ABBEL: Learning Natural Language Belief States for Efficient Memory Interaction

Large Language Models (LLMs) struggle to manage long and complex interactions due to context window limitations. The common heuristic method of recursive summarization (context compaction) significantly degrades performance, especially in human assistance domains where high-quality data is scarce, such as collaborative code generation.

ABBEL (Acting through Belief Bottlenecks) proposes a solution by isolating and supervising the informative content of summaries as natural language belief states. Inspired by recursive Bayesian estimation, ABBEL periodically updates these belief states. The key concept is "belief grading," which evaluates the quality of these belief states by measuring their ability to reconstruct recent observations, thereby reducing the performance gap with full-context models and accelerating training. This approach enables more efficient memory management and improved interpretability.