowledge-guided Representation Decomposition
标题：通过知识引导的表示分解学习蛋白质结构功能关系
链接：https://arxiv.org/abs/2605.23960
备注：28 pages, 17 figures, icml 2026 regular
【7】PilotWiMAE: Pilot-Native Representation Learning for Wireless Channels
标题：PilotWiMAE：无线通道的Pilot-Native表示学习
链接：https://arxiv.org/abs/2605.22856
3D|3D重建等相关(1篇)
【1】3D Magnetic Field Reconstruction and Mapping with Physics-Informed Neural Networks
标题：使用物理信息神经网络进行3D磁场重建和绘制
链接：https://arxiv.org/abs/2605.25640
编码器(2篇)
【1】Hermite-NGP: Gradient-Augmented Hash Encoding for Learning PDEs
标题：Hermite-NGP：一种用于学习偏微分方程的递归增强哈希编码
链接：https://arxiv.org/abs/2605.24774
备注：Accepted by ICML 2026.Project page: https://jinjinhe2001.github.io/hermite-ngp/
【2】Phase-Aware Wavelet-Based-Scattering Encoder-Decoder for Dense Predictions
标题：用于密集预测的相感知基于子波的散射编码器-解码器
链接：https://arxiv.org/abs/2605.24621
备注：21 pages, 16 figures, 10 tables
优化|敛散性(12篇)
【1】Optimal and Order-optimal Gated Priority-based Greedy Policies for Two-layer Multi-item Order Fulfillment
标题：两层多物品订单履行的最优和最优基于优先级的门控贪婪策略
链接：https://arxiv.org/abs/2605.25888
【2】DeepSeekMath Meets Order Book: Group-Aware Policy Optimization for High-Frequency Directional Trading
标题：DeepSeekMath符合订单书：高频定向交易的群体意识政策优化
链接：https://arxiv.org/abs/2605.25527
备注：9 pages, 3 figures
【3】EMA-Nesterov: Stabilizing Nesterov's Lookahead for Accelerated Deep Learning Optimization
标题：EMA-Nesterov：稳定Nesterov的展望以加速深度学习优化
链接：https://arxiv.org/abs/2605.25395
备注：25 page, 10 figures
【4】Label-NTK Alignments and A Tighter Convergence Bound in the NTK Regime
标题：标签-NTK路线和NTK制度中更严格的收敛界限
链接：https://arxiv.org/abs/2605.25275
【5】Investigating the Interplay between Contextual and Parametric Chain-of-Thought Faithfulness under Optimization
标题：研究优化下上下文和参数思维链忠诚度之间的相互作用
链接：https://arxiv.org/abs/2605.24960
备注：The first two authors contributed equally and share first-authorship
【6】Global linear convergence of entropy-regularized softmax policy gradient beyond tabular MDPs
标题：表格式MDPs之外的信息规范化softmax政策梯度的全球线性收敛
链接：https://arxiv.org/abs/2605.24939
【7】Zeroth-Order Nonconvex Nonsmooth Optimization with Heavy-Tailed Noise
标题：具有重尾噪音的零阶非凸非光滑优化
链接：https://arxiv.org/abs/2605.24513
【8】Generative OOD-regularized Model-based Policy Optimization
标题：生成式OOD正规化基于模型的策略优化
链接：https://arxiv.org/abs/2605.24405
【9】BoxLitE: A Faithful Knowledge Base Embedding Based on Convex Optimization
标题：BoxLitE：基于凸优化的忠实知识库嵌入
链接：https://arxiv.org/abs/2605.23937
备注：28 pages. Full version of paper accepted to KR 2026 (23nd International Conference on Principles of Knowledge Representation and Reasoning). Track: KR meets Machine Learning and Explanation
【10】Implicit Binarization via Complex Phase Dynamics in Combinatorial Optimization
标题：组合优化中的复相动力学隐式二值化
链接：https://arxiv.org/abs/2605.24502
备注：27 pages, 5 figures
【11】Multicalibration Boosting: Theory, Convergence, and Transferability
标题：多校准增强：理论、收敛和可移植性
链接：https://arxiv.org/abs/2605.24364
【12】Optimal Non-Asymptotic Edgeworth Expansions for Multivariate Neural Network Outputs
标题：多元神经网络输出的最优非渐进Edgeworth展开
链接：https://arxiv.org/abs/2605.24072
备注：34 pages, 2 figures
预测|估计(23篇)
【1】Predicting Stock Price Direction on Earnings Announcement Days using Multi-modal Deep Learning
标题：使用多模式深度学习预测盈利公告日的股价方向
链接：https://arxiv.org/abs/2605.25894
【2】Stochastic Estimation of the Layer-wise Hessian Trace for Monitoring Neural-network Training
标题：用于监测神经网络训练的分层Hessian轨迹的随机估计
链接：https://arxiv.org/abs/2605.25674
备注：9 pages, 1 table
【3】Learning Latent Dynamical Causal Processes for Single-Cell Perturbation Prediction
标题：单细胞微扰预测的潜在动态因果过程
链接：https://arxiv.org/abs/2605.25581
备注：Accepted to SIGKDD 2026 AI4Science Track
【4】UWM-JEPA: Predictive World Models That Imagine in Belief Space
标题：UWM-JEPA：在信仰空间中想象的预测世界模型
链接：https://arxiv.org/abs/2605.25313
备注：14 pages, 6 figures, 7 tables. Code and data: https://github.com/santoshkumarradha/uwm-jepa
【5】Eureka: Intelligent Feature Engineering for Enterprise AI Cloud Resource Demand Prediction
标题：尤里卡：企业人工智能云资源需求预测的智能特征工程
链接：https://arxiv.org/abs/2605.25297
备注：13 pages, accepted at DASFAA 2026 (International Conference on Database Systems for Advanced Applications)
【6】AME-TS: Anchored Mixture-of-Experts for Time Series Forecasting
标题：AME-TS：用于时间序列预测的锚定混合专家
链接：https://arxiv.org/abs/2605.25166
【7】Benchmarking non-conformity score functions in conformal prediction
标题：适形预测中的不符合评分函数基准
链接：https://arxiv.org/abs/2605.24983
备注：3 tables, 1 supplementary table, 1 supplementary figure
【8】RealBench: Benchmarking Data-Driven Numerical Weather Forecasting Under Operational Conditions and Extreme Event Challenges
标题：RealBench：在操作条件和极端事件挑战下对数据驱动的数值天气预报进行基准测试
