ABBEL

ABBEL is a framework that replaces a language model’s full interaction history with compact belief states, isolating and supervising the information retained in each belief. Belief updates replace recursive summarization; beliefs are graded by how well they enable the model to reconstruct the latest observation, penalizing poor retention. In collaborative coding (CollabBench), ABBEL–rec‑BG recovers about 50 % of the performance gap to full‑context models while using only 50 % of the peak tokens and training in half as many steps. In the word‑guessing game Combination Lock, domain‑knowledge belief grading allows ABBEL to approach or surpass full‑context performance with fewer learning steps. Memory peaks are substantially lower than full‑context, and peak‑belief penalties in multi‑objective QA reduce memory usage without significant loss of accuracy. EY SUPPORTS European Cities· Entity summary generated by Graphiti