When Should Models Change Their Minds? Contextual Belief Management in Large Language Models
For researchers working on long-horizon LLM interactions, this paper identifies and measures a critical failure mode (CBM) and provides two effective mitigation strategies.
The paper introduces Contextual Belief Management (CBM) as a challenge for LLMs in long-horizon interactions, and proposes BeliefTrack benchmark to measure it. Using reinforcement learning with belief-state rewards, they reduce failure rates by 70.9% on average, and representation-level steering reduces failures by 46.1%.
Long-horizon interactions require language models to manage accumulating information: when to update their state, when to preserve their state, and what to ignore. We study this challenge as \textbf{Contextual Belief Management (CBM)}: maintaining a predicted belief state aligned with formal evidence while isolating task-irrelevant noise. To make CBM measurable, we introduce BeliefTrack, a closed-world benchmark spanning Rule Discovery and Circuit Diagnosis, where a finite belief space and symbolic verifiers enable exact turn-level evaluation. BeliefTrack diagnoses three failures: Failed Stay, Failed Update, and Failed Isolation. Across multiple LLMs, vanilla models exhibit severe CBM failures, while explicit belief-tracking prompts provide limited gains. In contrast, reinforcement learning with belief-state rewards reduces failure rates by 70.9\% on average. Further probing reveals latent belief-state dynamics behind these failures, and representation-level steering reduces failure rates by 46.1\% across two tasks\footnote{Code is coming soon at https://github.com/zjunlp/CBM.