Entropic Claim Resolution: Uncertainty-Driven Evidence Selection for RAG

arXiv cs.AI / 3/31/2026

💬 OpinionDeveloper Stack & InfrastructureIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper argues that conventional relevance-based dense retrieval in RAG can fail when queries are ambiguous or evidence conflicts, because relevance alone doesn’t resolve epistemic uncertainty.
It proposes Entropic Claim Resolution (ECR), an inference-time algorithm that selects evidence by minimizing entropy over competing semantic answer hypotheses rather than simply maximizing query-document similarity.
ECR sequentially chooses atomic evidence using a decision-theoretic Expected Entropy Reduction (EER) criterion and stops when the system satisfies a defined epistemic sufficiency condition (H ≤ epsilon) under coherence constraints.
The authors integrate ECR into a production-grade multi-strategy retrieval pipeline (CSGR++) and provide theoretical analysis of the approach’s properties.
The work reframes RAG from retrieving “most relevant” documents to retrieving “most discriminative” evidence that reduces uncertainty in the final answer.

Abstract

Current Retrieval-Augmented Generation (RAG) systems predominantly rely on relevance-based dense retrieval, sequentially fetching documents to maximize semantic similarity with the query. However, in knowledge-intensive and real-world scenarios characterized by conflicting evidence or fundamental query ambiguity, relevance alone is insufficient for resolving epistemic uncertainty. We introduce Entropic Claim Resolution (ECR), a novel inference-time algorithm that reframes RAG reasoning as entropy minimization over competing semantic answer hypotheses. Unlike action-driven agentic frameworks (e.g., ReAct) or fixed-pipeline RAG architectures, ECR sequentially selects atomic evidence claims by maximizing Expected Entropy Reduction (EER), a decision-theoretic criterion for the value of information. The process dynamically terminates when the system reaches a mathematically defined state of epistemic sufficiency (H <= epsilon, subject to epistemic coherence). We integrate ECR into a production-grade multi-strategy retrieval pipeline (CSGR++) and analyze its theoretical properties. Our framework provides a rigorous foundation for uncertainty-aware evidence selection, shifting the paradigm from retrieving what is most relevant to retrieving what is most discriminative.