Research Papers 论文研究 5h ago Updated 58m ago 更新于 58分钟前 47

GRACE: Graph-Grounded Reflective Agent Copilot Engine for Expert-in-the-Loop Knowledge Expansion GRACE:面向专家在环知识扩展的图基座反思型智能体协同引擎

GRACE is a framework that deconstructs LLM responses into atomic claims and grounds them against trusted knowledge priors using a weighted bipartite graph structure Claims are classified as Grounded, Refuted, or Boundary through weighted centrality analysis, enabling detection of both hallucinations and novel/contested claims at the knowledge frontier A Return on Attention (RoA) objective function optimizes expert resource allocation by deferring claims to human review only when priority-weighte GRACE框架将LLM响应分解为原子声明,在加权二分图中与可信知识先验进行grounding,解决标准RAG无法追踪跨文档证据关系和量化不确定性的问题 通过加权中心性分析将声明分类为Grounded(有依据)、Refuted(被反驳)或Boundary(边界/新颖),不仅识别幻觉还能发现知识前沿的新颖或争议性声明 引入Return on Attention (RoA)目标函数,仅在优先级加权不确定性超过验证成本时才将声明提交专家审查,实现人机资源的高效分配 专家验证的声明被提升为新的证据锚点,形成validator-LLM进化循环,持续扩展知识库 在多个语言模型和通用/领域数据集上的评估显示,

62
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • GRACE is a framework that deconstructs LLM responses into atomic claims and grounds them against trusted knowledge priors using a weighted bipartite graph structure
  • Claims are classified as Grounded, Refuted, or Boundary through weighted centrality analysis, enabling detection of both hallucinations and novel/contested claims at the knowledge frontier
  • A Return on Attention (RoA) objective function optimizes expert resource allocation by deferring claims to human review only when priority-weighted uncertainty exceeds verification cost
  • Verified claims are promoted as new evidence anchors, creating an evolutionary validator-LLM loop that iteratively expands the knowledge base
  • GRACE outperforms standard RAG baselines across multiple language models and datasets spanning general and domain-specific knowledge

Why It Matters

This work addresses a critical limitation of current LLM deployment: standard RAG pipelines retrieve isolated passages without tracking cross-document evidence relationships or quantifying uncertainty, leaving hallucinations inadequately addressed. By shifting hallucination mitigation from the generation level to the system level through graph-structured representations and expert-in-the-loop verification, GRACE offers a scalable pathway for deploying LLMs in high-stakes domains where accuracy and traceability are non-negotiable.

Technical Details

  • Graph-Grounded Claim Decomposition: LLM outputs are decomposed into atomic claims and mapped onto a weighted bipartite graph connecting claims to trusted knowledge priors, with edge weights encoding the closeness of each claim to its corresponding prior
  • Weighted Centrality Classification: Using graph centrality analysis on the weighted bipartite structure, claims are classified into three categories—Grounded (well-supported), Refuted (contradicted), or Boundary (novel/contested at the knowledge frontier)—enabling nuanced uncertainty quantification beyond binary verification
  • Return on Attention (RoA) Objective: A formal optimization framework that determines whether a claim should be deferred to expert review based on the condition that priority-weighted uncertainty exceeds the cost of verification, ensuring efficient allocation of human or agent resources
  • Evolutionary Knowledge Expansion Loop: Expert-verified claims are promoted to new evidence anchors within the knowledge base, closing a validator-LLM feedback loop that iteratively expands and refines the trusted knowledge corpus across deployment cycles
  • Empirical Evaluation: Tested across multiple language models on datasets spanning both general knowledge and domain-specific applications, demonstrating superior retrieval performance compared to RAG baselines and effective selection of boundary knowledge for expert verification

Industry Insight

  • The system-level approach to hallucination mitigation—rather than generation-level fixes—represents a paradigm shift that organizations should adopt for high-stakes applications in healthcare, legal, and financial domains where traceable, verifiable outputs are essential
  • The RoA framework provides a practical cost-benefit model for human-in-the-loop deployment, enabling organizations to optimize the trade-off between verification thoroughness and operational cost, which is critical for scaling expert-assisted AI systems
  • The evolutionary knowledge expansion loop suggests that organizations can build compounding value from expert verification efforts over time, turning one-time review costs into permanent knowledge assets that improve system performance iteratively

TL;DR

  • GRACE框架将LLM响应分解为原子声明,在加权二分图中与可信知识先验进行grounding,解决标准RAG无法追踪跨文档证据关系和量化不确定性的问题
  • 通过加权中心性分析将声明分类为Grounded(有依据)、Refuted(被反驳)或Boundary(边界/新颖),不仅识别幻觉还能发现知识前沿的新颖或争议性声明
  • 引入Return on Attention (RoA)目标函数,仅在优先级加权不确定性超过验证成本时才将声明提交专家审查,实现人机资源的高效分配
  • 专家验证的声明被提升为新的证据锚点,形成validator-LLM进化循环,持续扩展知识库
  • 在多个语言模型和通用/领域数据集上的评估显示,GRACE的知识库检索性能优于RAG基线,RoA框架能有效选择有价值的边界知识

为什么值得看

GRACE框架代表了从生成层面到系统层面的幻觉缓解新思路,通过图结构表示和专家-in-the-loop验证的协同作用,为高风险场景下的LLM部署提供了更可靠的解决方案。其RoA目标函数为人机协作的资源分配提供了量化决策框架。

技术解析

  • 核心架构:GRACE采用加权二分图结构,将LLM生成的原子声明与可信知识先验进行匹配,边权重编码声明与先验的接近程度,支持跨文档证据关系追踪
  • 声明分类机制:基于加权中心性分析,将声明分为Grounded、Refuted和Boundary三类,Boundary类别特别用于识别模型知识前沿的新颖或争议性声明
  • RoA目标函数:Return on Attention机制通过比较优先级加权不确定性与验证成本,智能决策是否将声明提交专家审查,优化人工/agent资源分配
  • 进化循环设计:专家验证后的声明被提升为新的证据锚点,形成validator-LLM持续学习循环,实现知识库的迭代扩展

行业启示

  • 幻觉缓解应从生成层面转向系统层面,图结构表示结合专家验证可在系统级而非单点生成级解决问题
  • 引入量化决策框架(如RoA)为人机协作提供可操作的资源分配策略,平衡验证成本与知识价值
  • 构建可持续的知识进化循环,通过专家反馈持续扩展可信知识库,形成LLM与人类专家的协同增强机制

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。

RAG 检索增强生成 Agent Agent LLM 大模型 Research 科学研究