|
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417篇
大模型相关(56篇)
【1】A Behavioural and Representational Evaluation of Goal-Directedness in Language Model Agents
标题：语言模型主体中目标导向性的行为和表示评估
链接：https://arxiv.org/abs/2602.08964
摘要：Understanding an agent's goals helps explain and predict its behaviour, yet there is no established methodology for reliably attributing goals to agentic systems. We propose a framework for evaluating goal-directedness that integrates behavioural evaluation with interpretability-based analyses of models' internal representations. As a case study, we examine an LLM agent navigating a 2D grid world toward a goal state. Behaviourally, we evaluate the agent against an optimal policy across varying grid sizes, obstacle densities, and goal structures, finding that performance scales with task difficulty while remaining robust to difficulty-preserving transformations and complex goal structures. We then use probing methods to decode the agent's internal representations of the environment state and its multi-step action plans. We find that the LLM agent non-linearly encodes a coarse spatial map of the environment, preserving approximate task-relevant cues about its position and the goal location; that its actions are broadly consistent with these internal representations; and that reasoning reorganises them, shifting from broader environment structural cues toward information supporting immediate action selection. Our findings support the view that introspective examination is required beyond behavioural evaluations to characterise how agents represent and pursue their objectives.
【2】GSS: Gated Subspace Steering for Selective Memorization Mitigation in LLMs
标题：GSS：用于LLM中选择性子空间控制的选择性子空间控制
链接：https://arxiv.org/abs/2602.08901
备注：34 pages, 12 figures
摘要：Large language models (LLMs) can memorize and reproduce training sequences verbatim -- a tendency that undermines both generalization and privacy. Existing mitigation methods apply interventions uniformly, degrading performance on the majority of tokens that generalize normally. We show empirically that memorization is sparse, intermittent, and token-conditioned, suggesting that effective mitigation requires context-aware intervention rather than static parameter modification. To this end, we propose a novel and effective selective memorization mitigation method -- Gated Subspace Steering (GSS), which decomposes intervention into a probe (detecting memorization-relevant activations) and a steer (applying targeted correction only when the probe exceeds a threshold). The optimal probe-steer pair emerges from a principled optimization framework based on optimal subspace steering. Experiments on four benchmarks show GSS matches or exceeds state-of-the-art memorization reduction while requiring $100-1000 \times$ less compute than optimization-based alternatives. Furthermore, we provide new theoretical insights into the geometry of memorization in neural representations.
【3】AnomSeer: Reinforcing Multimodal LLMs to Reason for Time-Series Anomaly Detection
标题：AnomSeer：加强多峰LLM以推理时间序列异常检测
链接：https://arxiv.org/abs/2602.08868
备注：Preprint
摘要：Time-series anomaly detection (TSAD) with multimodal large language models (MLLMs) is an emerging area, yet a persistent challenge remains: MLLMs rely on coarse time-series heuristics but struggle with multi-dimensional, detailed reasoning, which is vital for understanding complex time-series data. We present AnomSeer to address this by reinforcing the model to ground its reasoning in precise, structural details of time series, unifying anomaly classification, localization, and explanation. At its core, an expert chain-of-thought trace is generated to provide a verifiable, fine-grained reasoning from classical analyses (e.g., statistical measures, frequency transforms). Building on this, we propose a novel time-series grounded policy optimization (TimerPO) that incorporates two additional components beyond standard reinforcement learning: a time-series grounded advantage based on optimal transport and an orthogonal projection to ensure this auxiliary granular signal does not interfere with the primary detection objective. Across diverse anomaly scenarios, AnomSeer, with Qwen2.5-VL-3B/7B-Instruct, outperforms larger commercial baselines (e.g., GPT-4o) in classification and localization accuracy, particularly on point- and frequency-driven exceptions. Moreover, it produces plausible time-series reasoning traces that support its conclusions.
【4】Dr. MAS: Stable Reinforcement Learning for Multi-Agent LLM Systems
标题：MAS博士：多智能体LLM系统的稳定强化学习
链接：https://arxiv.org/abs/2602.08847
备注：Preprint
摘要 ：Multi-agent LLM systems enable advanced reasoning and tool use via role specialization, yet reliable reinforcement learning (RL) post-training for such systems remains difficult. In this work, we theoretically pinpoint a key reason for training instability when extending group-based RL to multi-agent LLM systems. We show that under GRPO-style optimization, a global normalization baseline may deviate from diverse agents' reward distributions, which ultimately leads to gradient-norm instability. Based on this finding, we propose Dr. MAS, a simple and stable RL training recipe for multi-agent LLM systems. Dr. MAS uses an agent-wise remedy: normalizing advantages per agent using each agent's own reward statistics, which calibrates gradient scales and dramatically stabilizes training, both theoretically and empirically. Beyond the algorithm, Dr. MAS provides an end-to-end RL training framework for multi-agent LLM systems, supporting scalable orchestration, flexible per-agent LLM serving and optimization configs, and shared resource scheduling of LLM actor backends. We evaluate Dr. MAS on multi-agent math reasoning and multi-turn search benchmarks using Qwen2.5 and Qwen3 series models. Dr. MAS achieves clear gains over vanilla GRPO (e.g., +5.6\% avg@16 and +4.6\% pass@16 on math, and +15.2\% avg@16 and +13.1\% pass@16 on search) while largely eliminating gradient spikes. Moreover, it remains highly effective under heterogeneous agent-model assignments while improving efficiency.
【5】AMEM4Rec: Leveraging Cross-User Similarity for Memory Evolution in Agentic LLM Recommenders
标题：AEM 4Rec：在大型LLM推荐中利用跨用户相似性实现内存进化
链接：https://arxiv.org/abs/2602.08837
摘要：Agentic systems powered by Large Language Models (LLMs) have shown strong potential in recommender systems but remain hindered by several challenges. Fine-tuning LLMs is parameter-inefficient, and prompt-based agentic reasoning is limited by context length and hallucination risk. Moreover, existing agentic recommendation systems predominantly leverages semantic knowledge while neglecting the collaborative filtering (CF) signals essential for implicit preference modeling. To address these limitations, we propose AMEM4Rec, an agentic LLM-based recommender that learns collaborative signals in an end-to-end manner through cross-user memory evolution. AMEM4Rec stores abstract user behavior patterns from user histories in a global memory pool. Within this pool, memories are linked to similar existing ones and iteratively evolved to reinforce shared cross-user patterns, enabling the system to become aware of CF signals without relying on a pre-trained CF model. Extensive experiments on Amazon and MIND datasets show that AMEM4Rec consistently outperforms state-of-the-art LLM-based recommenders, demonstrating the effectiveness of evolving memory-guided collaborative filtering.
【6】FlexMoRE: A Flexible Mixture of Rank-heterogeneous Experts for Efficient Federatedly-trained Large Language Models
标题：FlexMoRE：一种灵活的等级异类专家混合，用于高效的联邦训练大型语言模型
链接：https://arxiv.org/abs/2602.08818
摘要：Recent advances in mixture-of-experts architectures have shown that individual experts models can be trained federatedly, i.e., in isolation from other experts by using a common base model to facilitate coordination. However, we hypothesize that full-sized experts may not be necessary for all domains and that instead low-rank adapters may be sufficient. Here, we introduce FlexMoRE, a Flexible Mixture of Rank-heterogenous Experts, which may be either full-sized experts or adapters of a suitable rank. We systematically investigate the trade-off between expert rank and downstream task performance by evaluating $6$ experts with ranks $2^0$ to $2^{14}$ resulting in experiments covering 150 mixtures (96 with 2 experts, 54 with 7 experts) that are evaluated across $120$ tasks. For our experiments, we build on FlexOlmo and turn its pre-trained experts into low-rank versions. Our regression analysis from expert rank to downstream task performance reveals that the best-performing rank is substantially higher for reasoning-heavy benchmarks than for knowledge-heavy benchmarks. These findings on rank sensitivity come with direct implications for memory efficiency: Using optimal ranks, FlexMoRE yields improved downstream task performance (average score $47.18$) compared to the baseline FlexOlmo-style mixture of full-sized experts (average score $45.46$) at less than one third the parameters ($10.75$B for FlexMoRE vs. $33.27$B for FlexOlmo). All code will be made available.
【7】How2Everything: Mining the Web for How-To Procedures to Evaluate and Improve LLMs
标题：How 2Everything：挖掘网络中的操作方法来评估和改进LLM
链接：https://arxiv.org/abs/2602.08808
备注：53 pages, 22 figures
摘要：Generating step-by-step "how-to" procedures is a key LLM capability: how-to advice is commonly requested in chatbots, and step-by-step planning is critical for reasoning over complex tasks. Yet, measuring and improving procedural validity at scale on real-world tasks remains challenging and understudied. To address this, we introduce How2Everything, a scalable framework to evaluate and improve goal-conditioned procedure generation. Our framework includes How2Mine, which mines 351K procedures from 980K web pages across 14 topics and readily scales to larger corpora. From this pool we build How2Bench, a 7K-example evaluation set balanced across topics. To reliably score model outputs, we develop How2Score, an evaluation protocol that uses an LLM judge to detect whether a generation contains any critical failure that would prevent achieving the goal. For low-cost, reproducible evaluation, we distill a frontier model into an open 8B model, achieving 80.5% agreement with human annotators. How2Bench reveals clear scaling trends across model sizes and training stages, providing signal early in pretraining. Finally, RL using How2Score as a reward improves performance on How2Bench by >10 points across three models without systematic regressions on standard benchmarks, with gains robust to superficial source-document memorization or format compliance. Taken together, How2Everything shows how pretraining web data can support a closed loop of capability evaluation and improvement at scale.
【8】QUOKA: Query-Oriented KV Selection For Efficient LLM Prefill
标题：QUOKA：面向查询的KN选择，以实现高效的LLM预编写
链接：https://arxiv.org/abs/2602.08722
摘要 ：We present QUOKA: Query-oriented KV selection for efficient attention, a training-free and hardware agnostic sparse attention algorithm for accelerating transformer inference under chunked prefill. While many queries focus on a smaller group of keys in the attention operator, we observe that queries with low cosine similarity with respect to the mean query interact more strongly with more keys and have the greatest contribution to final attention logits. By prioritizing these low cosine similarity queries, the behavior of full attention during the prefill stage can be closely approximated. QUOKA leverages this observation, accelerating attention by (1) first retaining a small set of representative queries and (2) then subselectin the keys most aligned with those queries. Through experiments on Needle-In-A-Haystack, LongBench, RULER, and Math500, we show that, while realizing a 3x reduction in time-to-first-token, 5x speedup in attention on Nvidia GPUs and up to nearly a 7x speedup on Intel Xeon CPUs, QUOKA achieves near-baseline accuracy, utilizing 88% fewer key-value pairs per attention evaluation.
【9】Reasoning aligns language models to human cognition
标题：推理使语言模型与人类认知保持一致
链接：https://arxiv.org/abs/2602.08693
备注：38 pages, 4 main figures, multiple appendix figures
摘要：Do language models make decisions under uncertainty like humans do, and what role does chain-of-thought (CoT) reasoning play in the underlying decision process? We introduce an active probabilistic reasoning task that cleanly separates sampling (actively acquiring evidence) from inference (integrating evidence toward a decision). Benchmarking humans and a broad set of contemporary large language models against near-optimal reference policies reveals a consistent pattern: extended reasoning is the key determinant of strong performance, driving large gains in inference and producing belief trajectories that become strikingly human-like, while yielding only modest improvements in active sampling. To explain these differences, we fit a mechanistic model that captures systematic deviations from optimal behavior via four interpretable latent variables: memory, strategy, choice bias, and occlusion awareness. This model places humans and models in a shared low-dimensional cognitive space, reproduces behavioral signatures across agents, and shows how chain-of-thought shifts language models toward human-like regimes of evidence accumulation and belief-to-choice mapping, tightening alignment in inference while leaving a persistent gap in information acquisition.
【10】Learning to Judge: LLMs Designing and Applying Evaluation Rubrics
标题：学习判断：法学硕士设计和应用评估指标
链接：https://arxiv.org/abs/2602.08672
备注：Accepted at EACL 2026 Findings
摘要：Large language models (LLMs) are increasingly used as evaluators for natural language generation, applying human-defined rubrics to assess system outputs. However, human rubrics are often static and misaligned with how models internally represent language quality. We introduce GER-Eval (Generating Evaluation Rubrics for Evaluation) to investigate whether LLMs can design and apply their own evaluation rubrics. We evaluate the semantic coherence and scoring reliability of LLM-defined criteria and their alignment with human criteria. LLMs reliably generate interpretable and task-aware evaluation dimensions and apply them consistently within models, but their scoring reliability degrades in factual and knowledge-intensive settings. Closed-source models such as GPT-4o achieve higher agreement and cross-model generalization than open-weight models such as Llama. Our findings position evaluation as a learned linguistic capability of LLMs, consistent within models but fragmented across them, and call for new methods that jointly model human and LLM evaluative language to improve reliability and interpretability.
【11】Sparse Models, Sparse Safety: Unsafe Routes in Mixture-of-Experts LLMs
标题：稀疏模型，稀疏安全：混合专家LL中的不安全路线
链接：https://arxiv.org/abs/2602.08621
摘要：By introducing routers to selectively activate experts in Transformer layers, the mixture-of-experts (MoE) architecture significantly reduces computational costs in large language models (LLMs) while maintaining competitive performance, especially for models with massive parameters. However, prior work has largely focused on utility and efficiency, leaving the safety risks associated with this sparse architecture underexplored. In this work, we show that the safety of MoE LLMs is as sparse as their architecture by discovering unsafe routes: routing configurations that, once activated, convert safe outputs into harmful ones. Specifically, we first introduce the Router Safety importance score (RoSais) to quantify the safety criticality of each layer's router. Manipulation of only the high-RoSais router(s) can flip the default route into an unsafe one. For instance, on JailbreakBench, masking 5 routers in DeepSeek-V2-Lite increases attack success rate (ASR) by over 4$\times$ to 0.79, highlighting an inherent risk that router manipulation may naturally occur in MoE LLMs. We further propose a Fine-grained token-layer-wise Stochastic Optimization framework to discover more concrete Unsafe Routes (F-SOUR), which explicitly considers the sequentiality and dynamics of input tokens. Across four representative MoE LLM families, F-SOUR achieves an average ASR of 0.90 and 0.98 on JailbreakBench and AdvBench, respectively. Finally, we outline defensive perspectives, including safety-aware route disabling and router training, as promising directions to safeguard MoE LLMs. We hope our work can inform future red-teaming and safeguarding of MoE LLMs. Our code is provided in https://github.com/TrustAIRLab/UnsafeMoE.
