Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction

BAIR · 1 month ago · 11

Researchers developed ABBEL, a framework that teaches language models to maintain and update natural-language belief states instead of storing full interaction history, improving performance on long-horizon tasks like collaborative coding. On CollabBench, ABBEL with reconstruction-based belief grading reduced the performance gap versus full-context models by 50% while requiring 50% fewer training steps and using 6,000 context tokens versus 14,000. This enables LLMs to handle extended interactions more efficiently while preserving task performance where data for training summarization is limited.