看学术

GUI-JEPA: Candidate-Conditioned JEPA Latents as a Fast Action Policy for Mobile GUI Agents

RubricAtlas: Mining Type-Level Rubrics from Expert Reports to Evaluate Deep Research Agents

Nothing Left Behind: Finding Evidence from a Large Randomized Evaluation Corpus

Developing a Speech Language Model for Basque

arXiv 论文评测决策模型用于 RAG 知识冲突分型的效果

论文研究用 typed decision models 替代 LLM 对 RAG 检索源冲突做分型分类，基于 CONFLICTS 基准和原论文 prompt，对一家推理网关上 7 家提供商的 11 个决策模型与 Gemini 系列在相同输入上评测。

TrawMem: Reframing Agent Memory as Task-Oriented Role-Adaptive Workspaces

DocXEval：多模态文档理解可解释性系统评测框架

DocXEval 提出首个面向多模态（文本、版面、图像）文档理解的可解释性（XAI）评测框架，系统评测 13 种 XAI 方法、3 类任务、8 个数据集、4 个模型与 17 项评测指标，发现仅少数方法在全部指标上表现稳定，其中 Occlusion 整体最忠实、最可解释。元评测显示多项忠实度与复杂度指标存在中到强一致性，同时识别出部分不可靠指标。代码已公开发布以支持可复现性。

JevRec: Linguistic Encoding with Jev-like Decision Modeling for LLM-based Generative Recommendation

Self-Keyed In-Parameter Memorization for External Memory-Free Dialogue Recall in Frozen Language Models

Nonlinear Direct Feedback Alignment for Scalable Backpropagation-Free Training

Neuronal Identity as an Organizational Basis for Analyzing Neural Population Dynamics

Generative Refinement Learning for Temporal Interpolation in Scientific Machine Learning

Robust LLM Data Selection via Content-Invariant Gradient Attribution

Frequency-Uniform Log-Regret Reinforcement Learning for Diffusions

Batching the Reflective Optimization Loop: Parallel Proposals Make GEPA Faster and Better

Does LLM Pre-Training Typically Occur at the Edge of Stability?

Select-TTT: Query-Conditioned Selective Fast Weight Updates for In-Place Test-Time Training

API-TAIL：通过温度对齐似然尾部检测LLM API中的模型替换

arXiv论文提出API-TAIL，一种检测商业大语言模型API部分模型替换的典型性审计方法。该方法先对齐单一有效温度以吸收批次级分布偏移，再标记超出阈值的尾部响应，从而在合法解码配置构成的复合零假设下检测低比例替换。在真实商业API上的评测显示，API-TAIL能控制误报并以较小查询预算检测低速率替换，优于基于rank和MMD的现有审计方法。

LLM knowledge for improved Out-of-Distribution Tabular Prediction - A Cross-Validated Regularisation Approach

SetGRPO: Calibrating Grader-Induced Policy Optimization

Variable-Length Feedback Coding With Universal Decoding and Bandit Channel Selection

TAPESTRY: A Single Backbone for Video Generation and Robot Control

Joint Distributional Copula Regression for Insurance Claim Frequency and Severity via Boosting

When All Rollouts Fail: Policy-Compatible Joint Decoding for Hard-Problem Reasoning

Boundary Geometry Determines Optimal Shallow ReLU Approximation under Weighted Coefficient Constraints

Evaluating AlphaEarth Embeddings for Crop Classification: Separability, Temporal Dynamics, and Cross-Regional Transferability

HorizON: A Multi-Horizon Deep Operator Network for Property Prediction of Organic Battery Stacks

Intrinsic Information-Theoretic Analysis of ReLU Nets

Sample-and-Score: Surrogate-Guided Coordinated Spatial Reuse in IEEE 802.11 Networks

Tail Cues, Principal Corrections: Plug-and-Play Rectification for Open-Set Test-Time Adaptation

RAG-CSI: Global Local Retrieval via Joint State and Dynamics Matching for CSI Prediction

Hyperparameter and Neural Architecture Search via Evolutionary Optimization for Reinforcement Learning: HPNAS-RL

PREEN: Exemplar-Based Color Style Transfer with Smartphone Editor Presets

HiLo-RD: Cross-Budget Self-Distillation for Masked Diffusion Language Models

图像扩散模型 LoRA 适配的来源签名检测协议（arXiv 预印本）

一篇 arXiv 预印本提出一套协议，用于检测在线服务是否在使用标准图像扩散模型的 LoRA 适配版本，并在开放世界场景下（候选适配与替代模型未知）区分基础模型、LoRA 适配与其他模型家族。该方法基于假设检验，仅利用模型生成图像集合的内在维度构建黑盒签名，区别于以往在已知候选模型列表内做模型识别的工作。论文目前处于 Under review 状态，作者为 Tobias Cheung。

Gaussian Process Bandits for Constraint Black Box Lever Optimization

TDLM: Adaptive Temporal Integration for Diffusion Language Models

Algorithmic Monoculture in Machine Learning-based Site-Specific Fertilizer Recommendation

DCKV: Learned Transform Coding of the Prompt KV Cache in Hybrid Language Models

Provably Efficient Q-learning for Robust Reinforcement Learning