链接：https://arxiv.org/abs/2605.24945
备注：35 pages, 22 figures
【9】Factorize to Generalize: Retrieval-Guided Invariant-Dynamic Decomposition for Time Series Forecasting
标题：因式分解推广：时间序列预测的检索引导不变动态分解
链接：https://arxiv.org/abs/2605.24911
【10】DBPnet: Damper Characteristics-Based Bayesian Physics-Informed Neural Network for Wheel Load Estimation
标题：DBPnet：用于车轮负载估计的基于减震器特征的Bayesian物理信息神经网络
链接：https://arxiv.org/abs/2605.24860
备注：14 pages, 12 figures, 6 tables
【11】Disentangled Double Machine Learning for Accurate Causal Effect Estimation
标题：分离双机器学习实现准确的因果效应估计
链接：https://arxiv.org/abs/2605.24808
备注：15 pages, 9 figures
【12】Deep ZakaiJ: Structured Filtering for Jump-Diffusion Time Series Forecasting
标题：Deep ZakaiJ：跳跃扩散时间序列预测的结构化过滤
链接：https://arxiv.org/abs/2605.24548
【13】Temporal Concept Drift in Legal Judgment Prediction: Neural Baselines Across Three Epochs of Ukrainian Court Decisions
标题：法律判决预测中的时间概念漂移：乌克兰法院判决三个时期的神经基线
链接：https://arxiv.org/abs/2605.24452
备注：17 pages, 6 tables, 5 figures. Dataset: https://huggingface.co/datasets/overthelex/ukrainian-court-decisions
【14】Assessing the Operational Viability of Foundation Models for Time Series Forecasting
标题：评估时间序列预测基础模型的操作可行性
链接：https://arxiv.org/abs/2605.24381
备注：21 pages, 8 Figures, Code available at [https://github.com/kavin-soni/timeseries-zeroshot-eval]
【15】Nano World Models: A Minimalist Implementation of Future Video Prediction
标题：Nano World模型：未来视频预测的极简实现
链接：https://arxiv.org/abs/2605.23993
备注：Project page: https://simchowitzlabpublic.github.io/nano-world-model/
【16】Algometrics: Forecasting Under Algorithmic Feedback
标题：数学测量：数学反馈下的预测
链接：https://arxiv.org/abs/2605.23978
【17】LETS Forecast: Learning Embedology for Time Series Forecasting
标题：LETS预测：学习时间序列预测的嵌入学
链接：https://arxiv.org/abs/2506.06454
备注：Accepted at International Conference on Machine Learning (ICML) 2025
【18】Multimodality Stacking with Blockwise missing values and application to the PIONeeR biomarkers study for prediction of resistance to immunotherapy
标题：利用Bestival缺失值进行多模式叠加以及在PIONeeR生物标志物研究中的应用，用于预测免疫治疗耐药性
链接：https://arxiv.org/abs/2605.25050
【19】Deep Learning-Enabled Prediction of Geoeffective CMEs Using SOHO and SDO Observations
标题：使用SOHO和NSO观测进行深度学习预测地球有效CME
链接：https://arxiv.org/abs/2605.24748
备注：23 pages, 12 figures, 4 tables
【20】AnnotateMissense: a genome-wide annotation and benchmarking framework for missense pathogenicity prediction
标题：AnnotateMisense：用于Misense致病性预测的全基因组注释和基准框架
链接：https://arxiv.org/abs/2605.24520
【21】Seeing Inside the Storm: Improving Nowcasting by Integrating Meteorological Drivers
标题：洞察风暴内幕：通过整合气象驱动因素来改进即时预报
链接：https://arxiv.org/abs/2605.24067
【22】Aurora Hunter: A Two-Stage Framework for Probabilistic Visibility Forecasting
标题：Aurora Hunter：概率能见度预报的两阶段框架
链接：https://arxiv.org/abs/2605.24038
【23】Game-Theoretic Modeling of Heterogeneous Investor Interactions for Stock Price Forecasting
标题：用于股价预测的异类投资者互动的博弈论模型
链接：https://arxiv.org/abs/2605.23953
备注：10 pages, 1 figure, intended for conference submission
其他神经网络|深度学习|模型|建模(54篇)
【1】From Model Scaling to System Scaling: Scaling the Harness in Agentic AI
标题：从模型扩展到系统扩展：扩展大型人工智能中的扩展
链接
：https://arxiv.org/abs/2605.26112
【2】AdvantageFlow: Advantage-Weighted Least Squares for RL in Flow Models
标题：MetallageFlow：流模型中RL的加权最小平方
链接：https://arxiv.org/abs/2605.26013
【3】Fuzzy PyTorch: Rapid Numerical Variability Evaluation for Deep Learning Models
标题：模糊PyTorch：深度学习模型的快速数值变异性评估
链接：https://arxiv.org/abs/2605.25991
备注：19 pages, 8 figures, Published in Transactions on Machine Learning Research (01/2026)
【4】Small Models, Strong Priors: Architectural Inductive Bias for Parameter-Efficient Neural PDE Solvers
标题：小模型，强先验：参数高效的神经PED求解器的架构归纳偏差
链接：https://arxiv.org/abs/2605.25949
【5】Universal Activation Verbalizer: A Unified Framework for Cross-Model Activation Explanation
标题：通用激活语言器：跨模型激活解释的统一框架
链接：https://arxiv.org/abs/2605.25903
备注：23 pages, 11 figures, 11 tables
【6】Concept Unlearning via Cross-Attention Activation Projection for Diffusion Models
标题：扩散模型的交叉注意激活投影概念忘记学习
链接：https://arxiv.org/abs/2605.25765
【7】Don't Retrain, Just Reuse: Recovering Dual-Target Molecules from Single-Target Diffusion Models
标题：不要重新训练，只需重复使用：从单目标扩散模型中恢复双目标分子
链接：https://arxiv.org/abs/2605.25681
【8】When In-Distribution Gains Fail: Evaluating Weak-to-Strong Reward Models under Preference Shift
标题：当分配内收益失败时：评估偏好转变下从弱到强的奖励模型
链接：https://arxiv.org/abs/2605.25629
备注：Code: https://anonymous.4open.science/r/w2s_reward_ood-682F
【9】Generalized Evidential Deep Learning: From a Bayesian Perspective
标题：广义证据深度学习：从Bayesian的角度来看
