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

This article introduces a novel approach to contrastive decoding in large language models (LLMs) by shifting from a context-aware paradigm to a conflict-aware paradigm that dynamically balances the authority of parametric priors and external context based on conflict signals. The authors present a new evaluation protocol, TriState-Bench, to assess the effectiveness of their method across three conflict states, and they propose Adaptive Regime Routing (ARR), which significantly improves error resistance from below 6 to a range of 16-33 while maintaining correction and agreement. This advancement is crucial for practitioners as it enhances the reliability of LLM outputs in scenarios where external context may conflict with learned knowledge, thereby improving model robustness in real-world applications.