看学术

FrontierCS 2: A Benchmark for Computer Science R&D

Mastering Model Fit: Applications of the InterModel Vigorish for Interpretable Machine Learning

CORD: Conflict-Aware Optimization for Joint RL and Distillation in Agents

The Price of Current-Effect Inference in Nonstationary Multi-Armed Bandits

SCDM: Scalable Causal Discovery in Nonlinear Temporal Systems with Meta-Learning

EPIC: Rethinking Trust Regions for Autoregressive LLM Policies via Exact-KL Prefix Importance-Ratio Clipping

Mitigating Model Collapse through Dreaming Learning

ExGrad：对抗训练中显式干净梯度相减，超越单纯抑制

ExGrad（外推梯度重组）通过构造干净输入、对抗输入与 AWP 梯度并测量其边界响应，在无约束的有界仿射域上求解三变量系数选择问题。在 CIFAR-10 上，负干净系数随仿射半径增大而迅速占主导；将干净系数置零并不足够，适度相减可提升 PGD-20，更强相减则以干净精度换取进一步增益。

Looped Uniform Diffusion Language Models

Commit or Abstain？System One 决策模型、LLM 与缺失的法律依据

研究将 CaseHOLD 扩展出弃权条件（5,314 题中 1,670 题替换正确判决），对比四个 System One 决策模型与六个生成式 LLM：无正确选项时生成式 LLM 仍有 38-86% 强行选一项，最谨慎的决策模型为 30%。

Loopholing All the Way: Language Modeling with pure Self-Conditioning

CulturalRiddles: A Multicultural Benchmark of Riddles

FaultGene: LLM-Assisted Fault-Augmented Test Generation for Hardware-in-the-Loop Validation of Autonomous Driving Systems

From Classifier Heads to Encoder Prototypes: Personalized Federated Learning with Head-Derived Clustering

Circuits of LLM Classifiers

TIDE: Tree-Induced Fine-Grained Diagnostic Evaluation from Model Behaviors

Orbitol: Crystal Generation on Wyckoff Orbits

HiddenLens: A Multi-Layer LLM Framework for Detecting and Resolving Hidden Assumptions in Software Requirements

VX-One：面向实时边界分诊的类型化校准决策头，实测未知攻击诚实性、级联灰区与漂移下的校准

VX-One 是一个 100K-10M 参数的小型多头网络，输出校准门控、含显式 unknown 选项的闭集选择与序数分数三类类型化决策原语，CPU 上 p50 约 0.7-1.7 ms。

Every Alarm Has a Cause: An Open Generator of Telecom Alarm Storms with Exact Root Causes, Bound to NORIA-O, RFC 8632 and TMF642

用稀疏自编码器构建可审计的 AI 文本检测器

研究者提出一条流水线，将微调后的神经文本检测器转换为基于人类可读概念的分类器：直接在检测器残差流上训练稀疏自编码器（SAE），为每个 SAE 特征分配自动解释标签，仅保留描述通过全局保真度阈值的特征，再在其上训练轻量分类器。

Agent Behavior Maps: Locating Agents by What They Would Do, Without Running Them

Velocity Ambiguity Profiles: Time-Resolved Bayes-Risk Diagnostics for Flow Matching

Code Comprehensor: Hierarchical Documentation and Cross-Modal Retrieval for Large Repositories

Minimax Optimality of the NPMLE for Gaussian Mixtures in Hellinger Distance

GuideEvolve: Self-Evolving Agent Framework with Expert Guidance and 3Ts Mechanisms

The role of locality and weight sharing in convolutional neural networks

Stationarity Constraints Improve Both Energies and Densities of Machine-Learned Exchange-Correlation Functionals

Self-normalized Bernstein Inequalities for Regenerative Processes and Applications to Learning Without Independence

A Quasipolynomial Algorithm for $\ell_\infty$ Contractions

论文：企业文档解析丢失格式线索，Agent 需求识别 F1 从 0.690 降至 0.150

Yikai Zhao 的论文指出，主流文档转换器都会丢失作者通过文本外观编码的语义线索，使 Agent 在识别政策文档中被作者特别标注的要求时，micro F1 从 0.690 降至 docling 的 0.150 和纯文本的 0.075，该差距在 5 个 Claude 模型和 GPT-5.6 上均可复现。

LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data

Robust Soft Skill Adaptation via Behavioral Localization in Tool-Using Agents

HGM-Dual: Lineage-Aware Guided Agent Harness Optimization for Long-Horizon Planning

SP-HLM: Sparse-Pose-Driven Motion Generation and Whole-Body Control for Humanoid Loco-Manipulation

Which Differences Matter? Controlled Tests of Rule Transfer, Joint Reliability, and Information Use in Decision Models

SWE-DIVE: Multi-Turn Coding Agent Evaluation with Dynamic User Simulation

From Reasoning to Action: Benchmarking Frontier AI Models on LIBERO-10 Robot Manipulation Tasks

Charts as Epistemic Actions for Experience-Driven Agentic Data Analysis

Parallel AI Agent Collaboration as Graph-Structured Distributed Block Optimization