Research Papers 论文研究 1d ago Updated 15h ago 更新于 15小时前 46

Learning Stateful Predictive Knowledge From Experience 从经验中学习状态预测性知识

Current LLM agents rely on trajectory-level reflection that produces brittle, path-dependent heuristics based on episodic hindsight rather than predictive foresight Stateful Knowledge Learning (SKL) shifts focus to maintaining explicit, declarative predictive assessments anchored to states Two scaling algorithms are introduced: self-distillation (SKL-SD) and reinforcement learning (SKL-RL) for autonomous knowledge extraction Stateful knowledge provides three key advantages: granularity, enhanced 提出Stateful Knowledge Learning (SKL)框架,将LLM agents从轨迹级反思转向维护状态锚定的预测性知识 设计了两种训练算法:基于自蒸馏的SKL-SD和基于强化学习的SKL-RL,使agent能自主从经验中提取状态化预测知识 在WebShop、ScienceWorld交互环境和ChessPuzzles复杂推理任务上验证,SKL显著优于现有基于反思的训练范式

60
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Current LLM agents rely on trajectory-level reflection that produces brittle, path-dependent heuristics based on episodic hindsight rather than predictive foresight
  • Stateful Knowledge Learning (SKL) shifts focus to maintaining explicit, declarative predictive assessments anchored to states
  • Two scaling algorithms are introduced: self-distillation (SKL-SD) and reinforcement learning (SKL-RL) for autonomous knowledge extraction
  • Stateful knowledge provides three key advantages: granularity, enhanced generalization, and knowledge bootstrapping
  • Experiments on WebShop, ScienceWorld, and ChessPuzzles show SKL significantly outperforms reflection-based training paradigms

Why It Matters

This work addresses a fundamental limitation in how LLM agents learn from experience—current reflection-based approaches produce fragile heuristics tied to specific trajectories. By enabling agents to extract and maintain state-grounded predictive knowledge, SKL offers a more robust and generalizable path toward autonomous learning systems that can transfer insights across diverse scenarios.

Technical Details

  • Core Concept: SKL replaces trajectory-level summarization with "Stateful Knowledge"—explicit declarative predictive assessments anchored to environmental states rather than episodic memories
  • Two Training Algorithms: SKL-SD uses self-distillation to transfer knowledge from experience trajectories, while SKL-RL employs reinforcement learning to train agents to autonomously extract and leverage state-grounded predictive knowledge for policy making
  • Benchmark Environments: Evaluated across interactive web navigation (WebShop), scientific reasoning (ScienceWorld), and complex strategic reasoning (ChessPuzzles)
  • Key Advantages Demonstrated: Stateful knowledge provides finer granularity than trajectory-level reflection, enhances generalization across unseen scenarios, and enables knowledge bootstrapping where learned predictions accelerate future learning
  • arXiv Reference: 2607.28638, submitted May 19, 2026, by Yan Song et al.

Industry Insight

  • The shift from episodic hindsight to predictive foresight represents a paradigm change in agent learning—organizations building LLM agents should prioritize stateful knowledge architectures over pure reflection-based approaches for better long-term adaptability
  • Self-distillation and RL-based knowledge extraction offer complementary paths: SKL-SD is more sample-efficient for resource-constrained settings, while SKL-RL scales better for complex environments requiring autonomous discovery
  • As interactive AI agents move toward real-world deployment, the ability to bootstrap knowledge from experience rather than relying on hand-engineered heuristics will be critical for reducing deployment costs and improving robustness

TL;DR

  • 提出Stateful Knowledge Learning (SKL)框架,将LLM agents从轨迹级反思转向维护状态锚定的预测性知识
  • 设计了两种训练算法:基于自蒸馏的SKL-SD和基于强化学习的SKL-RL,使agent能自主从经验中提取状态化预测知识
  • 在WebShop、ScienceWorld交互环境和ChessPuzzles复杂推理任务上验证,SKL显著优于现有基于反思的训练范式

为什么值得看

该研究揭示了当前LLM agent经验学习的核心缺陷——依赖"后见之明"而非"预测性前瞻",为agent学习机制提供了新的理论视角。SKL框架提出的状态化知识表示方法,有望推动agent从被动反思向主动预测性学习转变,对构建更鲁棒、泛化能力更强的智能体具有重要参考价值。

技术解析

  • 核心问题诊断:现有agent依赖轨迹级反思(trajectory-level reflection)提取经验,本质是情景后见之明(episodic hindsight),产生脆弱且路径依赖的启发式规则,缺乏预测性前瞻能力。
  • SKL框架设计:将agent焦点从轨迹级总结转移到维护"状态知识"(Stateful Knowledge)——即明确的、声明式的、锚定在状态上的预测性评估,实现知识粒度的细化和泛化能力的提升。
  • 两种训练算法:SKL-SD通过自蒸馏方式训练agent从经验中自主提取状态化预测知识;SKL-RL则利用强化学习框架,使agent学会利用状态知识指导策略制定。
  • 实验验证:在WebShop(电商交互)、ScienceWorld(科学探索)和ChessPuzzles(复杂推理)三个基准上测试,SKL方法显著超越当前基于反思的训练范式。

行业启示

  • Agent学习范式的转变:从"事后反思"到"事前预测"的知识构建思路,为agent经验学习提供了新的技术路线,值得在agent系统设计中优先考虑。
  • 状态化知识的工程价值:将预测性知识锚定在状态而非轨迹上,可实现知识的模块化复用和跨任务迁移,对降低agent训练成本和提升部署效率具有实际意义。
  • 自蒸馏与强化学习的结合:SKL-SD和SKL-RL两种算法路径展示了不同训练策略在状态知识学习上的适用性,为后续研究提供了可复用的方法论参考。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Agent Agent Research 科学研究 Training 训练