每日电讯 · 2026-07-02

今日前沿 · Top 5

按对你的价值排序 · 点击任一条进入正文与批注
  1. TRIAGE: Role-Typed Credit Assignment for Agentic Reinforcement Learning
    [huggingface] 用一个角色分类信号改进GRPO的信用分配,提升agent学习效率。
  2. Self-GC: Self-Governing Context for Long-Horizon LLM Agents
    [arxiv] 将LLM agent上下文管理从文本修剪升级为生命周期垃圾回收
  3. MemSyco-Bench: Benchmarking Sycophancy in Agent Memory
    [huggingface] 首个系统量化Agent记忆引发谄媚效应的评测基准。
  4. Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use
    [arxiv] 揭露LLM agent在开放世界中的脆弱性,提出扰动增强微调提升鲁棒性。
  5. Self-Evolving Agents with Anytime-Valid Certificates
    [arxiv] 通过可审计自适应门控实现自我进化agent,冻结基模型保持安全边界