Tag
TinyAstral evolves a frozen teacher's agent harness via recursive self-improvement with statistical acceptance tests, then distills the gains into a 4B student that matches the teacher's RoboTwin success (62.8% vs 63.6%) using about one fifth of the accelerator time.
The paper proposes TARGRS, a test-time detoxification framework that learns token-level, transition-aware reward guidance over toxicity trajectories in LLM latent space, achieving up to 69.38% toxicity reduction on RealToxicityPrompts and extending to vision-language models with minimal inference overhead.
DR Pixel is an open-source end-to-end multimodal deep research agent that retrieves and reasons directly over rendered document page images instead of OCR-extracted text, introducing counterfactual credit reweighting for agentic supervised finetuning. The approach outperforms OCR-based baselines on DRPixelBench across models like GPT-5.4 and Gemini-3.5-Flash, boosting retrieval recall by ~15 points and accuracy by 10+ points after training.
Hugging Face weekly roundup of top research papers covering world models, streaming video agents, robot policies, and efficient diffusion VLMs, highlighting works like Rollback, Hy-Nomas, OmniShowcase, PDRVLA, and StreamVibe.
Microsoft releases Mage-VL-Cowork, a compact 4B vision-language model that combines GUI interaction and terminal/code execution in a single policy, achieving 66.0% on OSWorld-Verified — the best result at the 4B scale — via a five-stage curriculum and online RL.
The paper introduces BedBench, a controlled benchmark of thirteen procedurally generated 3D reasoning tasks, showing that vision-language models can make correct individual judgments yet fail to recover and consistently apply exact per-color counts (7.2% vs. 93.3% human accuracy), and argues for evaluating exact correctness, consistency, and visual dependence rather than accuracy alone.
CaPtain is an open post-training pipeline that turns small open vision-language models into capable code-as-policy robot agents via SFT and multi-turn RL with a physical critic, achieving strong LIBERO and SimplerEnv results without frontier models or VLA backbones.
The paper proposes an online reinforcement learning framework that trains UI navigation agents on live iOS apps, decoupling them into a planning agent and a UI grounding model with a novel neighbor-conditioned value estimation method (GiGPPO) that improves critic stability and reduces inference costs. Their 8B RL agents outperform SFT baselines and match Gemini-3.1-Pro on real iOS apps while showing strong cross-platform generalization to Android.
Andon Labs 发布 Blueprint-Bench 2，评估 AI 在物理 3D 理解上的表现，结果显示 AI 已在该基准上追平人类水平。
DocSkill提出一种带技能编排策略的智能体框架，用于多语言多跳文档问答，通过MCTS离线探索轨迹并结合SFT/DPO训练策略，同时发布包含60K QA对、覆盖10种语言的MMHop-DocBench数据集。
A research paper exploring how vision-language models (VLMs) use and interpret visual symbols, finding that simple visuals like a red heart can implicitly steer model behavior (e.g., toward recursive coding solutions) and that models can develop creative visual communication strategies that diverge from human conventions.
论文提出 QViD，一种免训练的查询感知视觉token剪枝方法，通过发现查询-视觉交互矩阵的低秩结构并分解出主导潜在交互模式，在图像与视频理解基准上以相同token预算持续提升性能，尤其在高压缩率下增益显著。
Google 等机构提出 StreamVibe 框架，包含 StreamVibe-Bench 基准、StreamVibe-VL 流式主动模型与 StreamVibe-Harness 分层记忆机制，使视觉语言模型能从连续视频观测中推断隐式且动态变化的用户意图，并在响应时机与主动交互上分别提升 10.7% 和 12.3%。
StreamVibe introduces a benchmark (StreamVibe-Bench), a streaming proactive vision-language model (StreamVibe-VL), and a hierarchical memory harness to enable intent-aware interaction with implicit and evolving user goals, improving response quality by 10.7% over baselines and proactive interaction by 12.3%.
该论文提出“多证据整合差距”概念，发现视觉语言模型（如Gemma-3-27B）能正确恢复所有局部事实，但在需要联合约束满足的任务上全局有效率极低（仅19.75%），并通过隐藏态探针和激活替换实验揭示了跨层深度的可干预性。
该论文提出实体级分析作为补充视角来评估历史文档OCR质量，使用Berrutti档案馆175份文档及5,209条人工标注实体，比较了序列识别OCR系统、OCR专用视觉-语言模型和通用视觉-语言模型，发现不同评价标准下OCR排名会变化。
IRON-Meme is a human-annotated benchmark of 1,110 political memes with a social-science-grounded framework for classifying social identities, viewpoints, and narrative roles; evaluation of nine vision-language models shows frontier and reasoning models perform best, while template background knowledge does not improve role classification.
This paper formulates indoor radio map estimation as a collaborative in-context learning problem and proposes Poly-Ripple, a framework where multiple clients use large autoregressive vision models without model updates, employing ripple-ordered autoregressive modeling and importance-aware aggregation to outperform ICL baselines.
The paper proposes EvidenceConnect, a data synthesis framework that trains VLMs via RLVR to better connect visual evidence, yielding the ZwZ2 models with improved spatial perception, long-context understanding, and general visual reasoning across diverse downstream tasks.
本文提出 Outcome-Conditioned Visual Credit (OVC) 框架，通过分离视觉关联测量与 token 级信用分配，改进多模态强化学习中的奖励分配，在三个模型的六项基准上较 GSPO 提升 1.86-2.39 个百分点。