ABBEL: Training LLMs to Update Beliefs for Efficient Long-Term Interaction
ABBEL is a framework that improves LLM performance in long-horizon tasks by isolating and supervising the information content of summaries using natural-language belief states. It reduces the performance gap compared to full-context models by about 50% while cutting training steps in half and using significantly less memory.
Cursor
Alibaba/Qwen
BAIR (Berkeley AI)27.07 · 07:02
