Beyond Memory: PoS Maintains Explicit Belief States to Rescue Long-Horizon LLM Agents from Belief Trapping

核心背景与行业痛点

As LLM agents tackle extended multi-turn tasks across system diagnostics and software engineering, managing long-horizon context remains a brittle bottleneck. Prevailing frameworks treat context as raw interaction history or lossy semantic compression, neither of which guarantees a faithful mental model of changing world dynamics. Agents frequently fall into 'Belief Trapping'—a chronic failure mode where the model issues successive tool calls grounded in obsolete or disproven assumptions, thrashing in unproductive loops without driving true goal progression.

架构亮点与底层机制

Researchers from Tsinghua University and Nankai University present PoS (Possibility-of-State, arXiv:2610.01415), establishing an explicit belief framework at inference time:

1. Continual Explicit Belief State Maintenance: Moves beyond message logs by synthesizing structured belief states. Each belief pairs an up-to-date estimate of world state with unresolved task requirements, explicitly highlighting remaining information gaps.
2. Active Consistency Validation: Automatically cross-examines tool return signals against held beliefs, extinguishing hallucinated contradictions before they taint downstream reasoning.
3. Real-Time Belief Trapping Detection: Monitors trajectory velocity to flag periods where repetitive tool actions yield zero measurable progress toward target criteria.
4. Tailored Recovery Protocols: Deploys surgical recovery routines adapted to specific trapping failure typologies and unresolved requirement categories, rescuing stuck agents instantly.

权威 Benchmark 与实测跑分对比

Benchmarked across four challenging execution and system diagnosis benchmarks using three distinct foundation model backbones:

1. Top Overall Performance Across All Benchmarks: PoS captures the highest task success rate across every benchmark and across all three evaluated model families.
2. Essential Role of Consistency and Recovery: Ablation studies confirm that removing consistency validation or recovery modules triggers a 38%+ drop in success rates past 30 execution steps.
3. Robust Context-Scaling Resilience: Demonstrates near-flat performance stability across extended multi-turn horizons, completely avoiding the catastrophic degradation plaguing baseline agents.

开发者实战落地与开箱指南

The PoS inference engine is publicly available. Teams architecting long-running AIOps agents, site reliability copilots, or autonomous engineering bots can integrate PoS as an inference-time context middleware, replacing passive chat history with active belief-state management to guarantee goal alignment.