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

Haoming Xu, Weihong Xu, Zongrui Li, Mengru Wang, Yunzhi Yao, Chiyu Wu, Jin Shang, Yu Gong, Shumin Deng

arXiv:2605.30219 · 2026-05-29 공개 · arXiv · PDF

reinforcement-learning llm llm-evaluation representation-steering belief-state symbolic-verifiers contextual-belief-management belieftrack

Abstract

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 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.

한국어 요약

📋 한 줄 요약

**[LLM Belief Tracking / Long-Horizon]** BeliefTrack가 닫힌 세계에서 Failed Stay·Update·Isolation 3 실패 진단, vanilla LLM은 심각한 CBM 실패; belief-state reward RL이 평균 70.9% 실패 감소, representation steering이 두 task 평균 46.1% 감소.

🎯 핵심 기여도

💡 핵심 아이디어

Long-horizon 추론의 신뢰성은 belief state의 명시 관리(언제 stay·update·isolate)에 달려 있으며, closed-world의 symbolic verifier를 통해 정확히 측정 가능하고, prompting은 한계가 있는 반면 belief-state reward RL과 representation steering은 정량적으로 실패를 줄인다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Long-horizon LLM의 누락된 평가 축(belief management)을 형식화, closed-world의 symbolic verification으로 정밀 진단 가능, RL·representation steering의 효과를 정량 입증. **한계**: Closed-world·symbolic 평가가 실제 open-world 대화와 차이, 70.9% 감소도 잔여 실패 존재, RL·steering의 학습 비용·일반화는 추가 검증.

🚀 실용적 활용