【12】Stateless Yet Not Forgetful: Implicit Memory as a Hidden Channel in LLMs
标题：无状态但不健忘：内隐记忆作为LLM中的隐藏渠道
链接：https://arxiv.org/abs/2602.08563
备注：Accepted at IEEE SaTML 2026
摘要 ：Large language models (LLMs) are commonly treated as stateless: once an interaction ends, no information is assumed to persist unless it is explicitly stored and re-supplied. We challenge this assumption by introducing implicit memory-the ability of a model to carry state across otherwise independent interactions by encoding information in its own outputs and later recovering it when those outputs are reintroduced as input. This mechanism does not require any explicit memory module, yet it creates a persistent information channel across inference requests. As a concrete demonstration, we introduce a new class of temporal backdoors, which we call time bombs. Unlike conventional backdoors that activate on a single trigger input, time bombs activate only after a sequence of interactions satisfies hidden conditions accumulated via implicit memory. We show that such behavior can be induced today through straightforward prompting or fine-tuning. Beyond this case study, we analyze broader implications of implicit memory, including covert inter-agent communication, benchmark contamination, targeted manipulation, and training-data poisoning. Finally, we discuss detection challenges and outline directions for stress-testing and evaluation, with the goal of anticipating and controlling future developments. To promote future research, we release code and data at: https://github.com/microsoft/implicitMemory.
【13】Reinforcement Inference: Leveraging Uncertainty for Self-Correcting Language Model Reasoning
标题：强化推理：利用不确定性进行自我纠正语言模型推理
链接：https://arxiv.org/abs/2602.08520
摘要：Modern large language models (LLMs) are often evaluated and deployed under a \emph{one-shot, greedy} inference protocol, especially in professional settings that require deterministic behavior. This regime can systematically under-estimate a fixed model's true capability: many errors arise not from missing knowledge, but from premature commitment under internal ambiguity. We introduce \emph{Reinforcement Inference}, an entropy-aware inference-time control strategy that uses the model's own uncertainty to selectively invoke a second, more deliberate reasoning attempt, enabling stronger performance \emph{without any retraining}. On 12,032 MMLU-Pro questions across 14 subjects, using DeepSeek-v3.2 with deterministic decoding in a zero-shot setting, Reinforcement Inference improves accuracy from 60.72\% to 84.03\%, while only incurring 61.06\% additional inference calls. A 100\% re-asking ablation reaches 84.35\%, indicating that uncertainty-aware selection captures most of the attainable improvement with substantially less compute. Moreover, a \emph{prompt-only} ablation underperforms the baseline, suggesting that the gains are not explained by generic `` your output had high entropy, think step-by-step'' prompting alone. Beyond providing a practical inference-time upgrade, our results suggest a broader \emph{entropy-aware} paradigm for measuring and expanding model capability: because modern decoder-based models generate outputs autoregressively, entropy and related confidence measures arise naturally as first-class control signals during generation. The resulting gap between one-pass greedy inference and uncertainty-conditioned deliberation offers a diagnostic lens on an LLM's latent reasoning horizon and motivates future training objectives that explicitly constrain correctness--confidence alignment.
【14】Learning Self-Correction in Vision-Language Models via Rollout Augmentation
标题：通过推出增强在视觉语言模型中学习自我纠正
链接：https://arxiv.org/abs/2602.08503
备注：17 pages
摘要：Self-correction is essential for solving complex reasoning problems in vision-language models (VLMs). However, existing reinforcement learning (RL) methods struggle to learn it, as effective self-correction behaviors emerge only rarely, making learning signals extremely sparse. To address this challenge, we propose correction-specific rollouts (Octopus), an RL rollout augmentation framework that synthesizes dense self-correction examples by recombining existing rollouts. This augmentation simultaneously improves sample efficiency due to rollout reuse and stabilizes RL optimization through balanced supervision. Furthermore, we introduce a response-masking strategy that decouples self-correction from direct reasoning, avoiding signal conflicts and enabling both behaviors to be learned effectively. Building on this, we introduce Octopus-8B, a reasoning VLM with controllable self-correction capability. Across 7 benchmarks, it achieves SoTA performance among open-source VLMs, outperforming the best RLVR baseline by 1.0 score while requiring only $0.72\times$ training time per step.
【15】Modalities, a PyTorch-native Framework For Large-scale LLM Training and Research
标题：Modities，一个用于大规模LLM训练和研究的PyTorch原生框架
链接：https://arxiv.org/abs/2602.08387
摘要：Today's LLM (pre-) training and research workflows typically allocate a significant amount of compute to large-scale ablation studies. Despite the substantial compute costs of these ablations, existing open-source frameworks provide limited tooling for these experiments, often forcing researchers to write their own wrappers and scripts. We propose Modalities, an end-to-end PyTorch-native framework that integrates data-driven LLM research with large-scale model training from two angles. Firstly, by integrating state-of-the-art parallelization strategies, it enables both efficient pretraining and systematic ablations at trillion-token and billion-parameter scale. Secondly, Modalities adopts modular design with declarative, self-contained configuration, enabling reproducibility and extensibility levels that are difficult to achieve out-of-the-box with existing LLM training frameworks.
【16】The Chicken and Egg Dilemma: Co-optimizing Data and Model Configurations for LLMs
标题：鸡和蛋的困境：LLM的数据和模型协同优化
链接：https://arxiv.org/abs/2602.08351
摘要 ：Co-optimizing data and model configurations for training LLMs presents a classic chicken-and-egg dilemma: The best training data configuration (e.g., data mixture) for a downstream task depends on the chosen model configuration (e.g., model architecture), and vice versa. However, jointly optimizing both data and model configurations is often deemed intractable, and existing methods focus on either data or model optimization without considering their interaction. We introduce JoBS, an approach that uses a scaling-law-inspired performance predictor to aid Bayesian optimization (BO) in jointly optimizing LLM training data and model configurations efficiently. JoBS allocates a portion of the optimization budget to learn an LLM performance predictor that predicts how promising a training configuration is from a small number of training steps. The remaining budget is used to perform BO entirely with the predictor, effectively amortizing the cost of running full-training runs. We study JoBS's average regret and devise the optimal budget allocation to minimize regret. JoBS outperforms existing multi-fidelity BO baselines, as well as data and model optimization approaches across diverse LLM tasks under the same optimization budget.
【17】Towards Efficient Large Language Reasoning Models via Extreme-Ratio Chain-of-Thought Compression
标题：基于极值比思想链压缩的高效大型语言推理模型
链接：https://arxiv.org/abs/2602.08324
备注：15 pages, 7 figures
摘要：Chain-of-Thought (CoT) reasoning successfully enhances the reasoning capabilities of Large Language Models (LLMs), yet it incurs substantial computational overhead for inference. Existing CoT compression methods often suffer from a critical loss of logical fidelity at high compression ratios, resulting in significant performance degradation. To achieve high-fidelity, fast reasoning, we propose a novel EXTreme-RAtio Chain-of-Thought Compression framework, termed Extra-CoT, which aggressively reduces the token budget while preserving answer accuracy. To generate reliable, high-fidelity supervision, we first train a dedicated semantically-preserved compressor on mathematical CoT data with fine-grained annotations. An LLM is then fine-tuned on these compressed pairs via a mixed-ratio supervised fine-tuning (SFT), teaching it to follow a spectrum of compression budgets and providing a stable initialization for reinforcement learning (RL). We further propose Constrained and Hierarchical Ratio Policy Optimization (CHRPO) to explicitly incentivize question-solving ability under lower budgets by a hierarchical reward. Experiments on three mathematical reasoning benchmarks show the superiority of Extra-CoT. For example, on MATH-500 using Qwen3-1.7B, Extra-CoT achieves over 73\% token reduction with an accuracy improvement of 0.6\%, significantly outperforming state-of-the-art (SOTA) methods.
【18】Linearization Explains Fine-Tuning in Large Language Models
标题：线性化解释大型语言模型中的微调
链接：https://arxiv.org/abs/2602.08239
摘要：Parameter-Efficient Fine-Tuning (PEFT) is a popular class of techniques that strive to adapt large models in a scalable and resource-efficient manner. Yet, the mechanisms underlying their training performance and generalization remain underexplored. In this paper, we provide several insights into such fine-tuning through the lens of linearization. Fine-tuned models are often implicitly encouraged to remain close to the pretrained model. By making this explicit, using an Euclidean distance inductive bias in parameter space, we show that fine-tuning dynamics become equivalent to learning with the positive-definite neural tangent kernel (NTK). We specifically analyze how close the fully linear and the linearized fine-tuning optimizations are, based on the strength of the regularization. This allows us to be pragmatic about how good a model linearization is when fine-tuning large language models (LLMs). When linearization is a good model, our findings reveal a strong correlation between the eigenvalue spectrum of the NTK and the performance of model adaptation. Motivated by this, we give spectral perturbation bounds on the NTK induced by the choice of layers selected for fine-tuning. We empirically validate our theory on Low Rank Adaptation (LoRA) on LLMs. These insights not only characterize fine-tuning but also have the potential to enhance PEFT techniques, paving the way to better informed and more nimble adaptation in LLMs.
【19】InfiCoEvalChain: A Blockchain-Based Decentralized Framework for Collaborative LLM Evaluation
标题：InfiCoEvalChain：一个基于区块链的去中心化LLM协作评估框架
链接：https://arxiv.org/abs/2602.08229
摘要：The rapid advancement of large language models (LLMs) demands increasingly reliable evaluation, yet current centralized evaluation suffers from opacity, overfitting, and hardware-induced variance. Our empirical analysis reveals an alarming inconsistency in existing evaluations: the standard deviation across ten repeated runs of a single model on HumanEval (1.67) actually exceeds the performance gap among the top-10 models on the official leaderboard (0.91), rendering current rankings statistically precarious. To mitigate these instabilities, we propose a decentralized evaluation framework that enables hardware and parameter diversity through large-scale benchmarking across heterogeneous compute nodes. By leveraging the blockchain-based protocol, the framework incentivizes global contributors to act as independent validators, using a robust reward system to ensure evaluation integrity and discourage dishonest participation. This collective verification transforms evaluation from a "centralized black box" into a "decentralized endorsement" where multi-party consensus and diverse inference environments yield a more stable, representative metric. Experimental results demonstrate that the decentralized evaluation framework reduces the standard deviation across ten runs on the same model to 0.28. This significant improvement over conventional frameworks ensures higher statistical confidence in model rankings. We have completely implemented this platform and will soon release it to the community.
【20】DrugR: Optimizing Molecular Drugs through LLM-based Explicit Reasoning
标题：DrugR：通过基于LLM的显式推理优化分子药物
链接：https://arxiv.org/abs/2602.08213
【21】Spherical Steering: Geometry-Aware Activation Rotation for Language Models
标题：球形转向：语言模型的几何感知激活轮换
链接：https://arxiv.org/abs/2602.08169
备注：The code is at: https://github.com/chili-lab/Spherical-Steering
【22】The Confidence Manifold: Geometric Structure of Correctness Representations in Language Models
标题：置信度Manifold：语言模型中正确性表示的几何结构
链接：https://arxiv.org/abs/2602.08159
【23】Gender and Race Bias in Consumer Product Recommendations by Large Language Models
标题：大型语言模型在消费品推荐中的性别和种族偏见
链接：https://arxiv.org/abs/2602.08124
备注：Accepted at the 39th International Conference on Advanced Information Networking and Applications (AINA 2025)
【24】Online Domain-aware LLM Decoding for Continual Domain Evolution
标题：在线领域感知LLM解码以实现连续领域进化
链接：https://arxiv.org/abs/2602.08088
【25】Enhancing Bandit Algorithms with LLMs for Time-varying User Preferences in Streaming Recommendations
标题：使用LLM增强Bandit算法，以满足流媒体推荐中的时变用户偏好
链接：https://arxiv.org/abs/2602.08067
【26】Efficient and Adaptable Detection of Malicious LLM Prompts via Bootstrap Aggregation
标题：通过Bootstrap聚合高效且适应性地检测恶意LLM预测
链接：https://arxiv.org/abs/2602.08062
【27】Compiler-Assisted Speculative Sampling for Accelerated LLM Inference on Heterogeneous Edge Devices
标题：在异类边缘设备上加速LLM推理的操作员辅助推测采样
链接：https://arxiv.org/abs/2602.08060
备注：Accepted to AccML@HiPEAC 2026
【28】FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token Merging
标题：Flash VID：通过免训练的基于树的时空令牌合并的高效视频大型语言模型
链接：https://arxiv.org/abs/2602.08024
备注：Accepted by ICLR 2026 (Oral)
【29】Don't Always Pick the Highest-Performing Model: An Information Theoretic View of LLM Ensemble Selection
标题：不要总是选择表现最好的模式--从信息论的角度看LLM课程选择
链接：https://arxiv.org/abs/2602.08003
【30】SparseEval: Efficient Evaluation of Large Language Models by Sparse Optimization
标题：SparseEval：通过稀疏优化有效评估大型语言模型
链接：https://arxiv.org/abs/2602.07909
备注：ICLR2026
【31】Adaptive Acquisition Selection for Bayesian Optimization with Large Language Models
标题：大型语言模型的Bayesian优化的自适应获取选择
链接
：https://arxiv.org/abs/2602.07904
【32】rePIRL: Learn PRM with Inverse RL for LLM Reasoning
标题：rePIRL：使用反向RL学习PRM以进行LLM推理
链接：https://arxiv.org/abs/2602.07832
【33】CausalTAD: Injecting Causal Knowledge into Large Language Models for Tabular Anomaly Detection
标题：因果关系：将因果知识注入大型语言模型以进行表格异常检测
链接：https://arxiv.org/abs/2602.07798
【34】MaD-Mix: Multi-Modal Data Mixtures via Latent Space Coupling for Vision-Language Model Training
标题：MaD-Mix：通过潜在空间耦合进行多模式数据混合，用于视觉语言模型训练
链接：https://arxiv.org/abs/2602.07790
【35】Do We Need Adam? Surprisingly Strong and Sparse Reinforcement Learning with SGD in LLMs
标题：我们需要亚当吗？LLM中使用Singapore的惊人强大和稀疏强化学习
链接：https://arxiv.org/abs/2602.07729
【36】ParisKV: Fast and Drift-Robust KV-Cache Retrieval for Long-Context LLMs
标题：ParisKN：用于长上下文LLM的快速且漂移稳健的NV缓存检索
链接：https://arxiv.org/abs/2602.07721
备注：25 pages, 16 figures. Under review
【37】Efficient Table Retrieval and Understanding with Multimodal Large Language Models
标题：使用多模式大型语言模型高效的表检索和理解
链接：https://arxiv.org/abs/2602.07642
备注：Published at EACL 2026 Findings
【38】Astro: Activation-guided Structured Regularization for Outlier-Robust LLM Post-Training Quantization
标题：Astro：用于离群稳健LLM训练后量化的激活引导结构化
链接：https://arxiv.org/abs/2602.07596
【39】Improving Variable-Length Generation in Diffusion Language Models via Length Regularization
标题：利用长度正则化改进扩散语言模型中的变长生成
链接：https://arxiv.org/abs/2602.07546
备注：diffusion language models
【40】LLM-Guided Diagnostic Evidence Alignment for Medical Vision-Language Pretraining under Limited Pairing
标题：有限配对下医学视觉语言预训练的LLM引导诊断证据对齐
链接：https://arxiv.org/abs/2602.07540
【41】Scout Before You Attend: Sketch-and-Walk Sparse Attention for Efficient LLM Inference
标题：参加前先侦察：草图和步行分散注意力以实现高效的LLM推理
链接：https://arxiv.org/abs/2602.07397
【42】Efficient Post-Training Pruning of Large Language Models with Statistical Correction
标题：具有统计纠正的大型语言模型的有效训练后修剪
链接：https://arxiv.org/abs/2602.07375
备注：11 pages, 2 figures, 5 tables
【43】Controllable Value Alignment in Large Language Models through Neuron-Level Editing
标题：通过神经元级编辑实现大型语言模型中的可控值对齐
链接：https://arxiv.org/abs/2602.07356
【44】Revisiting Robustness for LLM Safety Alignment via Selective Geometry Control
标题：通过选择性几何控制重新审视LLM安全对齐的鲁棒性
链接：https://arxiv.org/abs/2602.07340
【45】Steer2Adapt: Dynamically Composing Steering Vectors Elicits Efficient Adaptation of LLMs
标题：Steer 2Adapt：动态构成引导载体实现LLM的高效自适应
链接：https://arxiv.org/abs/2602.07276
【46】Is there "Secret Sauce'' in Large Language Model Development?
