今日速览
| 序号 | 标题 | 来源 | 日期 | 主题 | 推荐等级 |
|---|---|---|---|---|---|
| 1 | Doomed from the Start: Early Abort of LLM Agent Episodes via a Recall-Controlled Probe Cascade | arXiv | 2026-07-07 | AI-Agent | 高 |
| 2 | AgentTether: Graph-Guided Diagnosis and Runtime Intervention for Reliable LLM Agent Operation | arXiv | 2026-07-07 | AI-Agent | 高 |
重点论文与技术动态
1. Doomed from the Start: Early Abort of LLM Agent Episodes via a Recall-Controlled Probe Cascade
- 来源：arXiv
- 日期：2026-07-07
- 作者/机构：Kai Ruan, Zihe Huang, Ziqi Zhou, Qianshan Wei, Xuan Wang, Hao Sun
- 主题标签：
AI-Agent,arXiv - 推荐等级：高
- 分类：cs.AI
一句话结论
该研究通过隐藏状态探针实现LLM代理任务的早期失败预测，显著减少计算资源浪费。
核心内容
- 失败可从代理内部隐藏表示早期预测：轻量级探针在隐藏激活上，能提前至第一轮交互预测失败，而仅基于行为的评分器效果接近随机。
- 构建了recall-controlled probe cascade框架：每轮一个分布无关校准门，联合优化每轮召回预算，确保成功任务在用户指定全局率下通过所有门，覆盖90%-97%召回目标。
- 在TextCraft数据集上，90%召回目标下节省47.1%（Qwen-2.5-7B）和37.2%（Llama-3.2-3B）计算资源，比单门策略高1.6-1.7倍，且隐藏状态已捕获行为信息。
方法与数据
- 方法：基于隐藏激活的轻量级探针和recall-controlled probe cascade框架。
- 数据：TextCraft数据集，测试了Qwen-2.5-7B和Llama-3.2-3B两个模型。
价值判断
- 值得关注：大幅降低LLM代理任务推理计算开销，提升部署效率。
- 可复用点：probe cascade框架可直接迁移至其他LLM代理系统，无需额外行为特征。
- 局限/待核查：高召回目标的样本复杂性需更多数据验证，摘要未详述具体局限。
摘要
Large language model (LLM) agents solving multi-step tasks frequently commit to trajectories that are doomed to fail, yet continue to consume substantial inference compute before the failure becomes observable. We show that failure is predictable early from the agent's internal representations: lightweight per-round probes on hidden activations anticipate eventual episode failure as early as the first interaction round, where scorers reading only the agent's observable behavior are barely better than chance. We turn this signal into a practical abort cascade: one distribution-free calibrated gate per round, with per-round recall budgets jointly searched so that eventually-successful episodes survive all gates at a user-specified global rate; this episode-level guarantee is the one that matters in deployment, since false-abort risk accumulates across gates. Across two agent models on TextCraft, the cascade meets every recall target from 90% to 97% and, at the 90% target, saves 47.1% +/- 10.3% (Qwen-2.5-7B) and 37.2% +/- 8.8% (Llama-3.2-3B) of inference compute, 1.6--1.7x the best single-gate policy. An otherwise-identical cascade reading only behavior saves roughly half as much, and adding behavioral features to the probe yields no further gain: the hidden states capture what behavior reveals. Finally, we characterize the sample complexity of certifying high recall targets, telling practitioners which recall promises their data can, and provably cannot, back. The code will be released soon.2. AgentTether: Graph-Guided Diagnosis and Runtime Intervention for Reliable LLM Agent Operation
- 来源：arXiv
- 日期：2026-07-07
- 作者/机构：Chenyu Zhao, Shenglin Zhang, Wenwei Gu, Yongqian Sun, Dan Pei, Chetan Bansal
- 主题标签：
AI-Agent,arXiv - 推荐等级：高
- 分类：cs.SE
一句话结论
AgentTether 提出了一种运行时修复框架，能自动诊断LLM代理故障并引导恢复，无需修改底层代理或环境。
核心内容
- 通过抽象运行轨迹为Transition Units并构建依赖感知的Critical Transition Graph，精准定位故障关键子轨迹，避免早期错误传播。
- 结合离线正常行为模型与运行时图检测器进行故障诊断，并利用Repair Memory提供行为范围的修复指导。
- 在Banking领域测试中，成功修复59.04%（Qwen）和65.12%（GPT）的初始失败任务，同时减少代理轮次和端到端令牌消耗。
方法与数据
- 方法：基于图的诊断框架，结合Repair Memory实现行为引导的运行时干预。
- 数据：在261个tau-bench任务上评估，覆盖三个领域（含Banking），使用Qwen3.7-max和GPT-5.4测试跨模型迁移。
价值判断
- 值得关注：解决LLM代理动态轨迹故障问题，区别于静态修复，提升生产可靠性。
- 可复用点：可作为独立层无缝集成现有代理部署，减少浪费的重新执行，无需重新训练代理。
- 局限/待核查：摘要未明确局限。