From Context-Aware to Conflict-Aware: Generalizing Contrastive Decoding for Knowledge Conflict in LLMs
Researchers have introduced a new framework called conflict-aware decoding to address knowledge conflicts in large language models. This method dynamically balances information from external context and the model's internal knowledge, unlike previous context-aware approaches that prioritized external information. The proposed technique, Adaptive Regime Routing (ARR), aims to resolve an inherent asymmetry in decoding regimes, improving the model's ability to handle disagreements between context and prior knowledge. AI
IMPACT Introduces a novel method to improve LLM reliability by better handling conflicting information sources.