链接：https://arxiv.org/abs/2602.07238
【47】ArcMark: Multi-bit LLM Watermark via Optimal Transport
标题：ArcMark：通过最佳传输的多位LLM水印
链接：https://arxiv.org/abs/2602.07235
【48】Adaptive Retrieval helps Reasoning in LLMs -- but mostly if it's not used
标题：自适应检索有助于LLM中的推理--但主要是在不使用时
链接：https://arxiv.org/abs/2602.07213
备注：Eurips Workshop on Principles of Generative Modeling (PriGM)
【49】The Optimal Token Baseline: Variance Reduction for Long-Horizon LLM-RL
标题：最佳代币基线：长期LLM-RL的方差降低
链接：https://arxiv.org/abs/2602.07078
【50】Neural Sentinel: Unified Vision Language Model (VLM) for License Plate Recognition with Human-in-the-Loop Continual Learning
标题：神经哨兵：用于车牌识别的统一视觉语言模型（VLM），采用人在环连续学习
链接：https://arxiv.org/abs/2602.07051
【51】DLLM-Searcher: Adapting Diffusion Large Language Model for Search Agents
标题：DLLM-Searcher：适应搜索代理的扩散大型语言模型
链接：https://arxiv.org/abs/2602.07035
【52】Neural Sabermetrics with World Model: Play-by-play Predictive Modeling with Large Language Model
标题：具有世界模型的神经Sabermetrics：具有大型语言模型的逐场预测建模
链接：https://arxiv.org/abs/2602.07030
【53】Fair Context Learning for Evidence-Balanced Test-Time Adaptation in Vision-Language Models
标题：视觉语言模型中证据平衡测试时适应的公平上下文学习
链接：https://arxiv.org/abs/2602.07027
【54】Steering to Say No: Configurable Refusal via Activation Steering in Vision Language Models
标题：转向说不：视觉语言模型中通过激活转向的可配置拒绝
链接：https://arxiv.org/abs/2602.07013
【55】Does Visual Rendering Bypass Tokenization? Investigating Script-Tokenizer Misalignment in Pixel-Based Language Models
标题：视觉渲染是否绕过代币化？调查基于像素的语言模型中的脚本与令牌器不一致
链接：https://arxiv.org/abs/2602.06973
备注：Submitted to ARR January
【56】Linguistic properties and model scale in brain encoding: from small to compressed language models
标题：大脑编码中的语言属性和模型规模：从小型语言模型到压缩语言模型
链接：https://arxiv.org/abs/2602.07547
备注：40 pages, 33 figures
Graph相关(图学习|图神经网络|图优化等)(18篇)
【1】Rethinking Graph Generalization through the Lens of Sharpness-Aware Minimization
标题：从敏锐度最小化的角度重新思考图形综合
链接：https://arxiv.org/abs/2602.08855
【2】A Graphop Analysis of Graph Neural Networks on Sparse Graphs: Generalization and Universal Approximation
标题：稀疏图上图神经网络的图分析：推广和普适逼近
链接：https://arxiv.org/abs/2602.08785
【3】HoGS: Homophily-Oriented Graph Synthesis for Local Differentially Private GNN Training
标题：HoGS：面向同性恋的图合成，用于本地差异私人GNN训练
链接：https://arxiv.org/abs/2602.08762
【4】Retrieval Pivot Attacks in Hybrid RAG: Measuring and Mitigating Amplified Leakage from Vector Seeds to Graph Expansion
标题：混合RAG中的检索枢纽攻击：测量和缓解从载体种子到图扩展的放大泄漏
链接：https://arxiv.org/abs/2602.08668
备注：18 pages, 5 figures
【5】Enhancing Genetic Algorithms with Graph Neural Networks: A Timetabling Case Study
标题：用图神经网络增强遗传算法：时间安排案例研究
链接：https://arxiv.org/abs/2602.08619
备注：Paper accepted to the International Conference on Applications of Evolutionary Computation (EvoApplications) 2026
【6】TFMLinker: Universal Link Predictor by Graph In-Context Learning with Tabular Foundation Models
标题：TFMLinker：通过使用表格基础模型的图内上下文学习实现的通用链接预测器
链接：https://arxiv.org/abs/2602.08592
【7】Incremental (k, z)-Clustering on Graphs
标题：图上的增量（k，z）-聚集
链接：https://arxiv.org/abs/2602.08542
备注：Abstract shortened to meet arXiv limits
【8】Bridging Academia and Industry: A Comprehensive Benchmark for Attributed Graph Clustering
标题：学术界和工业界的桥梁：归因图集群的综合基准
链接：https://arxiv.org/abs/2602.08519
【9】USBD: Universal Structural Basis Distillation for Source-Free Graph Domain Adaptation
标题：USBD：用于无源图域自适应的通用结构基蒸馏
链接：https://arxiv.org/abs/2602.08431
【10】Drop the mask! GAMM-A Taxonomy for Graph Attributes Missing Mechanisms
标题：摘下面具！GAMM-缺失机制的图属性分类
链接：https://arxiv.org/abs/2602.08407
【11】TAAM:Inductive Graph-Class Incremental Learning with Task-Aware Adaptive Modulation
标题：TAAM：具有任务感知自适应调制的归纳图形类增量学习
链接：https://arxiv.org/abs/2602.08036
【12】Quantifying Explanation Quality in Graph Neural Networks using Out-of-Distribution Generalization
标题：基于分布外泛化的图神经网络解释质量量化
链接：https://arxiv.org/abs/2602.07708
【13】GraphAgents: Knowledge Graph-Guided Agentic AI for Cross-Domain Materials Design
标题
：GraphAgents：用于跨领域材料设计的知识图引导的抽象人工智能
链接：https://arxiv.org/abs/2602.07491
【14】Bipartite Graph Attention-based Clustering for Large-scale scRNA-seq Data
标题：大规模scRN-seq数据的基于二部图注意力的集群
链接：https://arxiv.org/abs/2602.07475
【15】Graph homophily booster: Reimagining the role of discrete features in heterophilic graph learning
标题：图同质性助推器：重新想象离散特征在异质图学习中的作用
链接：https://arxiv.org/abs/2602.07256
备注：ICLR 2026
【16】Pro-ZD: A Transferable Graph Neural Network Approach for Proactive Zero-Day Threats Mitigation
标题：Pro-ZZ：一种用于主动缓解零日威胁的可转移图神经网络方法
链接：https://arxiv.org/abs/2602.07073
【17】Graph-Based Nearest-Neighbor Search without the Spread
标题：基于图的无扩散的最近邻搜索
链接：https://arxiv.org/abs/2602.06633
【18】Graph-based Semi-Supervised Learning via Maximum Discrimination
标题：通过最大区分的基于图的半监督学习
链接：https://arxiv.org/abs/2602.08042
Transformer(15篇)
【1】Diffusion-Inspired Reconfiguration of Transformers for Uncertainty Calibration
标题：用于不确定度校准的Transformer扩散启发重新配置
链接：https://arxiv.org/abs/2602.08920
【2】Understanding Dynamic Compute Allocation in Recurrent Transformers
标题：理解递归Transformer中的动态计算分配
链接：https://arxiv.org/abs/2602.08864
【3】Discovering Interpretable Algorithms by Decompiling Transformers to RASP
标题：通过将转换器反编译为RASP来发现可解释的算法
链接：https://arxiv.org/abs/2602.08857
备注：101 pages, 92 figures
【4】Central Dogma Transformer II: An AI Microscope for Understanding Cellular Regulatory Mechanisms
标题：中央教条Transformer II：了解细胞调节机制的人工智能显微镜
链接：https://arxiv.org/abs/2602.08751
备注：20 pages, 6 figures
【5】Trapped by simplicity: When Transformers fail to learn from noisy features
标题：被简单困住：当Transformer未能从嘈杂的特征中学习时
链接：https://arxiv.org/abs/2602.08695
备注：13+12 pages, 7 figures. Accepted at ICLR 2026
【6】Time-Delayed Transformers for Data-Driven Modeling of Low-Dimensional Dynamics
标题：用于低维动力学数据驱动建模的延时变形器
链接：https://arxiv.org/abs/2602.08478
【7】Low Rank Transformer for Multivariate Time Series Anomaly Detection and Localization
标题：基于低秩Transformer的多元时间序列异常检测与定位
链接：https://arxiv.org/abs/2602.08467
【8】Noise Stability of Transformer Models
标题：Transformer模型的噪音稳定性
链接：https://arxiv.org/abs/2602.08287
备注：Published in ICLR 2026
【9】Learning in Context, Guided by Choice: A Reward-Free Paradigm for Reinforcement Learning with Transformers
标题：受选择引导的背景学习：Transformer强化学习的免奖励范式
链接：https://arxiv.org/abs/2602.08244
【10】Thermodynamic Isomorphism of Transformers: A Lagrangian Approach to Attention Dynamics
标题：Transformer的热力学同质性：注意力动力学的拉格朗日方法
链接：https://arxiv.org/abs/2602.08216
备注：9 pages, 1 figure. Based on a thermodynamic framework for Transformer architectures. Derives the equation of state from first principles
【11】Approximating Matrix Functions with Deep Neural Networks and Transformers
标题：使用深度神经网络和变形器逼近矩阵函数
链接：https://arxiv.org/abs/2602.07800
【12】Gaussian Match-and-Copy: A Minimalist Benchmark for Studying Transformer Induction
标题：高斯匹配复制：研究Transformer感应的极简基准
链接：https://arxiv.org/abs/2602.07562
【13】Brep2Shape: Boundary and Shape Representation Alignment via Self-Supervised Transformers
标题：Brep 2Shape：通过自我监督Transformer进行边界和形状表示对齐
链接：https://arxiv.org/abs/2602.07429
【14】Parallel Track Transformers: Enabling Fast GPU Inference with Reduced Synchronization
标题：并行轨迹Transformer：通过减少同步来实现快速的图形处理器推理
链接：https://arxiv.org/abs/2602.07306
【15】Hybrid Dual-Path Linear Transformations for Efficient Transformer Architectures
标题：实现高效Transformer架构的混合双路径线性变换
链接：https://arxiv.org/abs/2602.07070
GAN|对抗|攻击|生成相关(13篇)
【1】StealthRL: Reinforcement Learning Paraphrase Attacks for Multi-Detector Evasion of AI-Text Detectors
标题：StealthRL：针对AI文本检测器的多检测器规避的强化学习重述攻击
链接：https://arxiv.org/abs/2602.08934
备注：Expanded version of a workshop submission. Code available
【2】Dashed Line Defense: Plug-And-Play Defense Against Adaptive Score-Based Query Attacks
标题：虚线防御：针对自适应基于分数的查询攻击的即插即用防御
链接：https://arxiv.org/abs/2602.08679
【3】Nansde-net: A neural sde framework for generating time series with memory
标题：Nansde-net：用于使用记忆生成时间序列的神经sde框架
链接：https://arxiv.org/abs/2602.08182
备注：PAKDD2026 Accepted
【4】Implicit Strategic Optimization: Rethinking Long-Horizon Decision-Making in Adversarial Poker Environments
标题：隐性战略优化：重新思考对抗性扑克环境中的长期决策
链接：https://arxiv.org/abs/2602.08041
【5】MARTI-MARS$^2$: Scaling Multi-Agent Self-Search via Reinforcement Learning for Code Generation
标题：MARTI-MARS$^2$：通过强化学习扩展多智能体自搜索以生成代码
链接：https://arxiv.org/abs/2602.07848
【6】Surprisal-Guided Selection: Compute-Optimal Test-Time Strategies for Execution-Grounded Code Generation
标题：惊喜引导的选择：基于执行的代码生成的计算最优测试时策略
链接：https://arxiv.org/abs/2602.07670
备注：13 pages, 7 figures, 11 tables. Preprint. Code: https://github.com/jbarnes850/test-time-training
【7】Unified Biomolecular Trajectory Generation via Pretrained Variational Bridge
标题：通过预训练变分桥统一生物分子轨迹生成
链接：https://arxiv.org/abs/2602.07588
备注：The Fourteenth International Conference on Learning Representations (ICLR 2026)
【8】Optimizing Few-Step Generation with Adaptive Matching Distillation
标题：利用自适应匹配蒸馏优化少步生成
链接：https://arxiv.org/abs/2602.07345
备注：25 pages, 15 figures, 11 tables
【9】Finding Connections: Membership Inference Attacks for the Multi-Table Synthetic Data Setting
标题：寻找联系：针对多表合成数据设置的成员推断攻击
链接：https://arxiv.org/abs/2602.07126
【10】Video-based Music Generation
标题：基于视频的音乐生成
链接：https://arxiv.org/abs/2602.07063
备注：PhD thesis, University of Porto
【11】TransConv-DDPM: Enhanced Diffusion Model for Generating Time-Series Data in Healthcare
标题：TransConv-DDPM：用于生成医疗保健中时间序列数据的增强型扩散模型
链接：https://arxiv.org/abs/2602.07033
备注：Previously published at IEEE COMPSAC 2025
【12】On Generation in Metric Spaces
标题：关于度量空间中的生成
链接：https://arxiv.org/abs/2602.07710
【13】Condition Errors Refinement in Autoregressive Image Generation with Diffusion Loss
标题：具有扩散损失的自回归图像生成中的条件误差细化
链接：https://arxiv.org/abs/2602.07022
备注：ICLR 2026
半/弱/无/有监督|不确定性|主动学习(7篇)
【1】LEFT: Learnable Fusion of Tri-view Tokens for Unsupervised Time Series Anomaly Detection
标题：左：用于无监督时间序列异常检测的三视图令牌的可学习融合
链接：https://arxiv.org/abs/2602.08638
【2】Estimating Aleatoric Uncertainty in the Causal Treatment Effect
标题：估计因果治疗效应中的性欲不确定性
链接：https://arxiv.org/abs/2602.08461
【3】Self-Supervised Bootstrapping of Action-Predictive Embodied Reasoning
标题：时间预测推理的自我监督引导
链接：https://arxiv.org/abs/2602.08167
【4】Variance-Gated Ensembles: An Epistemic-Aware Framework for Uncertainty Estimation
标题：方差门控集成：不确定性估计的认识意识框架
链接：https://arxiv.org/abs/2602.08142
【5】Automated rock joint trace mapping using a supervised learning model trained on synthetic data generated by parametric modelling
标题：使用在参数建模生成的合成数据上训练的监督学习模型来自动绘制岩石接缝轨迹
链接：https://arxiv.org/abs/2602.07590
备注：35 pages, 12 figures, 2 appendices
【6】Active Learning Using Aggregated Acquisition Functions: Accuracy and Sustainability Analysis
