Berkeley AI Research introduces ABBEL framework for efficient long-horizon LLM interaction

The ABBEL framework addresses the context scaling limitations of LLMs in long-horizon tasks by replacing full interaction history with natural-language belief states. This belief grading mechanism supervises the content of each belief state, improving performance. In the CollabBench collaborative code generation task, ABBEL reduced the performance gap compared to full-context models by about 50% and halved training steps, while also demonstrating effective belief updating and action selection in the Combination Lock test.