链接：https://arxiv.org/abs/2605.25599
备注：Submitted to ICML2026
【10】Missing Pattern Recognized Diffusion Imputation Model for Missing Not At Random
标题：缺失非随机缺失模式识别扩散插补模型
链接：https://arxiv.org/abs/2605.25439
【11】ViroBench: Benchmarking Nucleotide Foundation Models on Viral Genomics Tasks
标题：ViroBench：病毒基因组学任务中的核苷基金会模型基准
链接：https://arxiv.org/abs/2605.25388
备注：42 pages,15 figures
【12】PDEInvBench: A Comprehensive Dataset and Design Space Exploration of Neural Networks for PDE Inverse Problems
标题：PDEInvBench：用于PCE逆问题的神经网络的全面数据集和设计空间探索
链接：https://arxiv.org/abs/2605.25353
备注：37 total pages, 13 main pages, 20 figures, 8 tables. Published in Transactions on Machine Learning Research (TMLR), 2026
【13】Parallel Differentiable Reachability for Learning and Planning with Certified Neural Dynamics and Controllers
标题：使用认证的神经动力学和控制器实现学习和规划的并行可微可达性
链接：https://arxiv.org/abs/2605.25346
备注：Robotics: Science and Systems XXII (RSS 2026)
【14】Evolving Causal Regulatory Networks (ECR-Net)
标题：不断发展的因果监管网络（ECR-Net）
链接：https://arxiv.org/abs/2605.25211
备注：9 pages, 6 figures. Presents ECR-Net, an evolutionary framework for adaptive causal structure discovery under non-stationarity, with empirical evaluation against NOTEARS, PCMCI+, and related baselines
【15】Hide to Guide: Learning via Semantic Masking
标题：隐藏指南：通过语义掩蔽学习
链接：https://arxiv.org/abs/2605.25198
【16】Learning Treatment Effects during Resource Allocation via Priority-Queue Randomization
标题：通过优先级队列随机化进行资源分配期间的学习治疗效果
链接：https://arxiv.org/abs/2605.25169
【17】Leveraging Gauge Freedom for Learning Non-Gradient Population Dynamics of Stochastic Systems
标题：利用规范自由度学习随机系统的非梯度种群动力学
链接：https://arxiv.org/abs/2605.25107
【18】Random Neural Network Expressivity for Non-Linear Partial Differential Equations
标题：非线性偏方程的随机神经网络表现性
链接：https://arxiv.org/abs/2605.25057
【19】Learning, locomotion, and navigation of soft synthetic snakes in three-dimensional, heterogeneous environments
标题：软合成蛇在三维、异类环境中的学习、运动和导航
链接：https://arxiv.org/abs/2605.24985
备注：14 pages, 5 figures
【20】BandVQ: Band-Wise Vector-Quantized EEG Foundation Model
标题：BandVQ：逐带向量量化脑电基础模型
链接：https://arxiv.org/abs/2605.24921
备注：15 pages, 1 figure
【21】On the Impact of Class Imbalance on the Learning Dynamics of Deep Neural Networks:An Intuitive Insight
标题：阶级失衡对深度神经网络学习动态的影响：直观的见解
链接：https://arxiv.org/abs/2605.24908
备注：Conference
【22】IV-Net: A neural network for elliptic PDEs with random and highly varying coefficients
标题：IV-Net：具有随机和高度变化系数的椭圆形偏头痛的神经网络
链接：https://arxiv.org/abs/2605.24876
备注：36 pages
【23】A comparative study of accuracy and rollout stability of temporal surrogate models
标题：时态代理模型的准确性和展开稳定性的比较研究
链接：https://arxiv.org/abs/2605.24868
备注：24 pages, 18 figures, submitted to journal
【24】High-fidelity Modeling of Full-scale Pressurized Water Reactor Flow Fields for Machine Learning Applications
标题：机器学习应用的全尺寸压水反应堆流场高保真建模
链接：https://arxiv.org/abs/2605.24763
备注：30 pages, 10 figures, and 6 Tables
【25】A computational phase transition for learning-to-sample from Ising models
标题：从伊辛模型中学习样本的计算阶段转变
链接：https://arxiv.org/abs/2605.24752
【26】Feature Learning in Wide Neural Networks under $μ$P: Identifiability and Sparse-Dictionary Decomposition of the Mean-Field Limit
标题：$μ$P下的宽神经网络特征学习：平均场极限的可识别性和稀疏字典分解
链接：https://arxiv.org/abs/2605.24710
备注：86 pages
【27】Trajectory-Based Difficulty Scoring for Reliable Learning on Tabular Data
标题：基于轨迹的难度评分以实现表格数据的可靠学习
链接：https://arxiv.org/abs/2605.24680
【28】On the Stability and Realizability of Recurrent Polynomial Surrogate Ternary Logic Gate Networks
标题：循环多项替代三值逻辑门网络的稳定性和可实现性
链接：https://arxiv.org/abs/2605.24649
备注：9 pages, 3 figures. This work has been submitted to the IEEE for possible publication
【29】Beyond Fixed Points: Superpolynomial Capacity of Asymmetric Hopfield Networks
标题：超越不稳定点：非对称Hopfield网络的超方程容量
链接：https://arxiv.org/abs/2605.24611
【30】Lattice theory and algebraic models for deep convolutional learning based on mathematical morphology
标题：基于数学形态学的深度卷积学习的格论和代数模型
链接：https://arxiv.org/abs/2605.24608
【31】Learning to Reason Efficiently with A* Post-Training
标题：通过A* 训练后学习有效推理
链接：https://arxiv.org/abs/2605.24597
备注：Preprint
【32】ECHO: Terminal Agents Learn World Models for Free
标题：ECHO：航站楼代理免费学习世界模式
链接：https://arxiv.org/abs/2605.24517
【33】The Normalized Maximum Likelihood for Regular Non-Smooth Models: Measure-Theoretic Foundations and Geometric Sampling
标题：规则非光滑模型的正规化最大似然：测量理论基础和几何抽样
链接：https://arxiv.org/abs/2605.24477
【34】CAffNet: Hard Constraint-Affine Neural Networks
标题：CAffNet：硬约束仿射神经网络
链接：https://arxiv.org/abs/2605.24437
【35】The Model Is Not the Product: A Dual-Pillar Architecture for Local-First Psychological Coaching