标题：使用聚合获取功能的主动学习：准确性和可持续性分析
链接：https://arxiv.org/abs/2602.07440
【7】The Value of Variance: Mitigating Debate Collapse in Multi-Agent Systems via Uncertainty-Driven Policy Optimization
标题：方差的价值：通过不确定性驱动的政策优化缓解多主体系统中的辩论崩溃
链接：https://arxiv.org/abs/2602.07186
迁移|Zero/Few/One-Shot|自适应(14篇)
【1】ANCRe: Adaptive Neural Connection Reassignment for Efficient Depth Scaling
标题：ANCRe：自适应神经连接重新分配以实现高效深度缩放
链接：https://arxiv.org/abs/2602.09009
【2】StretchTime: Adaptive Time Series Forecasting via Symplectic Attention
标题：StretchTime：通过辛注意力的自适应时间序列预测
链接：https://arxiv.org/abs/2602.08983
【3】CompilerKV: Risk-Adaptive KV Compression via Offline Experience Compilation
标题：CompilerKN：通过离线体验编译进行风险自适应KV压缩
链接：https://arxiv.org/abs/2602.08686
【4】Beyond Correctness: Learning Robust Reasoning via Transfer
标题：超越正确性：通过转移学习稳健推理
链接：https://arxiv.org/abs/2602.08489
【5】V-ABFT: Variance-Based Adaptive Threshold for Fault-Tolerant Matrix Multiplication in Mixed-Precision Deep Learning
标题：V-ABFT：混合精度深度学习中用于容差矩阵相乘的基于方差的自适应阈值
链接：https://arxiv.org/abs/2602.08043
【6】AceGRPO: Adaptive Curriculum Enhanced Group Relative Policy Optimization for Autonomous Machine Learning Engineering
标题：AceGRPO：自主机器学习工程的自适应课程增强组相对政策优化
链接：https://arxiv.org/abs/2602.07906
备注：17 pages, 5 figures
【7】Efficient Adaptive Data Analysis over Dense Distributions
标题：稠密分布上的高效自适应数据分析
链接：https://arxiv.org/abs/2602.07732
备注：23 pages
【8】Analyzing and Guiding Zero-Shot Posterior Sampling in Diffusion Models
标题：扩散模型中的Zero-Shot后验抽样分析和指导
链接：https://arxiv.org/abs/2602.07715
【9】ODELoRA: Training Low-Rank Adaptation by Solving Ordinary Differential Equations
标题：ODELoRA：通过求解常微分方程训练低秩自适应
链接：https://arxiv.org/abs/2602.07479
备注：38 pages
【10】Laplacian-LoRA: Delaying Oversmoothing in Deep GCNs via Spectral Low-Rank Adaptation
标题：Laplacian-LoRA：通过频谱低阶自适应延迟深度GCN中的过平滑
链接：https://arxiv.org/abs/2602.07278
备注：4 pages
【11】Cerebellar-Inspired Residual Control for Fault Recovery: From Inference-Time Adaptation to Structural Consolidation
标题：应用于断层恢复的脑白质残余控制：从推理时间适应到结构整合
链接：https://arxiv.org/abs/2602.07227
【12】Online Learning for Uninformed Markov Games: Empirical Nash-Value Regret and Non-Stationarity Adaptation
标题：不知情的马尔科夫游戏在线学习：经验性的纳什值遗憾和非平稳性适应
链接：https://arxiv.org/abs/2602.07205
备注：36 pages
【13】Adaptive Matrix Online Learning through Smoothing with Guarantees for Nonsmooth Nonconvex Optimization
标题：通过保证非光滑非凸优化的平滑进行自适应矩阵在线学习
链接：https://arxiv.org/abs/2602.08232
备注：37 pages, 1 figure
【14】Adaptive Temporal Dynamics for Personalized Emotion Recognition: A Liquid Neural Network Approach
标题：个性化情绪识别的自适应时间动力学：一种液态神经网络方法
链接：https://arxiv.org/abs/2602.06997
强化学习(19篇)
【1】Learning to Coordinate via Quantum Entanglement in Multi-Agent Reinforcement Learning
标题：多智能体强化学习中通过量子纠缠学习协调
链接：https://arxiv.org/abs/2602.08965
【2】Learning the Value Systems of Societies with Preference-based Multi-objective Reinforcement Learning
标题：利用基于偏好的多目标强化学习社会的价值体系
链接：https://arxiv.org/abs/2602.08835
备注：18 pages, 3 figures. To be published in proceedings of the 25th International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS 2026). This is a full version that includes the supplementary material
【3】SoK: The Pitfalls of Deep Reinforcement Learning for Cybersecurity
标题：SoK：网络安全深度强化学习的陷阱
链接：https://arxiv.org/abs/2602.08690
【4】Learning To Sample From Diffusion Models Via Inverse Reinforcement Learning
标题：通过反向强化学习学习从扩散模型中采样
链接：https://arxiv.org/abs/2602.08689
备注：Preprint
【5】Breaking the Grid: Distance-Guided Reinforcement Learning in Large Discrete and Hybrid Action Spaces
标题：打破网格：大型离散和混合动作空间中的距离引导强化学习
链接：https://arxiv.org/abs/2602.08616
备注：26 pages, 8 figures
【6】Conditional Sequence Modeling for Safe Reinforcement Learning
标题：安全强化学习的条件序列建模
链接：https://arxiv.org/abs/2602.08584
【7】Contextual Rollout Bandits for Reinforcement Learning with Verifiable Rewards
标题：用于强化学习的上下文推出Bandits，具有可验证的奖励
链接：https://arxiv.org/abs/2602.08499
【8】Reinforcement Learning with Backtracking Feedback
标题：带回溯反馈的强化学习
链接：https://arxiv.org/abs/2602.08377
备注：NeurIPS 2025
【9】SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning
标题：SkillRL：通过回归技能增强强化学习来进化代理
链接：https://arxiv.org/abs/2602.08234
【10】Interpretable Failure Analysis in Multi-Agent Reinforcement Learning Systems
标题：多智能体强化学习系统中的可解释故障分析
链接：https://arxiv.org/abs/2602.08104
【11】Epigraph-Guided Flow Matching for Safe and Performant Offline Reinforcement Learning
标题：墓志铭引导的流匹配，实现安全且高效的离线强化学习
链接：https://arxiv.org/abs/2602.08054
备注：23 pages, 8 figures
【12】Efficient Anti-exploration via VQVAE and Fuzzy Clustering in Offline Reinforcement Learning
标题：离线强化学习中通过VQVAE和模糊集群实现高效反探索
链接：https://arxiv.org/abs/2602.07889
【13】Preference Conditioned Multi-Objective Reinforcement Learning: Decomposed, Diversity-Driven Policy Optimization
标题：偏好条件多目标强化学习：分解的、多元化驱动的政策优化
链接：https://arxiv.org/abs/2602.07764
【14】Efficient Planning in Reinforcement Learning via Model Introspection
标题：基于模型内省的强化学习有效规划
链接：https://arxiv.org/abs/2602.07719
【15】CoMI-IRL: Contrastive Multi-Intention Inverse Reinforcement Learning
标题：CoMI-IRL：对比多意图反向强化学习
链接：https://arxiv.org/abs/2602.07496
备注：14 pages, 6 figures
【16】Proximal Action Replacement for Behavior Cloning Actor-Critic in Offline Reinforcement Learning
标题：行为克隆的近端动作替代离线强化学习中的演员-批评者
链接：https://arxiv.org/abs/2602.07441
【17】High Fidelity Textual User Representation over Heterogeneous Sources via Reinforcement Learning
标题：通过强化学习实现异类源的高保真文本用户表示
链接：https://arxiv.org/abs/2602.07333
【18】Automating the Refinement of Reinforcement Learning Specifications
标题：自动细化强化学习规范
链接：https://arxiv.org/abs/2512.01047
【19】Deep Reinforcement Learning for Interference Suppression in RIS-Aided Space-Air-Ground Integrated Networks
标题：RIS辅助空地综合网络中干扰抑制的深度强化学习
链接：https://arxiv.org/abs/2602.06982
元学习(2篇)
【1】Is Meta-Path Attention an Explanation? Evidence of Alignment and Decoupling in Heterogeneous GNNs
标题：元路径注意力是解释吗？异类GNN中对齐和脱钩的证据
链接：https://arxiv.org/abs/2602.08500
【2】Amortising Inference and Meta-Learning Priors in Neural Networks
标题：神经网络中的推断和元学习先验
链接：https://arxiv.org/abs/2602.08782
备注：Accepted at ICLR 2026
符号|符号学习(2篇)
【1】Breaking the Simplification Bottleneck in Amortized Neural Symbolic Regression
标题：打破分期神经元符号回归简化瓶颈
链接：https://arxiv.org/abs/2602.08885
备注：main text: 8 pages, 7 figures appendix: 12 pages, 11 figures code available at https://github.com/psaegert/simplipy and https://github.com/psaegert/flash-ansr
【2】Interpretable Analytic Calabi-Yau Metrics via Symbolic Distillation
标题：通过符号蒸馏的可解释分析Calabi-Yau
链接：https://arxiv.org/abs/2602.07834
备注：31 pages, 7 figures
分层学习(1篇)
【1】Improving Detection of Rare Nodes in Hierarchical Multi-Label Learning
标题：改进分层多标签学习中稀有节点的检测
链接：https://arxiv.org/abs/2602.08986
备注：Accepted for publication in Transactions on Machine Learning Research (TMLR), 2026
医学相关(9篇)
【1】AMS-HD: Hyperdimensional Computing for Real-Time and Energy-Efficient Acute Mountain Sickness Detection
标题：AMS-HD：用于实时和节能的急性高山病检测的超维计算
链接：https://arxiv.org/abs/2602.08916
【2】A Causal Machine Learning Framework for Treatment Personalization in Clinical Trials: Application to Ulcerative Colitis
标题：临床试验中治疗个性化的因果机器学习框架：应用于溃疡性结肠炎
链接：https://arxiv.org/abs/2602.08171
【3】Multimodal normative modeling in Alzheimers Disease with introspective variational autoencoders
标题：使用内省变分自动编码器进行阿尔茨海默病的多模式规范建模
链接：https://arxiv.org/abs/2602.08077
备注：Conference on Health, Inference, and Learning (CHIL)
【4】Attention-Based Deep Learning for Early Parkinson's Disease Detection with Tabular Biomedical Data
标题：基于注意力的深度学习利用表格生物医学数据进行早期帕金森病检测
链接：https://arxiv.org/abs/2602.07933
【5】Dense Feature Learning via Linear Structure Preservation in Medical Data
标题：通过医疗数据中线性结构保留进行密集特征学习
链接：https://arxiv.org/abs/2602.07706
备注：ICLR Workshop
【6】MedVerse: Efficient and Reliable Medical Reasoning via DAG-Structured Parallel Execution
标题：MedVerse：通过DAB结构并行执行高效可靠的医学推理
链接
：https://arxiv.org/abs/2602.07529
【7】3D Transport-based Morphometry (3D-TBM) for medical image analysis
标题：用于医学图像分析的3D基于传输的形态测量术（3D-TBC）
链接：https://arxiv.org/abs/2602.07260
【8】Attention-Driven Framework for Non-Rigid Medical Image Registration
标题：非刚性医学图像配准的注意力驱动框架
链接：https://arxiv.org/abs/2602.07088
【9】High-fidelity 3D multi-slab diffusion MRI using Slab-shifting for Harmonized 3D Acquisition and Reconstruction with Profile Encoding Networks (SHARPEN)
标题：高保真3D多板扩散MRI使用片移进行协调3D采集和重建，并使用轮廓编码网络（SHARPEN）
链接：https://arxiv.org/abs/2602.07162
备注：26 pages, 11 figures + supplementary info
蒸馏|知识提取(4篇)
【1】RIFLE: Robust Distillation-based FL for Deep Model Deployment on Resource-Constrained IoT Networks
标题：RIFLE：基于蒸馏的稳健FL，用于在资源受限的物联网网络上部署深度模型
链接：https://arxiv.org/abs/2602.08446
备注：This paper has been accepted for publication in IEEE ICC 2026 and will be indexed in the IEEE Xplore Digital Library
【2】PAND: Prompt-Aware Neighborhood Distillation for Lightweight Fine-Grained Visual Classification
标题：PAND：用于轻量级细粒度视觉分类的预算感知邻里蒸馏
链接：https://arxiv.org/abs/2602.07768
备注：6pages, 3 figures, conference
【3】Pareto-guided Pipeline for Distilling Featherweight AI Agents in Mobile MOBA Games
标题：帕累托引导的移动MOBA游戏中提炼羽量级人工智能代理管道
链接：https://arxiv.org/abs/2602.07521
【4】FADE: Selective Forgetting via Sparse LoRA and Self-Distillation
标题：FADE：通过稀疏LoRA和自蒸馏的选择性遗忘
链接：https://arxiv.org/abs/2602.07058
推荐(4篇)
【1】Contrastive Learning for Diversity-Aware Product Recommendations in Retail
标题：零售业多元化产品推荐的对比学习
链接：https://arxiv.org/abs/2602.08886
【2】Learning to Alleviate Familiarity Bias in Video Recommendation
标题：学会减轻视频推荐中的熟悉度偏见
链接：https://arxiv.org/abs/2602.07987
备注：Accepted to the Companion Proceedings of the ACM Web Conference 2026 (WWW '26), April 13-17, 2026, Dubai, UAE
【3】MDL: A Unified Multi-Distribution Learner in Large-scale Industrial Recommendation through Tokenization
标题：MDL：通过代币化进行大规模工业推荐的统一多分布学习者
链接：https://arxiv.org/abs/2602.07520
备注：9 pages, 4 figures
【4】DSL: Understanding and Improving Softmax Recommender Systems with Competition-Aware Scaling
标题：SL：通过竞争感知扩展来了解和改进Softmax推荐系统
链接：https://arxiv.org/abs/2602.07206
聚类(2篇)
【1】VertCoHiRF: Decentralized Vertical Clustering Beyond k-means
标题：VertCoHiRF：超越k-means的去中心化垂直集群
链接：https://arxiv.org/abs/2602.07279
【2】Adjustment of Cluster-Then-Predict Framework for Multiport Scatterer Load Prediction
标题：多端口散布器负载预测的先确定后预测框架的调整
链接：https://arxiv.org/abs/2602.08129
自动驾驶|车辆|车道检测等(8篇)
【1】Robustness Is a Function, Not a Number: A Factorized Comprehensive Study of OOD Robustness in Vision-Based Driving
标题：稳健性是一个函数，而不是一个数字：基于视觉的驾驶中OOD稳健性的因子化综合研究
链接：https://arxiv.org/abs/2602.09018
【2】PACC: Protocol-Aware Cross-Layer Compression for Compact Network Traffic Representation
标题：PACC：用于紧凑网络流量表示的协议感知跨层压缩
链接：https://arxiv.org/abs/2602.08331
