From Context-Aware to Conflict-Aware: Generalizing Contrastive Decoding for Knowledge Conflict in LLMs

2026-06-08

AI Summary

This work addresses the reliability limitations of large language models when generating text based on external context, particularly when such context conflicts with the model’s internal priors. To overcome the shortcomings of conventional approaches that unilaterally prioritize external context, the authors propose a conflict-aware decoding paradigm featuring Adaptive Router Routing (ARR), which dynamically balances the authority of contextual information against internal knowledge. The study further uncovers the power-family structure underlying affine logit combinations in existing contrastive decoding methods and their inherent mechanism asymmetry, leveraging these insights to design a conflict signal detector and an adaptive routing strategy. Evaluated on the newly introduced TriState-Bench protocol, the proposed method significantly enhances robustness to erroneous context—raising exact match (EM) scores from below 6 to 16–33—while preserving performance in correction and consistency tasks.

Abstract

When large language models generate from retrieved or augmented contexts, conflicts between external context and parametric priors remain a central reliability bottleneck. Existing contrastive decoding methods follow a \emph{context-aware} paradigm that unilaterally amplifies context over parametric priors, overwriting correct priors when the context is erroneous. We generalize this to the \textbf{conflict-aware} paradigm that dynamically allocates authority between prior and context based on conflict signals, rather than presupposing context trustworthiness. We show that the affine combination of prior and context logits yields a \textbf{power family} with an inherent \textbf{regime asymmetry}: extrapolation amplifies errors unboundedly when the prior is correct, interpolation under-corrects when the context is correct, and no static regime covers both. Existing contrastive decoding methods are instances of this family, mostly extrapolative. To evaluate both conflict directions, we propose TriState-Bench, a model-aware evaluation protocol that calibrates per-model prior knowledge to measure three conflict states: correction, resistance, and agreement. To resolve the asymmetry, we propose Adaptive Regime Routing (ARR), which routes between regimes at each step, lifting resistance EM from below 6 to 16--33 without sacrificing correction or agreement. Our code is available at https://github.com/keith-Jiang/conflict-aware-decoding.

Problem

• knowledge conflict
• contrastive decoding
• context-aware
• large language models
• reliability bottleneck

Innovation

• conflict-aware decoding
• contrastive decoding
• adaptive regime routing
• knowledge conflict
• TriState-Bench