标题：模型不是产品：本地优先心理辅导的双支柱架构
链接：https://arxiv.org/abs/2605.24411
备注：10 pages, 3 figures
【36】ChainzRule: Sample-Efficient, Robust Deep Learning Across Tabular, NLP, and Vision Tasks
标题：ChainzRule：跨表格、NLP和Vision任务的样本高效、稳健的深度学习
链接：https://arxiv.org/abs/2605.24340
【37】Fourier Feature Pyramids for Physics-Informed Neural Networks
标题：物理信息神经网络的傅里叶特征金字塔
链接：https://arxiv.org/abs/2605.24278
【38】A lift for input-convex neural network training
标题：输入凸神经网络训练的提升
链接：https://arxiv.org/abs/2605.24274
【39】Optimizing Digital Therapeutic Interventions: Online Learning under Endogenous Adherence
标题：优化数字治疗干预：内生坚持下的在线学习
链接：https://arxiv.org/abs/2605.24261
备注：48 pages, 6 figures
【40】Incorporating Deep Learning Design in Database Queries
标题：在数据库中实现深度学习设计
链接：https://arxiv.org/abs/2605.24207
【41】Filtered Posterior Mean Collections: A Unified Framework for Analytical Models of Diffusion Generalization
标题：过滤后验均值集合：扩散推广分析模型的统一框架
链接：https://arxiv.org/abs/2605.24192
备注：27 Pages, 7 figures
【42】Verified SHAP: Provable Bounds for Exact Shapley Values of Neural Networks
标题：已验证的SHAP：神经网络精确Shapley值的可证明界限
链接：https://arxiv.org/abs/2605.24084
备注：Accepted at ICML 2026. 34 pages, 13 figures
【43】A Large-Scale Dataset and Benchmark: Do Protein-Ligand Models Learn Binding Sites or Just Binding Likelihood?
标题：大规模数据集和基准：蛋白配体模型是学习结合位点还是仅仅学习结合可能性？
链接：https://arxiv.org/abs/2605.24045
备注：Under Review for the NeurIPS 2026 Conference, Track on Evaluations and Datasets
【44】Towards Large Model Feature Coding
标题：走向大模型特征编码
链接：https://arxiv.org/abs/2605.24025
【45】TSFLora: Token-Compressed Split Fine-Tuning for Wireless Edge Networks
标题：TSFLora：用于无线边缘网络的令牌压缩拆分微调
链接：https://arxiv.org/abs/2605.23988
【46】The Model Parking Tax: Quantifying the Hidden Energy Cost of Always-On GPU Model Deployment
标题：停车税模型：量化Always-On图形处理器模型部署的隐藏能源成本
链接：https://arxiv.org/abs/2605.23918
备注：7 pages, 3 figures, 5 tables
【47】Machine Learning Multiscale Interactions
标题：机器学习多尺度交互
链接：https://arxiv.org/abs/2605.25710
【48】PAC Learning with Bandit Feedback: Sharp Sample Complexity in the Realizable Setting
标题：带Bandit反馈的PAC学习：可实现设置中的尖锐样本复杂性
链接：https://arxiv.org/abs/2605.25678
备注：18 pages
【49】Learning Sparse Compositional Functions with Norm-Constrained Neural Networks
标题：利用正规约束神经网络学习稀疏组合函数
链接：https://arxiv.org/abs/2605.25608
【50】QML-PipeGuard: Drift-Aware Behavioral Fingerprinting for Quantum Machine Learning Pipeline Integrity
标题：QML-PipeGuard：用于量子机器学习管道完整性的漂移感知行为指纹识别
链接：https://arxiv.org/abs/2605.25066
备注：54 pages, 12 Tables, 5 figures
【51】Learning dynamical systems with biochemically informed neural ordinary differential equations
标题：使用生物化学信息神经常微方程学习动力系统
链接：https://arxiv.org/abs/2605.24170
备注：23 pages, 13 figures, 4 tables
【52】Multitask learning with semiempirical orbital charges enables sample-efficient MLIPs
标题：具有半经验轨道电荷的多任务学习实现了样本高效的MLIP
链接：https://arxiv.org/abs/2605.24073
备注：16 pages, 6 figures
【53】Volatility Surface Reconstruction using Deep Learning under No-Arbitrage Constraints
标题：无套利约束下使用深度学习重建波动面
链接：https://arxiv.org/abs/2605.24031
备注：MSc thesis, Universidad de Buenos Aires, 2026. 94 pages, 27 figures
【54】Improving Ensemble CAPE Forecasts with a Diffusion Model Incorporating Aerosol Information
标题：利用吸收气溶胶信息的扩散模型改进Ensemble CAPE预测
链接：https://arxiv.org/abs/2605.24009
其他(106篇)
【1】Prism: A Plug-in Reproducible Infrastructure for Scalable Multimodal Continual Instruction Tuning
标题：Prism：用于可扩展多模式连续指令调优的插件可重复基础设施
链接：https://arxiv.org/abs/2605.26110
备注：Code is available at https://github.com/LAMDA-CL/Prism
【2】Conditional KRR: Injecting Unpenalized Features into Kernel Methods with Applications to Kernel Thresholding
标题：有条件KRR：将不受惩罚的特征注入到核心方法中并应用于核心保留
链接：https://arxiv.org/abs/2605.26067
备注：Accepted to ICML 2026
【3】Deployment-complete benchmarking
标题：部署完整基准测试
链接：https://arxiv.org/abs/2605.25997
备注：33 pages, 5 figures, 1 table; supplementary tables and code available
【4】Hidden in Plain Tokens: Simply Robust, Gradient-Free Watermark for Synthetic Audio
标题：隐藏在纯令牌中：用于合成音频的简单稳健、无干扰水印
链接：https://arxiv.org/abs/2605.25967
备注：Accepted to ICML 2026
【5】Mapping the Schedule x Bit-Width Boundary in Sub-100M Quantisation-Aware Training
标题：在Sub-100 M量化感知训练中映射时间表x位宽边界
链接：https://arxiv.org/abs/2605.25966
备注：20 pages, 6 figures, 4 tables. 1345 training runs total (720 + 625). Submitted for review at TMLR
【6】STaT: Resolving Shape Distortion in Non-Stationary Time Series via Tri-Modal Synergy
标题：STaT：通过三模协同解决非平稳时间序列中的形状失真
链接：https://arxiv.org/abs/2605.25943
【7】Multi-Agent Systems are Mixtures of Experts: Who Becomes an Influencer?
