MACR: Explicit Knowledge Conflict Resolution for LLMs

Huang Peng and four coauthors submitted a paper on 18 Jun 2026 proposing MACR, a framework that explicitly resolves conflicts between a large language model's internal parametric knowledge and external contextual information. The arXiv entry for the paper is arXiv:2606.20245 and the submission lists the work as 12 pages with 3 figures.

What is MACR?

MACR is a knowledge conflict resolution framework for LLM inference that combines confidence estimation and multi-agent reasoning, designed to handle unreliable internal and external knowledge simultaneously. The system first assesses an LLM's confidence using a modified semantic entropy measure, then either externalizes the model's internal knowledge as text or retrieves external knowledge to form basic contexts, and finally applies an inductive multi-agent reasoning pipeline to resolve inconsistencies.

MACR departs from approaches that assume one source is reliable and the other is not, instead moving beyond that binary choice by explicitly inducing rules, analyzing conflicts, and producing reconciled answers across all available contexts.

How does MACR work?

MACR operates in two stages: adaptive knowledge assessment and an inductive multi-agent reasoning framework with three specialized agents. The first stage uses a modified semantic entropy measure to quantify an LLM's confidence; when confidence is low the system retrieves external knowledge, when confidence is adequate it externalizes internal knowledge as textual context to use in downstream reasoning.

The second stage runs three specialized agents in an inductive multi-agent reasoning loop, where one agent induces explicit rules, a second analyzes potential conflicts among contexts, and a third resolves inconsistencies across all available contexts. The paper describes this flow as producing basic contexts for subsequent reasoning and delivering interpretable resolutions rather than privileging one source by default.

How does MACR compare to prior approaches?

Prior methods typically assume either the model or the provided context is reliable and avoid conflicts by privileging one source over the other. MACR explicitly challenges that assumption by quantifying confidence with semantic entropy and by orchestrating a dedicated three-agent resolution pipeline that reasons about and reconciles multiple sources simultaneously. The authors state that "MACR significantly outperforms state-of-the-art baselines across benchmarks," and they present the design and evaluation across the paper's datasets and figures.

Why it matters

LLM deployments increasingly mix parametric knowledge with user-supplied context, and errors can come from both sources. MACR addresses a practical gap by giving models a mechanism to assess when their internal knowledge is insufficient and to reconcile contradictory inputs rather than blindly trusting one side. That change could reduce factual errors in question answering and chain-of-thought scenarios where conflicting evidence appears, and it provides interpretable conflict resolutions that are useful for auditing and downstream validation.

What to watch

The arXiv page lists sections for code, data, and demos and toggles for platforms such as Hugging Face, Replicate, and DagsHub, so look for associated code releases or demos that reproduce the paper's claims. The authors submitted the paper on 18 Jun 2026 as arXiv:2606.20245; the next concrete signals will be any released code, benchmark files, or replication notes tied to the figures in the 12-page manuscript.