【3】Vision and language: Novel Representations and Artificial intelligence for Driving Scene Safety Assessment and Autonomous Vehicle Planning
标题：愿景和语言：驾驶场景安全评估和自动驾驶车辆规划的新型表示和人工智能
链接：https://arxiv.org/abs/2602.07680
【4】Looking and Listening Inside and Outside: Multimodal Artificial Intelligence Systems for Driver Safety Assessment and Intelligent Vehicle Decision-Making
标题：内外观察和聆听：用于驾驶员安全评估和智能车辆决策的多模式人工智能系统
链接：https://arxiv.org/abs/2602.07668
【5】Seeing Roads Through Words: A Language-Guided Framework for RGB-T Driving Scene Segmentation
标题：通过文字看路：用于RGB-T驱动场景分割的图形引导框架
链接：https://arxiv.org/abs/2602.07343
【6】RAPiD: Real-time Deterministic Trajectory Planning via Diffusion Behavior Priors for Safe and Efficient Autonomous Driving
标题：RAPiD：通过扩散行为先验进行实时确定性轨迹规划，实现安全有效的自动驾驶
链接：https://arxiv.org/abs/2602.07339
【7】Beyond Crash: Hijacking Your Autonomous Vehicle for Fun and Profit
标题：超越崩溃：劫持您的自动驾驶汽车以获取乐趣和利润
链接：https://arxiv.org/abs/2602.07249
【8】Extracting Root-Causal Brain Activity Driving Psychopathology from Resting State fMRI
标题：从静息状态fMRI中提取驱动精神病理的根本原因大脑活动
链接：https://arxiv.org/abs/2602.07233
点云|SLAM|雷达|激光|深度RGBD相关(2篇)
【1】On the Infinite Width and Depth Limits of Predictive Coding Networks
标题：关于预测编码网络的无限宽度和深度限制
链接：https://arxiv.org/abs/2602.07697
备注：31 pages, 27 figures
【2】The Median is Easier than it Looks: Approximation with a Constant-Depth, Linear-Width ReLU Network
标题：中位数比看起来更容易：用恒定深度、线性宽度ReLU网络进行逼近
链接：https://arxiv.org/abs/2602.07219
联邦学习|隐私保护|加密(4篇)
【1】ERIS: Enhancing Privacy and Communication Efficiency in Serverless Federated Learning
标题：ERIS：增强无服务器联邦学习中的隐私和通信效率
链接：https://arxiv.org/abs/2602.08617
【2】SDFed: Bridging Local Global Discrepancy via Subspace Refinement and Divergence Control in Federated Prompt Learning
标题：SDFed：通过联邦即时学习中的子空间细化和分歧控制弥合局部全局差异
链接：https://arxiv.org/abs/2602.08590
备注：13 pages, 6 figures
【3】Trust-Based Incentive Mechanisms in Semi-Decentralized Federated Learning Systems
标题：半分散联邦学习系统中基于信任的激励机制
链接：https://arxiv.org/abs/2602.08290
备注：To appear in the ICBTA 2025 Conference Proceedings and published as a volume of Lecture Notes in Networks and Systems by Springer
【4】Federated Learning with Profile Mapping under Distribution Shifts and Drifts
标题：分布偏移和漂移下基于轮廓映射的联邦学习
链接：https://arxiv.org/abs/2602.07671
备注：ICLR2026
推理|分析|理解|解释(30篇)
【1】Analysis of Converged 3D Gaussian Splatting Solutions: Density Effects and Prediction Limit
标题：收敛3D高斯飞溅解的分析：密度效应和预测极限
链接：https://arxiv.org/abs/2602.08909
【2】Positive Distribution Shift as a Framework for Understanding Tractable Learning
标题：正分布转移作为理解可持续学习的框架
链接：https://arxiv.org/abs/2602.08907
【3】Empirically Understanding the Value of Prediction in Allocation
标题：凭经验认识预测在分配中的价值
链接：https://arxiv.org/abs/2602.08786
【4】Foundation Inference Models for Ordinary Differential Equations
标题：常微方程的基础推理模型
链接：https://arxiv.org/abs/2602.08733
【5】Towards Understanding Multimodal Fine-Tuning: Spatial Features
标题：了解多模式微调：空间特征
链接：https://arxiv.org/abs/2602.08713
【6】The Theory and Practice of MAP Inference over Non-Convex Constraints
标题：非凸约束下MAP推理的理论与实践
链接：https://arxiv.org/abs/2602.08681
【7】Near-Oracle KV Selection via Pre-hoc Sparsity for Long-Context Inference
标题：通过预组织稀疏性进行近Oracle KN选择以进行长上下文推理
链接：https://arxiv.org/abs/2602.08329
备注：An effective method for accelerating LLM's inference via selective KV processing
【8】Sharp analysis of linear ensemble sampling
标题：线性集合抽样的夏普分析
链接：https://arxiv.org/abs/2602.08026
【9】Regret Analysis of Unichain Average Reward Constrained MDPs with General Parameterization
标题：通用参数化的单链平均回报约束MDPs的遗憾分析
链接：https://arxiv.org/abs/2602.08000
【10】When Is Compositional Reasoning Learnable from Verifiable Rewards?
标题：什么时候可以从可验证的奖励中学习成分推理？
链接：https://arxiv.org/abs/2602.07992
【11】An Explainable Multi-Task Similarity Measure: Integrating Accumulated Local Effects and Weighted Fréchet Distance
标题：一种可解释的多任务相似性度量：集成累积局部效应和加权Fréchet距离
链接：https://arxiv.org/abs/2602.07966
【12】GRAFT: Decoupling Ranking and Calibration for Survival Analysis
标题：GRAFT：生存分析的脱钩排名和校准
链接：https://arxiv.org/abs/2602.07884
【13】Learnable Chernoff Baselines for Inference-Time Alignment
标题：可学习的推理时间对齐基线
链接：https://arxiv.org/abs/2602.07738
【14】Data-Aware and Scalable Sensitivity Analysis for Decision Tree Ensembles
标题：决策树集成的数据感知和可扩展敏感性分析
链接：https://arxiv.org/abs/2602.07453
【15】Dichotomy of Feature Learning and Unlearning: Fast-Slow Analysis on Neural Networks with Stochastic Gradient Descent
标题：特征学习和取消学习的二分法：随机梯度下降神经网络的快-慢分析
链接：https://arxiv.org/abs/2602.07378
备注：40 pages
【16】XShare: Collaborative in-Batch Expert Sharing for Faster MoE Inference
标题：XShare：协作批量专家共享，以更快的MoE推理
链接：https://arxiv.org/abs/2602.07265
【17】Latent Target Score Matching, with an application to Simulation-Based Inference
标题：潜在目标分数匹配，应用于基于模拟的推理
链接：https://arxiv.org/abs/2602.07189
备注：Machine Learning and the Physical Sciences Workshop, NeurIPS 2025
【18】Landscaper: Understanding Loss Landscapes Through Multi-Dimensional Topological Analysis
标题：景观设计师：通过多维布局分析了解损失景观
链接：https://arxiv.org/abs/2602.07135
【19】Reasoning-Augmented Representations for Multimodal Retrieval
标题：多模式检索的推理增强表示
链接：https://arxiv.org/abs/2602.07125
【20】AVERE: Improving Audiovisual Emotion Reasoning with Preference Optimization
标题：AVRE：通过偏好优化改进视听情感推理
链接：https://arxiv.org/abs/2602.07054
备注：Accepted as a conference paper at ICLR 2026. Project page: https://avere-iclr.github.io
【21】OMNI-Dent: Towards an Accessible and Explainable AI Framework for Automated Dental Diagnosis
标题：OMNI-Dent：迈向自动牙科诊断的可访问且可解释的人工智能框架
链接：https://arxiv.org/abs/2602.07041
【22】Lagged backward-compatible physics-informed neural networks for unsaturated soil consolidation analysis
标题：用于非饱和土压实分析的滞后向后相容物理信息神经网络
链接：https://arxiv.org/abs/2602.07031
【23】Scalable spatial point process models for forensic footwear analysis
标题：用于法医鞋类分析的可扩展空间点过程模型
链接：https://arxiv.org/abs/2602.07006
【24】Bridging the Knowledge Void: Inference-time Acquisition of Unfamiliar Programming Languages for Coding Tasks
标题：弥合知识真空：推理时获取用于编码任务的陌生编程语言
链接：https://arxiv.org/abs/2602.06976
【25】GAAVI: Global Asymptotic Anytime Valid Inference for the Conditional Mean Function
标题：GAavi：条件均值函数的全局渐进随时有效推断
链接：https://arxiv.org/abs/2602.08096
【26】Scalable Mean-Field Variational Inference via Preconditioned Primal-Dual Optimization
标题：通过预条件原始-二元优化的可扩展平均场变分推理
链接：https://arxiv.org/abs/2602.07632
【27】Statistical inference after variable selection in Cox models: A simulation study
标题：Cox模型中变量选择后的统计推断：模拟研究
链接：https://arxiv.org/abs/2602.07477
【28】Fast and Robust Likelihood-Guided Diffusion Posterior Sampling with Amortized Variational Inference
标题：具有摊销变分推理的快速稳健似然引导扩散后验抽样
链接：https://arxiv.org/abs/2602.07102
【29】BayesFlow 2.0: Multi-Backend Amortized Bayesian Inference in Python
标题：BayesFlow 2.0：Python中的多后台摊销Bayesian推理
链接：https://arxiv.org/abs/2602.07098
【30】LatentChem: From Textual CoT to Latent Thinking in Chemical Reasoning
标题：LatentChem：从文本CoT到化学推理中的潜在思维
链接：https://arxiv.org/abs/2602.07075
检测相关(8篇)
【1】Multimodal Learning for Arcing Detection in Pantograph-Catenary Systems
标题：多模式学习用于受电弓-接触网系统中的弧线检测
链接：https://arxiv.org/abs/2602.08792
【2】ManifoldKV: Training-Free KV Cache Compression via Euclidean Outlier Detection
标题：ManifoldKV：通过欧几里德离群点检测的免训练KV缓存压缩
链接：https://arxiv.org/abs/2602.08343
备注：18 pages, 5 figures, 18 tables
【3】Evasion of IoT Malware Detection via Dummy Code Injection
标题：通过伪代码注入逃避物联网恶意软件检测
链接：https://arxiv.org/abs/2602.08170
【4】MMLSv2: A Multimodal Dataset for Martian Landslide Detection in Remote Sensing Imagery
标题：MMLSv 2：用于遥感图像中火星滑坡检测的多峰数据集
链接：https://arxiv.org/abs/2602.08112
【5】Spectral Guardrails for Agents in the Wild: Detecting Tool Use Hallucinations via Attention Topology
标题：野生智能体的光谱护栏：通过注意力拓扑检测工具使用幻觉
链接：https://arxiv.org/abs/2602.08082
备注：32 pages, 2 fgures, 18 tables
【6】TASTE: Task-Aware Out-of-Distribution Detection via Stein Operators
标题：TASTE：通过Stein操作员的任务感知分发外检测
链接：https://arxiv.org/abs/2602.07640
【7】Cutting Through the Noise: On-the-fly Outlier Detection for Robust Training of Machine Learning Interatomic Potentials
标题：穿透噪声：用于机器学习原子间势鲁棒训练的动态离群值检测
链接：https://arxiv.org/abs/2602.08849
备注：12 pages, 6 figures
【8】Fundamental Limits of Community Detection in Contextual Multi-Layer Stochastic Block Models
标题：上下文多层随机块模型中社区检测的基本局限性
链接：https://arxiv.org/abs/2602.08173
备注：49 pages, 5 figures
分类|识别(10篇)
【1】ShapeCond: Fast Shapelet-Guided Dataset Condensation for Time Series Classification
标题：ShapeCond：用于时间序列分类的快速Shapelet引导数据集浓缩
链接：https://arxiv.org/abs/2602.09008
备注：Code at: https://github.com/lunaaa95/ShapeCond
【2】GEMSS: A Variational Bayesian Method for Discovering Multiple Sparse Solutions in Classification and Regression Problems
标题：GEMPS：一种用于发现分类和回归问题中多个稀疏解的变分Bayesian方法
链接：https://arxiv.org/abs/2602.08913
【3】Redundancy-Free View Alignment for Multimodal Human Activity Recognition with Arbitrarily Missing Views
标题：具有潜在缺失视图的多模式人类活动识别的无冗余视图对齐
链接：https://arxiv.org/abs/2602.08755
【4】Enhanced Food Category Recognition under Illumination-Induced Domain Shift
标题：光照诱导域转移下增强食品类别识别
链接：https://arxiv.org/abs/2602.08491
【5】Gesture Matters: Pedestrian Gesture Recognition for AVs Through Skeleton Pose Evaluation
标题：手势很重要：通过骨架姿势评估识别AV的行人手势
链接：https://arxiv.org/abs/2602.08479
备注：9th International Conference on Instrumentation, Control, and Automation (ICA)
【6】Enhancing Time Series Classification with Diversity-Driven Neural Network Ensembles
标题：利用多样性驱动的神经网络集成增强时间序列分类
链接：https://arxiv.org/abs/2602.07579
备注：Published in IEEE IJCNN 2025 proceedings. 10 pages, 8 figures
【7】Fair Decisions from Calibrated Scores: Achieving Optimal Classification While Satisfying Sufficiency
标题：来自校准分数的公平决策：在满足充分性的同时实现最佳分类
链接：https://arxiv.org/abs/2602.07285
【8】Speech Emotion Recognition Leveraging OpenAI's Whisper Representations and Attentive Pooling Methods
标题：利用OpenAI的Whisper表示和细心的池化方法的语音情感识别
链接：https://arxiv.org/abs/2602.06000
【9】Empirical Study of Observable Sets in Multiclass Quantum Classification
标题：多类量子分类中可观测集的实证研究
链接：https://arxiv.org/abs/2602.08485
备注：13 pages, 11 figures
【10】Hybrid Deep Learning Framework for CSI-Based Activity Recognition in Bandwidth-Constrained Wi-Fi Sensing
标题：用于带宽限制Wi-Fi感知中基于CSC的活动识别的混合深度学习框架
链接：https://arxiv.org/abs/2602.06983
备注：6 pages, 6 figures
表征(8篇)
【1】Circuit Representations of Random Forests with Applications to XAI
标题：随机森林的电路表示及其在XAI中的应用
链接：https://arxiv.org/abs/2602.08362
【2】On Improving Neurosymbolic Learning by Exploiting the Representation Space
标题：利用表象空间改善神经符号学习
链接：https://arxiv.org/abs/2602.07973
【3】Efficient Representations are Controllable Representations
标题：有效的表示是可控的表示
链接：https://arxiv.org/abs/2602.07828