标题：多智能体系统是专家的混合体：谁成为影响者？
链接：https://arxiv.org/abs/2605.25929
【8】Does Continued Pretraining on a Learner Corpus Improve Automated Essay Scoring on English Proficiency Tests? Evidence from EFCAMDAT
标题：学习者数据库的持续预训练是否会提高英语能力测试的自动论文评分？来自EFCAMDAT的证据
链接：https://arxiv.org/abs/2605.25924
备注：16 pages, 3 figures, 10 tables, including references and appendices
【9】Reading the Finetuning Prior: Verbatim Content Recovery via Contrastive Decoding Diffing
标题：阅读Finetuning Prior：通过对比解码差异进行逐字内容恢复
链接：https://arxiv.org/abs/2605.25902
【10】The Timing Dependencies of Trust: Speed, Accuracy, and cBCI Neuro-Decoupling in Human-AI Teams
标题：信任的时间依赖：人类-人工智能团队中的速度、准确性和cBCI神经脱钩
链接：https://arxiv.org/abs/2605.25868
【11】Branched Signature Kernel Solvers for ODEs with rough Single-Trajectory signals
标题：粗糙单轨信号常微分方程的分支签名核求解器
链接：https://arxiv.org/abs/2605.25826
备注：39 pages, 12 figures
【12】On the Benefits of Free Exploration for Regret Minimization in Multi-Armed Bandits
标题：论自由探索对减少多臂强盗后悔的好处
链接：https://arxiv.org/abs/2605.25789
备注：55 pages
【13】Invariant-Based Weight Sharing for Message Passing
标题：基于不变的消息传递权重共享
链接：https://arxiv.org/abs/2605.25750
备注：13 pages main paper + 30 pages references and appendix
【14】The Behavioral Credibility Trilemma: When Calibrated Autonomy Becomes Impossible
标题：行为可信度三角困境：当校准的自主性变得不可能
链接：https://arxiv.org/abs/2605.25739
备注：48 pages, 3 figures
【15】FLOATBench: A Dataset and Benchmark for Floating Offshore Wind Turbine Tower Fatigue
标题：FLOATBench：浮动海上风力涡轮机塔疲劳的数据集和基准
链接：https://arxiv.org/abs/2605.25717
【16】Evaluating passing decision-making in professional football: An enhanced MPNN approach to Receiver Selection
标题：评估职业足球中的传球决策：增强的MPNN方法来选择接球手
链接：https://arxiv.org/abs/2605.25696
【17】Insuring Every Action: An Authority Frontier Framework for Runtime Actuarial Control of Autonomous AI Agents
标题：确保每一项行动：自主人工智能代理的精算控制的权威前沿框架
链接：https://arxiv.org/abs/2605.25632
备注：35 pages, 4 figures, 11 tables. Companion paper on the mathematical foundations: SSRN 6761960
【18】CUA-Gym: Scaling Verifiable Training Environments and Tasks for Computer-Use Agents
标题：CUA-Gym：扩展计算机使用代理的可验证训练环境和任务
链接：https://arxiv.org/abs/2605.25624
【19】Towards the Connection between Activation Sparsity and Flat Minima
标题：激活稀疏性和平坦极小值之间的联系
链接：https://arxiv.org/abs/2605.25612
【20】RotMoLE: Enhancing Mixture of Low-Rank Experts through Rotational Gating Mechanism
标题：RotMoLE：通过旋转门控机制增强低级别专家的混合
链接：https://arxiv.org/abs/2605.25565
【21】BC Protocol: Structured Dual-Expert Dialogue for Eliciting High-Quality Chain-of-Thought Post-Training Data
标题：BC协议：结构化双专家对话，用于获取高质量的思想链训练后数据
链接：https://arxiv.org/abs/2605.25549
【22】Relative Repairability: A Calibration-Based Diagnostic for High-Sparsity Post-Pruning Allocation
标题：相对可修复性：高稀疏性修剪后分配的基于校准的诊断
链接：https://arxiv.org/abs/2605.25508
【23】Accelerated Dynamic Importance Weighting with Versatile Divergence-Minimizing Estimators
标题：多用途最小方差估计的加速动态重要性加权
链接：https://arxiv.org/abs/2605.25499
【24】SafetyRepro: Configuration-Conditional Rank Instability on Alignment Benchmarks
标题：SafetyRepro：调整基准上的条件排名不稳定
链接：https://arxiv.org/abs/2605.25492
【25】JacQuant: STE-Free Quantization-Aware Training via Learned Jacobian Surrogates
标题：JacQuant：通过习得的Jacobian代理人进行无STED量化感知训练
链接：https://arxiv.org/abs/2605.25469
【26】Autoregression-Free Neural Operators for Time-Dependent PDEs
标题：时变偏微分方程的自回归自由神经算子
链接：https://arxiv.org/abs/2605.25413
备注：Submitted to IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)
【27】A Context Augmented Multi-Play Multi-Armed Bandit Algorithm for Fast Channel Allocation in Opportunistic Spectrum Access
标题：一种上下文增强的多播放多臂Bandit算法用于均衡频谱接入中快速通道分配
链接：https://arxiv.org/abs/2605.25391
备注：Accepted by ISCC'24
【28】Not only where, But when: Temporal Scheduling for RLVR
标题：不仅在哪里，而且在何时：WLVR的临时调度
链接：https://arxiv.org/abs/2605.25381
备注：Github: https://github.com/Jinghaoleven/RLVR-Schedule
【29】Certified Robustness from Approximate Gaussian Mixture Structures in Pretrained Latent Spaces
标题：预训练潜空间中近似高斯混合结构的鲁棒性认证
链接：https://arxiv.org/abs/2605.25352
【30】ERNIE-Image Technical Report
标题：ERNIE-Image技术报告
链接：https://arxiv.org/abs/2605.25347
【31】Algorithms with Polynomially-Improved Approximation Factors for the $2 \rightarrow q$ Norm, and Applications
标题：2美元的具有多项改进逼近因子的算法 ightarrow q$ Norm和应用
链接：https://arxiv.org/abs/2605.25303
【32】A Principled Self-Referenced Early Stopping Approach for Deep Image Prior
