BELIEFWEAVER improves long-horizon agent recovery from earlier failures
Original title: Failure Recovery via Belief Updates and World Modeling for Long-Horizon Agents
A distilled 9B student gained 9.8 points over the native baseline and cut estimated teacher API cost by 54.8%.
BELIEFWEAVER keeps a relational belief state over frontiers, attempts, observations, and checkpoints, updating priorities from execution evidence while an implicit world model scores repair hypotheses. Across three terminal benchmarks it raised aggregate final score by 5.1 percentage points over native Terminus-2 continuation. Recurring world-model reasoning was distilled into a 9B student with supervised fine-tuning and on-policy distillation.
Unconfirmed: Score and cost figures are the authors' estimates; the three benchmarks are not named.
- Diagnostic evidence gates execution; attempt evidence updates related frontier priorities.
- The framework also uses rollback and checkpoints when recovering from earlier failures.
- The 9B student reduced estimated teacher API cost by 54.8% versus the matched teacher.
- Exploration efficiency showed positive trends on the three benchmarks.
From the original article
1 min read at huggingface.co
Long-horizon agents often struggle to recover from failures whose causes lie several steps earlier, beyond what local reflection or backtracking can resolve. We introduce BELIEFWEAVER, a framework that combines persistent evidence across attempts with prospective reasoning over repair directions. It maintains a relational belief state over frontiers, attempts, observations, and checkpoints.
By Xiyang Wu, Zhichao Liu, Shengxin zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.