AI论文速递 2026年08月04日（HuggingFace Daily Papers）

数据来源：https://huggingface.co/papers 采集时间：2026-08-04

📌 重点关注

1. LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger | arXiv — 【重点关注】 Multimodal agents for visual question answering increasingly operate as multi...

   💡 证据溯源约束Agent推理链，对构建可信AI Agent极具参考价值

2. From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement | arXiv — 【重点关注】 Reinforcement Learning with Verifiable Rewards (RLVR) has driven recent progr...

   💡 自验证奖励让LLM在开放任务中自我进化，RL新范式值得跟踪

3. Constitutional Midtraining: Content Presence Drives Alignment Gains | arXiv — 【重点关注】 Post-training alignment is often shallow, eroding under fine-tuning. Whether ...

   💡 中训练阶段注入对齐优于传统后训练，对齐工程新思路

📋 其他值得关注

1. Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions | arXiv — Deep Research agents extend LLM-based assistants into long-horizon workflows ...
2. EMBL AI Librarian: Life-Sciences Knowledge Layer for AI Agents | arXiv — The web is increasingly accessed by AI agents rather than humans. Every agent...
3. Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability | arXiv — Deployed LLM agents increasingly keep their long-term memory as a filesystem:...
4. SGTP: Sampling-based Game-Theoretic Planning for Real-Time Multi-Vehicle Autonomous Racing | arXiv — Autonomous multi-vehicle racing requires real-time planning of diverse compet...
5. See2Think: Do Multimodal Models Really Use Intermediate Visual States? | arXiv — Multimodal large language models increasingly use sketches, annotations, tool...
6. ScrambleToolBench: Agents Search Exhaustively Even When Their Own Map Points to the Next Step | arXiv — To operate robustly in open-world environments, autonomous agents should be a...
7. Weak-to-Strong On-Policy Distillation | arXiv — On-policy distillation (OPD), which aligns a student with the teacher's token...