Research Papers 论文研究 4h ago Updated 2h ago 更新于 2小时前 47

Personalizing Large Language Model Agents with Small Policy Models 使用小型策略模型个性化大型语言模型智能体

FABLE (Factorized Adaptive Bandit Layer for Execution) is introduced as a lightweight policy layer that personalizes frozen LLM agents without fine-tuning the underlying model The approach factorizes execution decisions into memory, information-acquisition, and response components, enabling feedback-driven adaptation of user-specific preferences Bayesian contextual Thompson sampling learns residual preferences relative to a fixed default-and-cost score, with theoretical regret bounds under a lin 提出FABLE(Factorized Adaptive Bandit Layer for Execution)框架,通过轻量级外部策略层实现冻结LLM agent的个性化,无需微调主模型 将agent执行决策分解为记忆、信息获取和响应三个维度,通过贝叶斯上下文Thompson采样学习用户特定的残差偏好 在理论层面提供期望遗憾界保证和任何时刻有效的虚假推广控制机制 在个性化推理、受控反馈和可执行工具使用评估中,FABLE在偏好敏感行为上优于纯规则控制,同时保持端到端任务性能竞争力

62
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • FABLE (Factorized Adaptive Bandit Layer for Execution) is introduced as a lightweight policy layer that personalizes frozen LLM agents without fine-tuning the underlying model
  • The approach factorizes execution decisions into memory, information-acquisition, and response components, enabling feedback-driven adaptation of user-specific preferences
  • Bayesian contextual Thompson sampling learns residual preferences relative to a fixed default-and-cost score, with theoretical regret bounds under a linear residual-reward model
  • The method includes feasibility filtering, characterization of unidentifiable preferences under persistent constraints, and anytime-valid false-promotion control
  • Empirical evaluations across personalized reasoning, controlled feedback, and executable tool-use tasks demonstrate improved preference-sensitive behaviors while maintaining competitive end-to-end task performance

Why It Matters

This work addresses a critical bottleneck in deploying LLM agents: the difficulty of personalizing execution decisions without expensive fine-tuning or access to proprietary model internals. By treating personalization as an online learning problem with scalar feedback, FABLE offers a practical, model-agnostic solution that can be layered onto existing black-box agents. This is particularly relevant as the industry moves toward deploying personalized AI assistants at scale.

Technical Details

  • FABLE operates as an external policy layer that wraps a potentially black-box host agent, learning per-user execution policies from scalar feedback observed only for executed actions
  • The framework factorizes agent decisions into three components—memory retrieval, information-acquisition (e.g., asking clarifying questions), and response generation—allowing feedback to selectively update related choices
  • Action exploration is constrained through an externally specified feasible set, and user-specific preferences are learned as residuals against a fixed default-and-cost baseline using Bayesian contextual Thompson sampling
  • Theoretical guarantees include an expected-regret bound against the best feasible action under a linear residual-reward model, along with characterization of preferences that remain unidentifiable under persistent feasibility constraints
  • The method provides anytime-valid false-promotion control to prevent spurious preference learning, and evaluations span personalized-reasoning benchmarks, controlled-feedback scenarios, and executable tool-use tasks

Industry Insight

  • FABLE's model-agnostic design enables personalization of proprietary or closed-source LLM agents without requiring fine-tuning access, lowering the barrier for enterprise deployment of customized AI assistants
  • The factorization of execution decisions into independent components suggests a modular architecture for personalization that could be extended to other agent capabilities beyond memory and tool use
  • The theoretical framework combining regret bounds with false-promotion control offers a principled approach to safe online personalization, which could inform industry standards for user-specific AI behavior adaptation

TL;DR

  • 提出FABLE(Factorized Adaptive Bandit Layer for Execution)框架,通过轻量级外部策略层实现冻结LLM agent的个性化,无需微调主模型
  • 将agent执行决策分解为记忆、信息获取和响应三个维度,通过贝叶斯上下文Thompson采样学习用户特定的残差偏好
  • 在理论层面提供期望遗憾界保证和任何时刻有效的虚假推广控制机制
  • 在个性化推理、受控反馈和可执行工具使用评估中,FABLE在偏好敏感行为上优于纯规则控制,同时保持端到端任务性能竞争力

为什么值得看

该研究解决了LLM agent个性化适应的核心痛点:如何在无法微调或访问内部参数的情况下,从用户反馈中学习个性化执行策略。这对专有系统和黑盒API场景具有直接应用价值,为agent个性化提供了理论保证与工程可行性的平衡方案。

技术解析

  • FABLE架构:作为轻量级策略层部署在可能为黑盒的主agent外部,不修改或微调内部模型,仅通过外部反馈信号学习用户偏好
  • 决策分解机制:将agent的执行决策分解为记忆检索、信息获取(如是否询问澄清问题)和响应生成三个独立维度,使反馈信号能够精准更新相关决策
  • 贝叶斯上下文Thompson采样:在固定默认-成本评分基础上,学习用户特定的残差偏好,通过可行集过滤确保探索动作符合外部约束
  • 理论保证:在线性残差奖励模型假设下,校准变体继承了对最优可行动作的期望遗憾界;同时刻画了持久可行集约束下不可识别的偏好,并提供任何时刻有效的虚假推广控制
  • 评估基准:涵盖个性化推理、受控反馈和可执行工具使用三类任务,验证了方法在偏好敏感行为上的改进

行业启示

  • 为无法微调的专有LLM系统提供了可行的个性化路径,降低了个性化部署的成本和门槛
  • 分解式策略学习框架可推广至其他需要多维度决策优化的agent系统,具有架构复用价值
  • 理论保证(遗憾界、虚假推广控制)为agent个性化在高风险场景(如医疗、金融)的应用提供了可信度基础

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 Fine-tuning 微调 Research 科学研究 Training 训练