【4】TerraBind: Fast and Accurate Binding Affinity Prediction through Coarse Structural Representations
标题：TerraBind：通过粗结构表示快速准确地预测结合亲和力
链接：https://arxiv.org/abs/2602.07735
备注：31 pages, 14 figures
【5】Escaping Spectral Bias without Backpropagation: Fast Implicit Neural Representations with Extreme Learning Machines
标题：无需反向传播即可摆脱谱偏差：使用极限学习机器的快速隐式神经表示
链接：https://arxiv.org/abs/2602.07603
【6】Probing Neural TSP Representations for Prescriptive Decision Support
标题：探索TPS神经表示以提供规定性决策支持
链接：https://arxiv.org/abs/2602.07216
备注：Submitted to ICML 2026
【7】Electron-Informed Coarse-Graining Molecular Representation Learning for Real-World Molecular Physics
标题：现实世界分子物理的电子信息粗粒度分子表示学习
链接：https://arxiv.org/abs/2602.07087
备注：KDD 2025 Research Track
【8】Financial Bond Similarity Search Using Representation Learning
标题：利用表示学习进行金融债券相似性搜索
链接：https://arxiv.org/abs/2602.07020
备注：22 pages, 18 figures, 1 table
优化|敛散性(25篇)
【1】ARO: A New Lens On Matrix Optimization For Large Models
标题：ARO：大型型号矩阵优化的新镜头
链接：https://arxiv.org/abs/2602.09006
【2】Distributionally Robust Optimization via Generative Ambiguity Modeling
标题：基于生成模糊建模的分布鲁棒优化
链接：https://arxiv.org/abs/2602.08976
【3】Near-optimal Swap Regret Minimization for Convex Losses
标题：凸损失的近最优交换后悔最小化
链接：https://arxiv.org/abs/2602.08862
【4】Robust Policy Optimization to Prevent Catastrophic Forgetting
标题：稳健的政策优化以防止灾难性遗忘
链接：https://arxiv.org/abs/2602.08813
【5】Default Machine Learning Hyperparameters Do Not Provide Informative Initialization for Bayesian Optimization
标题：默认机器学习超参数不为Bayesian优化提供信息性数据集
链接：https://arxiv.org/abs/2602.08774
【6】Data Reconstruction: Identifiability and Optimization with Sample Splitting
标题：数据重建：通过样本拆分的可识别性和优化
链接：https://arxiv.org/abs/2602.08723
【7】Predicting Future Utility: Global Combinatorial Optimization for Task-Agnostic KV Cache Eviction
标题：预测未来效用：任务不可知的KV缓存驱逐的全局组合优化
链接：https://arxiv.org/abs/2602.08585
【8】Causal Schrödinger Bridges: Constrained Optimal Transport on Structural Manifolds
标题：因果Schrödinger桥：结构上的约束最优输运
链接：https://arxiv.org/abs/2602.08535
备注：12 pages, 7 figures
【9】Learning Credal Ensembles via Distributionally Robust Optimization
标题：通过分布鲁棒优化学习Credal合奏
链接：https://arxiv.org/abs/2602.08470
备注：32 pages
【10】All ERMs Can Fail in Stochastic Convex Optimization Lower Bounds in Linear Dimension
标题：所有ERM都可能在线性维度的随机凸优化下限中失败
链接：https://arxiv.org/abs/2602.08350
【11】TextResNet: Decoupling and Routing Optimization Signals in Compound AI Systems via Deep Residual Tuning
标题：TextResNet：通过深度残差调整在复合AI系统中解耦和路由优化信号
链接：https://arxiv.org/abs/2602.08306
【12】Constraint-Aware Generative Auto-bidding via Pareto-Prioritized Regret Optimization
标题：基于帕累托优先后悔优化的约束感知生成自动竞价
链接：https://arxiv.org/abs/2602.08261
【13】CADO: From Imitation to Cost Minimization for Heatmap-based Solvers in Combinatorial Optimization
标题：CADO：组合优化中基于热图的求解器从模仿到成本最小化
链接：https://arxiv.org/abs/2602.08210
【14】Beyond Optimization: Intelligence as Metric-Topology Factorization under Geometric Incompleteness
标题：超越优化：智能作为几何不完全性下的度量-拓因分解
链接：https://arxiv.org/abs/2602.07974
【15】MemFly: On-the-Fly Memory Optimization via Information Bottleneck
标题：MemFly：通过信息瓶颈进行实时内存优化
链接：https://arxiv.org/abs/2602.07885
【16】Fairness Aware Reward Optimization
标题：公平意识的奖励优化
链接：https://arxiv.org/abs/2602.07799
【17】Achieving Optimal Static and Dynamic Regret Simultaneously in Bandits with Deterministic Losses
标题：在具有确定性损失的盗贼中同时实现最佳静态和动态遗憾
链接：https://arxiv.org/abs/2602.07418
【18】Nonparametric Bayesian Optimization for General Rewards
标题：一般奖励的非参数Bayesian优化
链接：https://arxiv.org/abs/2602.07411
【19】Optimization of Precipitate Segmentation Through Linear Genetic Programming of Image Processing
标题：图像处理线性遗传规划优化图像分割
链接：https://arxiv.org/abs/2602.07310
备注：39 pages, 12 figures, 1 table
【20】Hybrid Feedback-Guided Optimal Learning for Wireless Interactive Panoramic Scene Delivery
标题：无线交互式全景场景交付的混合反馈引导最佳学习
链接：https://arxiv.org/abs/2602.07273
备注：Submitting to ToN
【21】BONSAI: Bayesian Optimization with Natural Simplicity and Interpretability
标题：BONSAI：具有自然简单性和可解释性的Bayesian优化
链接：https://arxiv.org/abs/2602.07144
备注：26 pages
【22】MolLIBRA: Genetic Molecular Optimization with Multi-Fingerprint Surrogates and Text-Molecule Aligned Critic
标题：MolLIBRA：具有多指纹替代者和文本分子对齐批评者的遗传分子优化
链接：https://arxiv.org/abs/2602.07002
【23】Online monotone density estimation and log-optimal calibration
标题：在线单调密度估计和log最优校准
链接：https://arxiv.org/abs/2602.08927
备注：28 pages, 1 figure
【24】Differentiable Logical Programming for Quantum Circuit Discovery and Optimization
标题：量子电路发现和优化的可区分逻辑编程
链接：https://arxiv.org/abs/2602.08880
【25】Fast Model Selection and Stable Optimization for Softmax-Gated Multinomial-Logistic Mixture of Experts Models
标题：软最大门控多元逻辑混合专家模型的快速模型选择和稳定优化
链接：https://arxiv.org/abs/2602.07997
备注：TrungKhang Tran and TrungTin Nguyen are co-first authors
预测|估计(8篇)
【1】Discrete Bridges for Mutual Information Estimation
标题：用于互信息估计的离散桥
链接：https://arxiv.org/abs/2602.08894
【2】FreqLens: Interpretable Frequency Attribution for Time Series Forecasting
标题：FreqLens：用于时间序列预测的可解释频率属性
链接：https://arxiv.org/abs/2602.08768
【3】Two-Stage Data Synthesization: A Statistics-Driven Restricted Trade-off between Privacy and Prediction
标题：两阶段数据合成：统计驱动的隐私和预测之间的限制权衡
链接：https://arxiv.org/abs/2602.08657
【4】ForecastOcc: Vision-based Semantic Occupancy Forecasting
标题：ForecastOcc：基于视觉的语义占用预测
链接：https://arxiv.org/abs/2602.08006
【5】AI-Driven Predictive Modelling for Groundwater Salinization in Israel
标题：人工智能驱动的以色列地下水盐碱化预测建模
链接：https://arxiv.org/abs/2602.07478
备注：60 pages, 9 figures in main text and 6 figures in appendix, 2 tables, 3 Appendix
【6】BRIDGE: Predicting Human Task Completion Time From Model Performance
标题：BRIDGE：根据模型性能预测人工任务完成时间
链接：https://arxiv.org/abs/2602.07267
【7】Estimation of Fish Catch Using Sentinel-2, 3 and XGBoost-Kernel-Based Kernel Ridge Regression
标题：使用Sentinel-2、3和XGBoost基于核的核岭回归估计鱼类捕捞量
链接：https://arxiv.org/abs/2602.08511
备注：Manuscript
【8】Flow-Based Conformal Predictive Distributions
标题：基于流的保形预测分布
链接：https://arxiv.org/abs/2602.07633
备注：9 pages, 6 figures, 10 appendix pages
其他神经网络|深度学习|模型|建模(62篇)
【1】Contact-Anchored Policies: Contact Conditioning Creates Strong Robot Utility Models
标题：接触锚定政策：接触条件反射创造强大的机器人实用模型
链接：https://arxiv.org/abs/2602.09017
【2】Learning Potentials for Dynamic Matching and Application to Heart Transplantation
标题：动态匹配的学习潜力及其在心脏移植中的应用
链接：https://arxiv.org/abs/2602.08878
【3】Bayesian Preference Learning for Test-Time Steerable Reward Models
标题：测试时可控奖励模型的Bayesian偏好学习
链接：https://arxiv.org/abs/2602.08819
备注：Preprint
【4】Efficient Deep Learning for Biometrics: Overview, Challenges and Trends in Ear of Frugal AI
标题：生物识别的有效深度学习：概述，挑战和趋势
链接：https://arxiv.org/abs/2602.08809
备注：8 pages, 2 figures, accepted at the 2025 IEEE SDS conference
【5】Equalized Generative Treatment: Matching f-divergences for Fairness in Generative Models
标题：均衡生成处理：生成模型中匹配f-偏差以实现公平
链接：https://arxiv.org/abs/2602.08660
【6】Projected Gradient Ascent for Efficient Reward-Guided Updates with One-Step Generative Models
标题：通过一步生成模型实现高效的奖励引导更新的投影梯度上升
链接：https://arxiv.org/abs/2602.08646
【7】Modeling Score Approximation Errors in Diffusion Models via Forward SPDEs
标题：通过正向SPDES建模扩散模型中的得分逼近误差
链接：https://arxiv.org/abs/2602.08579
【8】M-Loss: Quantifying Model Merging Compatibility with Limited Unlabeled Data
标题：M-Loss：量化模型与有限的未标记数据融合兼容性
链接：https://arxiv.org/abs/2602.08564
备注：Code available at https://github.com/languangduan/mLoss
【9】GOT-Edit: Geometry-Aware Generic Object Tracking via Online Model Editing
标题：GOT-Edit：通过在线模型编辑实现几何感知通用对象跟踪
链接：https://arxiv.org/abs/2602.08550
备注：ICLR 2026. This is a preprint version. The camera-ready version will be updated soon
【10】Do physics-informed neural networks (PINNs) need to be deep? Shallow PINNs using the Levenberg-Marquardt algorithm
标题：基于物理的神经网络（PINN）需要深入吗？使用Levenberg-Marquardt算法的浅PINN
链接：https://arxiv.org/abs/2602.08515
【11】The Connection between Kriging and Large Neural Networks
标题：克里格法和大型神经网络之间的联系
链接：https://arxiv.org/abs/2602.08427
【12】Radial Müntz-Szász Networks: Neural Architectures with Learnable Power Bases for Multidimensional Singularities
标题：辐射Müntz-Szász网络：具有多维奇异性可学习功率基的神经架构
链接：https://arxiv.org/abs/2602.08419
备注：47 pages, 13 figures
【13】Learning Human-Like Badminton Skills for Humanoid Robots
标题：类人机器人学习类人羽毛球技能
链接：https://arxiv.org/abs/2602.08370
备注：10 pages, 4 figures
【14】Regime Change Hypothesis: Foundations for Decoupled Dynamics in Neural Network Training
标题：政权变革假说：神经网络训练中脱钩动力学的基础
链接：https://arxiv.org/abs/2602.08333
备注：8 pages, 1 figure
【15】Interaction-Grounded Learning for Contextual Markov Decision Processes with Personalized Feedback
标题：具有个性化反馈的上下文马尔可夫决策过程的基于交互的学习
链接：https://arxiv.org/abs/2602.08307
【16】Grokking in Linear Models for Logistic Regression
标题：逻辑回归线性模型中的探索
链接：https://arxiv.org/abs/2602.08302
【17】When Do Multi-Agent Systems Outperform? Analysing the Learning Efficiency of Agentic Systems
标题：多代理系统何时表现出色？统计系统的学习效率分析
链接：https://arxiv.org/abs/2602.08272
【18】Sparsity-Aware Evolution for Model Merging
标题：模型合并的稀疏意识进化
链接：https://arxiv.org/abs/2602.08218
【19】Interpretable Dynamic Network Modeling of Tensor Time Series via Kronecker Time-Varying Graphical Lasso
标题：基于Kronecker时变图形套索的张量时间序列可解释动态网络建模
链接：https://arxiv.org/abs/2602.08197
备注：Accepted at ACM Web Conference 2026 (WWW2026)
【20】Dreaming in Code for Curriculum Learning in Open-Ended Worlds
标题：开放世界中的课程学习代码中的梦想
链接：https://arxiv.org/abs/2602.08194
备注：11 pages (main text), 90 pages total. Project page: https://konstantinosmitsides.github.io/dreaming-in-code
【21】Reliable and Responsible Foundation Models: A Comprehensive Survey
标题：可靠和负责任的基金会模式：全面调查
链接：https://arxiv.org/abs/2602.08145
备注：TMLR camera-ready version
【22】Online Bayesian Imbalanced Learning with Bregman-Calibrated Deep Networks
标题：使用Bregman校准的深度网络进行在线Bayesian不平衡学习
链接：https://arxiv.org/abs/2602.08128
【23】Efficient Distribution Learning with Error Bounds in Wasserstein Distance
标题：具有Wasserstein距离误差界的高效分布学习
链接：https://arxiv.org/abs/2602.08063
【24】Horizon Imagination: Efficient On-Policy Training in Diffusion World Models
标题：地平线想象力：扩散世界模型中的有效政策训练
链接：https://arxiv.org/abs/2602.08032
备注：This paper will be published in the ICLR 2026 proceedings
【25】Learning-guided Kansa collocation for forward and inverse PDEs beyond linearity
标题：学习引导的Kansa搭配，用于超越线性的正向和反向偏置
链接：https://arxiv.org/abs/2602.07970
备注：Fangcheng Zhong and Chenliang Zhou are co-corresponding authors
【26】A Thermodynamic Theory of Learning Part II: Critical Period Closure and Continual Learning Failure
标题：学习的热力学理论第二部分：关键期关闭和持续学习失败
链接：https://arxiv.org/abs/2602.07950
备注：Part II of a series entitled "A Thermodynamic Theory of Learning."
