Long-term tasks require models to remember hundreds of interaction steps. Keeping full history in context becomes impossible as tasks grow longer. Previous methods used self-summarization to compress this history. However, compressed summaries often fail to match the quality of full context.
ABBEL introduces belief states to replace raw interaction histories. The system updates these summaries based on new information from each step. Researchers use a method called belief grading to supervise the summary content. This process rewards summaries that retain critical information for future decisions.
In collaborative coding tests, ABBEL closed half the performance gap with full context models. The approach required 50% fewer training steps to reach similar results. Memory usage dropped significantly compared to keeping all tokens in context. ABBEL also performed well on a word-guessing game using domain-specific grading.
This method helps agents operate efficiently without sacrificing intelligence. Users can deploy agents that handle long tasks with lower resource costs.
