When Should Models Change Their Minds? Contextual Belief Management in Large Language Models

Researchers introduced Contextual Belief Management and the BeliefTrack benchmark to measure how language models update, preserve, and filter beliefs over long contexts.

Using exact turn-level evaluation, the study identified three common failure modes in LLMs during belief tracking and context management.

The results show that belief-reward reinforcement learning and representation-level steering can substantially reduce these errors.

The work provides a concrete way to test long-horizon reasoning and improve reliability in evolving dialogue settings.