标题：深度图像先验的原则性自指早期停止方法
链接：https://arxiv.org/abs/2605.25299
备注：35 pages, 10 figures, 14 tables
【33】Constraint-Anchored Attribution: Feasibility-Certified Counterfactuals and Bonferroni-PAC Sufficient Subsets for Neural CO Policies
标题：约束锚定归因：神经CO政策的可信度认证反事实和Bonferroni-PAC充分子集
链接：https://arxiv.org/abs/2605.25235
备注：4 pages, 1 figure, Reference implementation: https://github.com/sohaibafifi/neuro-co-cax (MIT)
【34】A Blended Likelihood Approach for Achieving Fairness Using Naive Bayes
标题：使用天真的Bayes实现公平的混合可能性方法
链接：https://arxiv.org/abs/2605.25228
【35】Beyond the Frontier: Stochastic Backtracking for Efficient Test-Time Scaling
标题：超越前沿：随机回溯以实现有效的测试时间扩展
链接：https://arxiv.org/abs/2605.25143
【36】Optimizing Multidimensional Scaling in Gini Metric Spaces
标题：基尼度量空间中的多维缩放优化
链接：https://arxiv.org/abs/2605.25124
【37】Revisiting Pre-Propagation GNNs: Robust Diffusion Operators and Hidden-State Re-Propagation
标题：重温预传播GNN：鲁棒扩散运算符和隐藏状态重传播
链接：https://arxiv.org/abs/2605.25111
【38】MimirRAG: A Multi-Agent RAG Framework for Financial Data Retrieval with Metadata Integration
标题：MimirRAG：一个用于具有元数据集成的金融数据检索的多代理RAG框架
链接：https://arxiv.org/abs/2605.25030
【39】Mitigating Gradient Pathology in PINNs through Aligned Constraint
标题：通过对齐约束缓解PINN中的梯度病理
链接：https://arxiv.org/abs/2605.25001
【40】Bridging the Gap: Enabling Soft Actor Critic for High Performance Legged Locomotion
标题：弥合差距：为高性能腿部运动打造软演员评论家
链接：https://arxiv.org/abs/2605.24975
【41】ARCANE-PedSynth: Synthetic Multi-Pedestrian Datasets with Behavioural Crossing Annotations
标题：ARCANE-PedSynth：具有行为交叉注释的合成多行人数据集
链接：https://arxiv.org/abs/2605.24950
【42】Riemannian-Manifold Steering: Geometry-Aware Generative Autoencoders for Label-Free Steering
标题：Riemann-Manifold转向：用于无标签转向的几何感知生成自动编码器
链接：https://arxiv.org/abs/2605.24942
【43】Quaternion Self-Attention with Shared Scores
标题：具有共享分数的四元数自我注意力
链接：https://arxiv.org/abs/2605.24920
备注：26 pages, 6 figures and 15 tables. Accepted at ICML2026
【44】MVR-cache: Optimizing Semantic Caching via Multi-Vector Retrieval and Learned Prompt Segmentation
标题：MVR缓存：通过多载体检索和习得提示分割优化语义缓存
链接：https://arxiv.org/abs/2605.24914
备注：Published in ICML 2026
【45】Towards a Universal Causal Reasoner
标题：走向普遍因果推理
链接：https://arxiv.org/abs/2605.24873
【46】The Perception-Physics Paradox: Probing Scientific Alignment with TC-Bench
标题：知觉-物理悖论：与TC长凳探索科学一致性
链接：https://arxiv.org/abs/2605.24782
备注：Accepted at ICML 2026
【47】Complement Submodular Information Measures for Balanced and Robust Data Selection
标题：补充子模块信息测量以实现平衡和稳健的数据选择
链接：https://arxiv.org/abs/2605.24779
【48】Motion-Compensated Weight Compression
标题：运动补偿重量压缩
链接：https://arxiv.org/abs/2605.24754
备注：54 pages, 17 tables, 6 Figures
【49】MDIA: A Multi-Agent Diagnostic Intelligence Pipeline on HealthBench Professional
标题：MDIA：HealthBench Professional上的多代理诊断智能管道
链接：https://arxiv.org/abs/2605.24699
备注：33 pages, 10 figures
【50】Sum of Costs Diffusion with Dynamic Guidance for Motion Planning
标题：动态引导下的运动规划成本总和扩散
链接：https://arxiv.org/abs/2605.24690
备注：Accepted at the Frontiers of Optimization for Robotics Workshop at the IEEE International Conference of Robotics & Automation (ICRA), 2026
【51】WLNO: Wavelet-Laplace Neural Operator for Solving Partial Differential Equations
标题：WLNO：用于求解偏微方程的Wavelet-Lapplace神经运算符
链接：https://arxiv.org/abs/2605.24658
【52】WINO: A Weak-Form Physics Informed Neural Operator for Hyperelasticity on Variable Domains
标题：WINO：一种弱形式物理信息神经运算符，用于变量域超弹性
链接：https://arxiv.org/abs/2605.24651
【53】Demystifying the Mythos or Disrupting Bugonomics? From Zero-Day Asymmetry to Defender Remediation Throughput
标题：揭开神话还是颠覆昆虫经济学？从零日不对称到捍卫者补救能力
链接：https://arxiv.org/abs/2605.24632
【54】Beyond Generative Priors: Minority Sampling with JEPA-Guided Diffusion
标题：超越生成先验：使用JEPA引导的扩散进行少数抽样
链接：https://arxiv.org/abs/2605.24631
备注：ICML 2026, 21 pages, 9 figures
【55】LAPLEX: The FFT of Learnable Laplace Kernels
标题：LAPFLEX：可学习拉普拉斯核的快速傅里叶变换
链接：https://arxiv.org/abs/2605.24584
【56】An Effective-Rank Audit of Alignment-Induced Activation Shifts: Confound Control, Constructive Calibration, and Limits
标题：对对齐诱导的激活转变的有效等级审计：混淆控制、建设性校准和限制
链接：https://arxiv.org/abs/2605.24583
备注：18 pages, 1 figure, 21 tables. Code, data, and an immutable Zenodo archive are available at https://github.com/Nakammura/effective-rank-audit (DOI: 10.5281/zenodo.20341445)
