今日速览
| 序号 | 标题 | 来源 | 日期 | 主题 | 推荐等级 |
|---|---|---|---|---|---|
| 1 | QVal: Cheaply Evaluating Dense Supervision Signals for Long-Horizon LLM Agents | arXiv | 2026-06-30 | AI-Agent | 高 |
| 2 | Generative Skill Composition for LLM Agents | arXiv | 2026-06-30 | AI-Agent | 高 |
重点论文与技术动态
1. QVal: Cheaply Evaluating Dense Supervision Signals for Long-Horizon LLM Agents
- 来源：arXiv
- 日期：2026-06-30
- 作者/机构：Sergio Hernández-Gutiérrez, Matteo Merler, Ilze Amanda Auzina, Joschka Strüber, Ameya Prabhu, Matthias Bethge
- 主题标签：
AI-Agent,arXiv - 推荐等级：高
- 分类：cs.LG, cs.AI, cs.CL
一句话结论
QVal通过Q值对齐直接评估密集监督信号，无需训练即可实现方法公平比较。
一段话。
该测试平台避免了传统评估中训练工程干扰，使密集监督方法能在统一基准上高效比较，显著降低评估成本。
核心内容
- 现有密集监督方法评估依赖下游训练性能，成本高且混淆监督质量与工程因素，导致方法不可比。
- QVal测量信号与强参考策略Q值的对齐度，实现训练-free评估，分离信号质量与工程选择。
- 实验表明简单提示基线持续优于文献方法，性能按方法家族显著聚类，且跨模型大小、环境和模态一致。
方法与数据
- QVal基于Q值对齐评估信号，无需训练。
- 基准测试21种方法，覆盖4环境、7方法家族、1200+实验及6个开源模型骨干。
价值判断
- 值得关注：解决密集监督评估瓶颈，推动方法公平比较与研究进展。
- 可复用点：设计可扩展至新环境和方法，支持研究者训练前迭代。
- 局限/待核查：摘要未明确。
摘要
LLM agents increasingly act over long horizons, where a single trajectory can contain hundreds or thousands of actions. In these settings, outcome-only rewards provide too sparse guidance, failing to inform the model about the goodness of intermediate actions. Dense supervision methods aim to solve this problem by scoring intermediate steps, from intrinsic confidence to self-distillation and embedding similarities. However, it is common practice to evaluate them by measuring the downstream performance of a training pipeline that integrates them. This is expensive, conflates supervision quality with training engineering confounders, and renders different methodological families requiring distinct training setups incomparable. As a result, dense supervision methods are rarely benchmarked on common ground. We introduce QVal, a training-free testbed for directly evaluating dense supervision signals. Given a state-action pair, QVal measures how well a method's score is Q-aligned: whether it orders actions according to the Q-values of a strong reference-policy. This lets us compare signals before any training run and separate signal quality from other engineering choices. We instantiate QVal as QVal-v1.0, benchmarking 21 dense supervision methods across four diverse environments and seven methodological families, with over 1.2K evaluation experiments across six open-weight model backbones. We find that simple prompting baselines consistently outperform recent dense supervision methods from the literature, and that performance clusters strongly by family. These findings hold across model sizes, environments, and observation modalities. QVal is designed to be easily extensible to new environments and methods, enabling researchers to iterate on dense supervision methods before any training run.2. Generative Skill Composition for LLM Agents
- 来源：arXiv
- 日期：2026-06-30
- 作者/机构：Xinyu Zhao, Zhen Tan, Vaishnav Tadiparthi, Nakul Agarwal, Kwonjoon Lee, Ehsan Moradi Pari
- 主题标签：
AI-Agent,arXiv - 推荐等级：高
- 分类：cs.CL
一句话结论
SkillComposer通过结构化技能组合显著提升LLM代理任务成功率，在SkillsBench测试中使GPT-5.2-Codex和Gemini-3-Pro-Preview的通过率分别提高23.1%和18.2%。
核心内容
- 技能组合需联合决策技能选择、数量和执行顺序，三者不可解耦，现有方法忽略其结构特性。
- SkillComposer将结构化技能组合实例化为任务条件的技能序列预测，使用约束自回归解码器实现子集、数量和顺序的联合生成。
- 评估显示，SkillComposer超越top-3检索方法，匹配黄金技能检索上界，且提示令牌成本更低。
方法与数据
- 方法：采用约束自回归解码器预测技能序列，自然捕获技能间依赖。
- 数据：基于真实人类策划的技能库构建训练集。
价值判断
- 值得关注：有效解决技能库增长引发的技能组合瓶颈问题。
- 可复用点：技能序列预测框架可推广至其他LLM代理任务。
- 局限/待核查：摘要未明确。