【27】Safety Alignment as Continual Learning: Mitigating the Alignment Tax via Orthogonal Gradient Projection
标题：安全调整作为持续学习：通过垂直梯度投影减轻调整税
链接：https://arxiv.org/abs/2602.07892
【28】Dynamic Load Model for Data Centers with Pattern-Consistent Calibration
标题：基于模式一致性校正的数据中心动态负载模型
链接：https://arxiv.org/abs/2602.07859
备注：10 pages, 13 figures
【29】TodoEvolve: Learning to Architect Agent Planning Systems
标题：TodoEvolve：学习构建代理规划系统
链接：https://arxiv.org/abs/2602.07839
【30】Spectral Gating Networks
标题：光谱门控网络
链接：https://arxiv.org/abs/2602.07679
【31】Debugging code world models
标题：收件箱代码世界模型
链接：https://arxiv.org/abs/2602.07672
备注：8 pages, 4 figures, under review in conference
【32】SleepMaMi: A Universal Sleep Foundation Model for Integrating Macro- and Micro-structures
标题：SleepMaMi：集成宏观和微观结构的通用睡眠基金会模型
链接：https://arxiv.org/abs/2602.07628
备注：8 pages, Appendix 9 pages
【33】Dense Neural Networks are not Universal Approximators
标题：密集神经网络不是通用逼近器
链接：https://arxiv.org/abs/2602.07618
【34】SERE: Similarity-based Expert Re-routing for Efficient Batch Decoding in MoE Models
标题：SERE：基于相似性的专家重新路由，用于MoE模型中的高效批量解码
链接：https://arxiv.org/abs/2602.07616
备注：Published as a conference paper at ICLR 2026
【35】Object-Oriented Transition Modeling with Inductive Logic Programming
标题：使用归纳逻辑编程的面向对象转换建模
链接：https://arxiv.org/abs/2602.07602
备注：46 pages, 26 figures
【36】Evaluating Object-Centric Models beyond Object Discovery
标题：评估对象发现之外的以对象为中心的模型
链接：https://arxiv.org/abs/2602.07532
备注：Project Page: https://guided-sa.github.io/eval-ocl/
【37】PALMS: Pavlovian Associative Learning Models Simulator
标题：PALMS：巴甫洛夫联想学习模型模拟器
链接：https://arxiv.org/abs/2602.07519
【38】Physical Analog Kolmogorov-Arnold Networks based on Reconfigurable Nonlinear-Processing Units
标题：基于可重配置非线性处理单元的物理模拟Kolmogorov-Arnold网络
链接：https://arxiv.org/abs/2602.07518
【39】Hyperparameter Transfer Laws for Non-Recurrent Multi-Path Neural Networks
标题：非回归多路径神经网络的超参数传输律
链接：https://arxiv.org/abs/2602.07494
【40】Learning Molecular Chirality via Chiral Determinant Kernels
标题：通过手征决定核学习分子手征
链接：https://arxiv.org/abs/2602.07415
备注：Accepted at the ICLR 2026
【41】BitLogic: Training Framework for Gradient-Based FPGA-Native Neural Networks
标题：BitLogic：基于对象的FPGA-Native神经网络的训练框架
链接：https://arxiv.org/abs/2602.07400
【42】Privately Learning Decision Lists and a Differentially Private Winnow
标题：私人学习决策列表和不同的私人Winnow
链接：https://arxiv.org/abs/2602.07370
备注：27 pages, The 37th International Conference on Algorithmic Learning Theory
【43】FEM-Informed Hypergraph Neural Networks for Efficient Elastoplasticity
标题：基于FEM的超图神经网络实现高效弹塑性
链接：https://arxiv.org/abs/2602.07364
备注 ：43 pages, 26 figures, 8tables
【44】Scalable Dexterous Robot Learning with AR-based Remote Human-Robot Interactions
标题：基于AR的远程人机交互的可扩展灵巧机器人学习
链接：https://arxiv.org/abs/2602.07341
【45】Incorruptible Neural Networks: Training Models that can Generalize to Large Internal Perturbations
标题：坚不可摧的神经网络：可以推广到大内部扰动的训练模型
链接：https://arxiv.org/abs/2602.07320
【46】Cross-View World Models
标题：交叉视角世界模型
链接：https://arxiv.org/abs/2602.07277
备注：12 pages, 7 figures
【47】tLoRA: Efficient Multi-LoRA Training with Elastic Shared Super-Models
标题：tLoRA：使用弹性共享超模型的高效多LoRA训练
链接：https://arxiv.org/abs/2602.07263
【48】Fault-Tolerant Evaluation for Sample-Efficient Model Performance Estimators
标题：样本高效模型性能估计器的容差评估
链接：https://arxiv.org/abs/2602.07226
【49】Automated Modernization of Machine Learning Engineering Notebooks for Reproducibility
标题：机器学习工程笔记本的自动化现代化以实现再现
链接：https://arxiv.org/abs/2602.07195
【50】Systematic Performance Assessment of Deep Material Networks for Multiscale Material Modeling
标题：用于多尺度材料建模的深度材料网络的系统性能评估
链接：https://arxiv.org/abs/2602.07192
【51】Learning Nonlinear Systems In-Context: From Synthetic Data to Real-World Motor Control
标题：在上下文中学习非线性系统：从合成数据到现实世界的电机控制
链接：https://arxiv.org/abs/2602.07173
备注：Accepted to be presented in IEEE ICASSP 2026
【52】Convex Dominance in Deep Learning I: A Scaling Law of Loss and Learning Rate
标题：深度学习中的凸优势I：损失和学习率的缩放定律
链接：https://arxiv.org/abs/2602.07145
备注：Part of a planned series to understand and leverage the convexity in deep learning. Accepted to ICLR 2026
【53】Featured Reproducing Kernel Banach Spaces for Learning and Neural Networks
标题：学习和神经网络的特征再生核Banach空间
链接：https://arxiv.org/abs/2602.07141
【54】Theory of Space: Can Foundation Models Construct Spatial Beliefs through Active Exploration?
标题：空间理论：基础模型能否通过主动探索构建空间信念？
链接：https://arxiv.org/abs/2602.07055
备注：published at iclr 2026
【55】Where Not to Learn: Prior-Aligned Training with Subset-based Attribution Constraints for Reliable Decision-Making
标题：哪里不可以学习：具有基于子集的归因约束的优先一致训练，以实现可靠的决策
链接：https://arxiv.org/abs/2602.07008
【56】Attractor Patch Networks: Reducing Catastrophic Forgetting with Routed Low-Rank Patch Experts
标题：吸引者补丁网络：通过路由低级别补丁专家减少灾难性遗忘
链接：https://arxiv.org/abs/2602.06993
备注：9 pages. Code (APN implementation in nanoGPT transformer): https://github.com/shankch/nanoGPT-apn (baseline: https://github.com/karpathy/nanoGPT) Data prep: https://github.com/karpathy/nanoGPT/tree/master/data/shakespeare_char and https://github.com/karpathy/nanoGPT/tree/master/data/shakespeare
【57】When do neural ordinary differential equations generalize on complex networks?
标题：神经常微方程什么时候能在复杂网络上推广？
链接：https://arxiv.org/abs/2602.08980
【58】Provably robust learning of regression neural networks using $β$-divergences
标题：使用$β$-偏差进行回归神经网络的可证明鲁棒学习
链接：https://arxiv.org/abs/2602.08933
备注：Pre-print, under review
【59】DNS: Data-driven Nonlinear Smoother for Complex Model-free Process
标题：DNS：数据驱动的非线性平滑器，用于复杂无模型流程
链接：https://arxiv.org/abs/2602.08560
【60】Trajectory Stitching for Solving Inverse Problems with Flow-Based Models
标题：基于流的模型求解反问题的轨迹缝合
链接：https://arxiv.org/abs/2602.08538
【61】Capturing the Topological Phase Transition and Thermodynamics of the 2D XY Model via Manifold-Aware Score-Based Generative Modeling
标题：通过基于Manifold感知分数的生成式建模捕捉2D XY模型的布局相转变和热力学
链接：https://arxiv.org/abs/2602.07548
备注：16 pages, 13 figures
【62】Machine learning enhanced data assimilation framework for multiscale carbonate rock characterization
标题：用于多尺度碳酸盐岩定性的机器学习增强数据同化框架
链接：https://arxiv.org/abs/2602.06989
其他(86篇)
【1】Next-Gen CAPTCHAs: Leveraging the Cognitive Gap for Scalable and Diverse GUI-Agent Defense
标题：下一代验证码：利用认知差距实现可扩展和多样化的GUI-Agent防御
链接：https://arxiv.org/abs/2602.09012
备注：Project page at https://greenoso.github.io/NextGen-CAPTCHAs_webpage/
【2】DirMoE: Dirichlet-routed Mixture of Experts
标题：DirMoE：Dirichlet路线的专家混合体
链接：https://arxiv.org/abs/2602.09001
【3】MotionCrafter: Dense Geometry and Motion Reconstruction with a 4D VAE
标题：MotionCrafter：使用4D VAE的密集几何和运动重建
链接：https://arxiv.org/abs/2602.08961
备注：Project page: https://ruijiezhu94.github.io/MotionCrafter_Page
【4】DynamiQ: Accelerating Gradient Synchronization using Compressed Multi-hop All-reduce
标题：DynamiQ：使用压缩多跳全归约加速梯度同步
链接：https://arxiv.org/abs/2602.08923
备注：18 pages, 18 figures
【5】Stress-Testing Alignment Audits With Prompt-Level Strategic Deception
标题：具有预算级战略欺骗的压力测试一致审计
链接：https://arxiv.org/abs/2602.08877
【6】Magnitude Distance: A Geometric Measure of Dataset Similarity
标题：幅度距离：数据集相似性的几何测量
链接：https://arxiv.org/abs/2602.08859
【7】Kirin: Improving ANN efficiency with SNN Hybridization
标题：麒麟：通过SNN杂交提高NN效率
链接：https://arxiv.org/abs/2602.08817
【8】Permissive-Washing in the Open AI Supply Chain: A Large-Scale Audit of License Integrity
标题：开放人工智能供应链中的许可清洗：对许可完整性的大规模审计
链接：https://arxiv.org/abs/2602.08816
备注：13 pages, 2 figures, 10 tables
【9】$\texttt{lrnnx}$: A library for Linear RNNs
标题：$ extttt {lrnnx}$：线性RNN的库
链接：https://arxiv.org/abs/2602.08810
备注：EACL Student Research Workshop 2026
【10】On the Expressive Power of GNNs for Boolean Satisfiability
标题：论GNN对布尔可满足性的表达能力
链接：https://arxiv.org/abs/2602.08745
备注：Accepted at ICLR 2026
【11】Welfarist Formulations for Diverse Similarity Search
标题：用于多样化相似性搜索的福利主义公式
链接：https://arxiv.org/abs/2602.08742
【12】LLaDA2.1: Speeding Up Text Diffusion via Token Editing
标题：LLaDA 2.1：通过代币编辑加速文本传播
链接：https://arxiv.org/abs/2602.08676
备注：11 pages, 3 figures
【13】From Robotics to Sepsis Treatment: Offline RL via Geometric Pessimism
标题：从机器人技术到败血症治疗：通过几何悲观主义进行离线RL
链接：https://arxiv.org/abs/2602.08655
备注：10 pages, 8 figures
【14】We Should Separate Memorization from Copyright
标题：我们应该将电子化与版权分开
链接：https://arxiv.org/abs/2602.08632
【15】CauScale: Neural Causal Discovery at Scale
标题：CauScale：神经因果发现的规模
链接：https://arxiv.org/abs/2602.08629
【16】FairRARI: A Plug and Play Framework for Fairness-Aware PageRank
标题：FairRARI：公平意识PageRank的即插即用框架
链接：https://arxiv.org/abs/2602.08589
【17】An arithmetic method algorithm optimizing k-nearest neighbors compared to regression algorithms and evaluated on real world data sources
标题：与回归算法相比，优化k近邻的算术方法算法，并在现实世界数据源上进行评估
链接：https://arxiv.org/abs/2602.08577
备注：Nature Scientific Reports
【18】Rho-Perfect: Correlation Ceiling For Subjective Evaluation Datasets
标题：Rho-Perfect：主观评估数据集的相关性上限
链接：https://arxiv.org/abs/2602.08552
【19】When Evaluation Becomes a Side Channel: Regime Leakage and Structural Mitigations for Alignment Assessment
标题：当评估成为侧渠道时：对齐评估的制度泄漏和结构缓解措施
链接：https://arxiv.org/abs/2602.08449
备注：25 pages, 4 figures,
【20】Altruism and Fair Objective in Mixed-Motive Markov games
标题：混合动机马尔科夫博弈中的利他主义和公平目标
链接：https://arxiv.org/abs/2602.08389
【21】OJBKQ: Objective-Joint Babai-Klein Quantization
标题：OJBKQ：双关节Babai-Klein量化
链接：https://arxiv.org/abs/2602.08376
【22】Grounding Generative Planners in Verifiable Logic: A Hybrid Architecture for Trustworthy Embodied AI