【57】EMA: Effort Metric Attention for Anatomical Effort-Guided Human Motion Diffusion
标题：EMA：解剖学努力引导的人体运动扩散的努力指标关注
链接：https://arxiv.org/abs/2605.24566
备注：Accepted at IEEE International Conference on Automatic Face and Gesture Recognition (FG 2026)
【58】The Multilingual Curse at the Retrieval Layer: Evidence from Amharic
标题：检索层的多语言诅咒：来自阿姆哈拉语的证据
链接：https://arxiv.org/abs/2605.24556
备注：10 pages, 4 tables. Accepted to the 1st Workshop on Multilinguality in the Era of Large Language Models (MeLLM) at ACL 2026
【59】RL with Learnable Textual Feedback: A Bilevel Approach
标题：具有可学习文本反馈的RL：两层方法
链接：https://arxiv.org/abs/2605.24547
【60】Rethinking Federated Unlearning via the Lens of Memorization
标题：从小型化的角度重新思考联邦取消学习
链接：https://arxiv.org/abs/2605.24545
备注：This paper has been accepted by SIGKDD 2026
【61】What Are We Actually Decoding? Source Attribution for Non-Invasive Brain-to-Language Retrieval
标题：我们实际上在解码什么？非侵入性脑语言检索的源归因
链接：https://arxiv.org/abs/2605.24524
备注：35 pages, 7 figures, 25 tables
【62】Φ-Noise: Training-Free Temporal Video Conditioning via Phase-Based Noise Manipulation
标题：μ-Noise：通过基于阶段的噪音操纵进行免训练的时间视频调节
链接：https://arxiv.org/abs/2605.24509
备注：Under Review; 26 pages, 21 figures
【63】SPACE: Unifying Symmetric and Asymmetric Routing Problems for Generalist Neural Solver
标题：空间：通用神经求解器统一对称和非对称路由问题
链接：https://arxiv.org/abs/2605.24484
【64】Smoother Action Chunking Flow Policy via Prior-Corrected Orthogonal Trust-Region Guidance
标题：通过预先纠正的垂直信托区域指南实现更平稳的行动分块流政策
链接：https://arxiv.org/abs/2605.24433
【65】Batch Normalization Amplifies Memorization and Privacy Risks
标题：批量规范化放大了小型化和隐私风险
链接：https://arxiv.org/abs/2605.24420
【66】A Unified Python Framework for Direct PPO-based Control of AHUs with Economizer Logic and CO2-Constrained Ventilation
标题：统一的Python框架，用于对AHU进行基于PSO的直接控制，具有节能器逻辑和CO2约束通风
链接：https://arxiv.org/abs/2605.24406
备注：10 pages, 7 figures
【67】GEESE: Genotype-aware End-to-End Spatio-temporal Embedding for Behavioral Phenotyping
标题：GESSE：基因型感知的端到端时空嵌入行为表型
链接：https://arxiv.org/abs/2605.24370
【68】Interdomain Attention: Beyond Token-Level Key-Value Memory
标题：域间注意力：超越代币级关键值记忆
链接：https://arxiv.org/abs/2605.24330
【69】From One-Pass SGD to Data Reuse: Mini-Batch Scaling Laws in Sketched Linear Regression
标题：从一遍新元到数据重用：草图线性回归中的小批量缩放定律
链接：https://arxiv.org/abs/2605.24316
备注：56 pages, 3 figures
【70】PrivFusion: A Privacy-preserving Multi-Agent Framework for Harmonizing Distributed Datasets
标题：PrivFusion：一个用于协调分布式数据集的保护隐私的多代理框架
链接：https://arxiv.org/abs/2605.24249
备注：Accepted by IEEE CBMS 2026
【71】Towards Evaluation Engineering: An Empirical Study of ML Evaluation Harnesses in the Wild
标题：走向评估工程：野外ML评估工具的实证研究
链接：https://arxiv.org/abs/2605.24213
【72】Characterizing the Representational Capacity of Neural Processes
标题：描述神经过程的代表能力
链接：https://arxiv.org/abs/2605.24210
备注：To appear at ProbML/AABI 2026
【73】Single View Seafloor Recovery from Imaging Sonar via Differentiable Rendering
标题：通过差异渲染从成像声纳中进行单视图海底恢复
链接：https://arxiv.org/abs/2605.24195
【74】AvalancheBench: Evaluating Enterprise Data Agents Through Latent World Recovery
标题：AvalancheBench：通过潜在的世界复苏评估企业数据代理
链接：https://arxiv.org/abs/2605.24183
【75】MAPLE: Multi-State Aggregated Policy Evaluation for AlphaZero in Imperfect-Information Games
标题：MAPLE：不完美信息游戏中AlphaZero的多州汇总政策评估
链接：https://arxiv.org/abs/2605.24139
备注：Accepted by the IEEE Conference on Games (IEEE CoG 2026)
【76】Not All Transitions Matter: Evidence from PPO
标题：并非所有的过渡都很重要：来自PPO的证据
链接：https://arxiv.org/abs/2605.24071
备注：12 pages, 5 figures. Submitted to 2026 8th Asia Conference on Machine Learning and Computing (ACMLC 2026)
【77】Feature Lottery? A Bifurcation Theory of Concept Emergence
标题：特色彩票？概念涌现的分歧理论
链接：https://arxiv.org/abs/2605.24057
【78】Cascade-KDE: Robust Time-Series Restoration under Out-of-Distribution Impulse Corruptions
标题：Cascade-TEK：分布外脉冲腐蚀下的鲁棒时间序列恢复
链接：https://arxiv.org/abs/2605.24055
【79】Hidden-State Privacy Has an Empty Middle
标题：隐藏状态隐私中间空虚
链接：https://arxiv.org/abs/2605.24042
备注：74 pages, 61 figures
【80】Iterative Refinement Neural Operators are Learned Fixed-Point Solvers: A Principled Approach to Spectral Bias Mitigation
标题：迭代细化神经运算符是习得的定点求解器：光谱偏差缓解的原则方法
链接：https://arxiv.org/abs/2605.24041
备注：47 pages; accepted to ICML 2026 as a Spotlight
【81】MGVQ: Synergizing Multi-dimensional Sensitivity-Aware and Gradient-Hessian Fusion for Vector Quantization
标题：MGVQ：协同多维敏感性感知和干扰-黑森融合进行载体量化
链接：https://arxiv.org/abs/2605.24019
【82】Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL
标题：具有分散奖励的差异指导：走向原则性的一步生成器RL
链接：https://arxiv.org/abs/2605.24001
【83】Low-Cost Labels, Reliable Choices: Rollout-Calibrated Hyper-Heuristics for Job Shop Scheduling
标题：低成本标签，可靠的选择：针对车间调度的推出校准超启发式方法
链接：https://arxiv.org/abs/2605.23957
【84】QUIVER: A Formal Framework for Quantifying Perturbation Propagation and Bifurcation in Compound AI Systems
标题：QUIVER：量化复合人工智能系统中扰动传播和分歧的形式框架
链接：https://arxiv.org/abs/2605.23956
【85】From Accuracy to Auditability: A Survey of Determinism in Financial AI Systems
标题：从准确性到可审计性：金融人工智能系统确定性调查
链接：https://arxiv.org/abs/2605.23955
【86】Quantum Frog: Emergent Cooperation and Difficulty Scaling in a Quantized-Time Cooperative Game
标题：量子青蛙：量化时间合作博弈中的紧急合作和难度缩放
链接：https://arxiv.org/abs/2605.23930
【87】High-Risk AI Systems and the Problem of Identity in the European AI Act
标题：高风险人工智能系统和欧洲人工智能法案中的身份问题
链接：https://arxiv.org/abs/2605.23922
备注：Accepted as a non-archival paper at The 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT '26), June 25-28, 2026, Montreal, QC, Canada
【88】Accelerating Bayesian inverse design in computational fluid dynamics using neural operators
标题：基于神经算子的计算流体力学贝叶斯逆设计加速
链接：https://arxiv.org/abs/2605.26059
【89】Minimax Limits of k-Fold Cross-Validation via Majority
标题：通过多数进行k重交叉验证的极大极小限度
链接：https://arxiv.org/abs/2605.25859
【90】Efficient Benchmarking Is Just Feature Selection and Multiple Regression
标题：有效的基准测试只是特征选择和多元回归
链接：https://arxiv.org/abs/2605.25773
备注：36 pages, 27 figures
【91】Decoding Stimulus Reconstruction-Based Auditory Attention Robustly in Unbalanced EEG Datasets
标题：不平衡脑电数据集中基于刺激重建的听觉注意力解码
链接：https://arxiv.org/abs/2605.25605
【92】Nonstationary Generalized Linear Bandits with Discounted Online Mirror Descent
标题：具有折扣在线镜像下降的非平稳广义线性盗贼
链接：https://arxiv.org/abs/2605.25590
【93】Rao-Blackwellized Score Matching on Manifolds
标题：Manifolds上的Rao-Blackwelized得分匹配
链接：https://arxiv.org/abs/2605.25567
备注：22 pages, 3 figures. SPIGM @ ICML 2026
【94】Guided Flow Matching for Forward and Inverse PDE Problems with Sparse Observations: Algorithm and Theory
标题：具有稀疏观测的正向和逆偏方程问题的引导流匹配：算法和理论
链接：https://arxiv.org/abs/2605.25509
备注：50 pages, 8 figures, 4 tables
【95】Mean-Shift PCA by Knockoff Mean
标题：通过击倒平均值的均值漂移PCA
链接：https://arxiv.org/abs/2605.25460
备注：ICML 2026
【96】Data-Specific Hyper-Parameter Design: A Paradigm Shift in Reservoir Computing
标题：特定于数据的超参数设计：储层计算的范式转变
链接：https://arxiv.org/abs/2605.25221
【97】Nyström Kernel Stein Discrepancy Tests
标题：Nyström核心Stein偏差测试
链接：https://arxiv.org/abs/2605.25173
【98】Rejoinder: The ICML 2023 Ranking Experiment: Examining Author Self-Assessment in ML/AI Peer Review
标题：反驳：ICML 2023排名实验：在ML/AI同行评审中检查作者自我评估
链接：https://arxiv.org/abs/2605.25172
备注：Rejoinder to the JASA Discussion of "The ICML 2023 Ranking Experiment: Examining Author Self-Assessment in ML/AI Peer Review" (arXiv:2408.13430)
【99】Estimating Mixture Distributions via Stochastic Mirror Descent
标题：通过随机镜像下降估计混合分布
链接：https://arxiv.org/abs/2605.24929
【100】Spiking the training data to correct for test set contamination
标题：峰值训练数据以纠正测试集污染
链接：https://arxiv.org/abs/2605.24818
【101】Lifted Schrödinger Bridges for Gaussian Mixture Endpoints: Projection Gaps and Path-Space Obstructions
标题：高斯混合端点的提升薛定汉桥：投影间隙和路径空间障碍
链接：https://arxiv.org/abs/2605.24795
备注：35 pages. Submitted to a journal; comments are welcome
【102】On the Sample Complexity of Robust Binary Hypothesis Testing
标题：稳健二元假设测试的样本复杂性
链接：https://arxiv.org/abs/2605.24741
备注：Comments welcome
【103】Fermi-Dirac machines as quantizations of neurons
标题：费米-狄拉克机作为神经元的量化
链接：https://arxiv.org/abs/2605.24386
备注：87 pages, 12 figures, 2 tables
【104】A Matched Spectral Benchmark of Quantum Inspired Feature Maps
标题：量子启发特征地图的匹配光谱基准
链接：https://arxiv.org/abs/2605.24324
【105】Causality as the Statistical Conscience of Artificial Intelligence: From Pearl's Ladder to Trustworthy Machines
标题：因果关系作为人工智能的统计良心：从珀尔的梯子到值得信赖的机器
链接：https://arxiv.org/abs/2605.24076
备注：18 pages, 4 figures, 1 table
【106】Quantification of atmospheric carbon dioxide from the Geostationary Operational Environmental Satellite (GOES East)
标题：地球同步业务环境卫星（GOES East）的大气二氧化碳量化
链接：https://arxiv.org/abs/2605.23991
备注：23 pages, 7 figures, 1 table
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/196894
本文地址：http://www.python88.com/topic/196894