标题：将生成规划器置于可验证逻辑中：值得信赖的人工智能的混合架构
链接：https://arxiv.org/abs/2602.08373
备注：Accepted to ICLR 2026. Project page. https://openreview.net/forum?id=wb05ver1k8¬eId=v1Ax8CwI71
【23】Dynamic Regret via Discounted-to-Dynamic Reduction with Applications to Curved Losses and Adam Optimizer
标题：通过折扣到动态还原的动态遗憾，并应用于曲线损失和Adam Optimizer
链接：https://arxiv.org/abs/2602.08372
【24】MemAdapter: Fast Alignment across Agent Memory Paradigms via Generative Subgraph Retrieval
标题：MemAdaptor：通过生成式子图检索跨代理内存范式快速对齐
链接：https://arxiv.org/abs/2602.08369
【25】Fast Flow Matching based Conditional Independence Tests for Causal Discovery
标题：基于快速流匹配的条件独立性测试用于因果发现
链接：https://arxiv.org/abs/2602.08315
【26】Inverting Data Transformations via Diffusion Sampling
标题：通过扩散抽样反数据变换
链接：https://arxiv.org/abs/2602.08267
备注：24 pages, 4 figures
【27】Distribution-Free Robust Functional Predict-Then-Optimize
标题：无分布稳健功能预测-然后优化
链接：https://arxiv.org/abs/2602.08215
【28】A second order regret bound for NormalHedge
标题：一个二阶的遗憾，走向正常对冲
链接：https://arxiv.org/abs/2602.08151
【29】Mutual information and task-relevant latent dimensionality
标题：互信息和任务相关潜在维度
链接：https://arxiv.org/abs/2602.08105
【30】Probability Hacking and the Design of Trustworthy ML for Signal Processing in C-UAS: A Scenario Based Method
标题：C-UAS中信号处理的概率黑客攻击和值得信赖的ML设计：一种基于场景的方法
链接：https://arxiv.org/abs/2602.08086
备注：6 pages, Pre-publication. Copyright 2026 IEEE. Peer Reviewed. Accepted at ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), scheduled for 3-8 May 2026 in Barcelona, Spain
【31】The CAPSARII Approach to Cyber-Secure Wearable, Ultra-Low-Power Networked Sensors for Soldier Health Monitoring
标题：CAPSARII方法，用于士兵健康监测的网络安全可穿戴、超低功耗网络传感器
链接：https://arxiv.org/abs/2602.08080
【32】SiameseNorm: Breaking the Barrier to Reconciling Pre/Post-Norm
标题：SiameseNorm：打破障碍，实现前规范/后规范
链接：https://arxiv.org/abs/2602.08064
【33】Interpretable Fuzzy Systems For Forward Osmosis Desalination
标题：用于正渗透淡化的可解释模糊系统
链接：https://arxiv.org/abs/2602.08050
备注：7 pages, 4 figures, FUZZ-IEEE 2025
【34】FIRE: Frobenius-Isometry Reinitialization for Balancing the Stability-Plasticity Tradeoff
标题：FIRE：Frobenius-等距重新初始化以平衡稳定性-塑性折衷
链接：https://arxiv.org/abs/2602.08040
备注：ICLR'26 (oral)
【35】The Benefits of Diversity: Combining Comparisons and Ratings for Efficient Scoring
标题：多样性的好处：结合比较和评级以实现高效评分
链接：https://arxiv.org/abs/2602.08033
备注：1 table, 5 figures, 8 pages
【36】The Rise of Sparse Mixture-of-Experts:A Survey from Algorithmic Foundations to Decentralized Architectures and Vertical Domain Applications
标题：稀疏混合专家的兴起：从数学基金会到去中心化架构和垂直领域应用程序的调查
链接：https://arxiv.org/abs/2602.08019
【37】A Unified Density Operator View of Flow Control and Merging
标题：流量控制和合并的统一密度运营商观点
链接：https://arxiv.org/abs/2602.08012
【38】From $O(mn)$ to $O(r^2)$: Two-Sided Low-Rank Communication for Adam in Distributed Training with Memory Efficiency
标题：从$O（mn）$到$O（r^2）$：Adam在分布式训练中的双边低秩通信
链接：https://arxiv.org/abs/2602.08007
【39】Tighter Information-Theoretic Generalization Bounds via a Novel Class of Change of Measure Inequalities
标题：通过一类新型的度量变化不等式来更严格的信息论推广界限
链接：https://arxiv.org/abs/2602.07999
备注：41 pages, 1 figure
【40】A Kinetic-Energy Perspective of Flow Matching
标题：流量匹配的动能视角
链接：https://arxiv.org/abs/2602.07928
【41】CausalArmor: Efficient Indirect Prompt Injection Guardrails via Causal Attribution
标题：凯瑟琳装甲：通过因果归因的高效间接即时注射护栏
链接：https://arxiv.org/abs/2602.07918
【42】CausalCompass: Evaluating the Robustness of Time-Series Causal Discovery in Misspecified Scenarios
标题：凯瑟琳指南针：评估错误指定场景中时间序列因果发现的稳健性
链接：https://arxiv.org/abs/2602.07915
【43】Harpoon: Generalised Manifold Guidance for Conditional Tabular Diffusion
标题：Harpoon：条件表格扩散的广义流形指导
链接：https://arxiv.org/abs/2602.07875
备注：Accepted at ICLR 2026
【44】Direct Soft-Policy Sampling via Langevin Dynamics
标题：通过Langevin Dynamics直接软政策抽样
链接：https://arxiv.org/abs/2602.07873
【45】Riemannian MeanFlow
标题：Riemann MeanFlow
链接：https://arxiv.org/abs/2602.07744
【46】The Laplacian Keyboard: Beyond the Linear Span
标题：拉普拉斯键盘：超越线性跨度
链接：https://arxiv.org/abs/2602.07730
备注：28 pages, 17 figures
【47】Towards Robust Scaling Laws for Optimizers
标题：优化器的鲁棒缩放定律
链接：https://arxiv.org/abs/2602.07712
【48】ElliCE: Efficient and Provably Robust Algorithmic Recourse via the Rashomon Sets
标题：ElliCE：通过罗生门集高效且可证明稳健的数学追索
链接：https://arxiv.org/abs/2602.07674
【49】Continuous Program Search
标题：持续程序搜索
链接：https://arxiv.org/abs/2602.07659
【50】Rational Transductors
标题：理性传感器
链接：https://arxiv.org/abs/2602.07599
【51】Beyond Arrow: From Impossibility to Possibilities in Multi-Criteria Benchmarking
标题：超越箭头：多标准基准中从不可能到可能
链接：https://arxiv.org/abs/2602.07593
【52】$\partial$CBDs: Differentiable Causal Block Diagrams
链接：https://arxiv.org/abs/2602.07581
【53】Compact Conformal Subgraphs
标题：紧凑共形子图
链接：https://arxiv.org/abs/2602.07530
【54】Deriving Neural Scaling Laws from the statistics of natural language
标题：从自然语言的统计信息中推导神经标度律
链接：https://arxiv.org/abs/2602.07488
【55】Bandit Allocational Instability
标题：盗贼配置不稳定
链接：https://arxiv.org/abs/2602.07472
【56】Learned Finite Element-based Regularization of the Inverse Problem in Electrocardiographic Imaging
标题：心电图成像反问题的基于学习有限单元的正规化
链接：https://arxiv.org/abs/2602.07466
【57】On the Importance of a Multi-Scale Calibration for Quantization
标题：论量化多尺度校准的重要性
链接：https://arxiv.org/abs/2602.07465
备注：ICASSP 2026
【58】Sign-Based Optimizers Are Effective Under Heavy-Tailed Noise
标题：基于符号的优化器在重尾噪音下有效
链接：https://arxiv.org/abs/2602.07425
备注：Code available at https://github.com/Dingzhen230/Heavy-tailed-Noise-in-LLMs
【59】UTOPIA: Unlearnable Tabular Data via Decoupled Shortcut Embedding
标题：UTOPIA：通过去耦合队列嵌入无法学习的表格数据
链接：https://arxiv.org/abs/2602.07358
【60】Semantic Search At LinkedIn
标题：LinkedIn上的语义搜索
链接：https://arxiv.org/abs/2602.07309
【61】Robust Ultra-High-Dimensional Variable Selection With Correlated Structure Using Group Testing
标题：使用群测试进行具有相关结构的鲁棒性超高维变量选择
链接：https://arxiv.org/abs/2602.07258
备注：57 Pages, 5 Figures, 4 Tables
【62】SpecAttn: Co-Designing Sparse Attention with Self-Speculative Decoding
标题：SpecAttn：与自我思考解码共同设计稀疏注意力
链接：https://arxiv.org/abs/2602.07223
【63】Collaborative and Efficient Fine-tuning: Leveraging Task Similarity
标题：协作且高效的微调：利用任务相似性
链接：https://arxiv.org/abs/2602.07218
【64】Exactly Computing do-Shapley Values
标题：精确计算do-Shapley值
链接：https://arxiv.org/abs/2602.07203
【65】Risk-Sensitive Exponential Actor Critic
标题：风险敏感指数演员评论家
链接：https://arxiv.org/abs/2602.07202
备注：To appear at AAAI 2026
【66】Free Energy Mixer
标题：自由能源搅拌机
链接：https://arxiv.org/abs/2602.07160
备注：Camera-ready version. Accepted at ICLR 2026
【67】Mimetic Initialization of MLPs
标题：MLP的模拟收件箱
链接：https://arxiv.org/abs/2602.07156
【68】Beyond Pooling: Matching for Robust Generalization under Data Heterogeneity
标题：超越池化：数据异类下的稳健概括匹配
链接：https://arxiv.org/abs/2602.07154
备注：AISTATS 2026
【69】On Randomness in Agentic Evals
标题：论爆炸事件中的随机性
链接：https://arxiv.org/abs/2602.07150
【70】TACIT: Transformation-Aware Capturing of Implicit Thought
标题：TACIT：具有转变意识的内隐思想捕捉
链接：https://arxiv.org/abs/2602.07061
备注：25 pages, 7 figures
【71】ShapBPT: Image Feature Attributions Using Data-Aware Binary Partition Trees
标题：ShapBPT：使用数据感知二进制分区树的图像特征属性
链接：https://arxiv.org/abs/2602.07047
备注：AAAI-2026
【72】AI for Sustainable Data Protection and Fair Algorithmic Management in Environmental Regulation
标题：人工智能可持续数据保护和环境监管中的公平统计管理
链接：https://arxiv.org/abs/2602.07021
备注：Presented at National Conference on Navigating The Intersection of Artificial Intelligence and Law: Ethical and Legal Horizons, 29 September 2024, pp. 91-106
【73】Curriculum-Learned Vanishing Stacked Residual PINNs for Hyperbolic PDE State Reconstruction
标题：双曲PDL状态重建的课程学习消失堆叠剩余PINN
链接：https://arxiv.org/abs/2602.06996
【74】NLP Sampling: Combining MCMC and NLP Methods for Diverse Constrained Sampling
标题：NLP采样：结合MCMC和NLP方法进行多样化约束采样
链接：https://arxiv.org/abs/2407.03035
【75】Universal Coefficients and Mayer-Vietoris Sequence for Groupoid Homology
标题：群群类同系的普适系数和Mayer-Vietoris序列
链接：https://arxiv.org/abs/2602.08998
备注：Master's thesis
【76】Winner's Curse Drives False Promises in Data-Driven Decisions: A Case Study in Refugee Matching
标题：赢家的诅咒导致数据驱动决策中的虚假承诺：难民匹配案例研究
链接：https://arxiv.org/abs/2602.08892
【77】Constructive conditional normalizing flows
标题：建设性的有条件正常化流程
链接：https://arxiv.org/abs/2602.08606
【78】Schrödinger bridge problem via empirical risk minimization
标题：基于经验风险最小化的薛定谔桥问题
链接：https://arxiv.org/abs/2602.08374
【79】Is Flow Matching Just Trajectory Replay for Sequential Data?
标题：流量匹配只是序列数据的轨迹回放吗？
链接：https://arxiv.org/abs/2602.08318
备注：51 pages
【80】A Statistical Framework for Alignment with Biased AI Feedback
标题：与有偏见的人工智能反馈保持一致的统计框架
链接：https://arxiv.org/abs/2602.08259
【81】Discrete Adjoint Schrödinger Bridge Sampler
标题：离散伴随薛定汉桥采样器
链接：https://arxiv.org/abs/2602.08243
【82】Information Geometry of Absorbing Markov-Chain and Discriminative Random Walks
标题：吸收马尔科夫链的信息几何和区分性随机游动
链接：https://arxiv.org/abs/2602.08185
【83】BFTS: Thompson Sampling with Bayesian Additive Regression Trees
标题：BFTS：使用Bayesian加法回归树的汤普森抽样
链接：https://arxiv.org/abs/2602.07767
【84】How does longer temporal context enhance multimodal narrative video processing in the brain?
标题：更长的时间背景如何增强大脑中的多模式叙事视频处理？
链接：https://arxiv.org/abs/2602.07570
备注：22 pages, 15 figures
【85】Discrete Adjoint Matching
标题：离散伴随匹配
链接：https://arxiv.org/abs/2602.07132
备注：ICLR 2026
【86】BERT Learns (and Teaches) Chemistry
标题：BERT学习（并教授）化学
链接：https://arxiv.org/abs/2007.16012
备注：10 pages, 5 figures
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/192800
本文地址：http://www.python88.com/topic/192800