Failure Recovery via Belief Updates and World Modeling for Long-Horizon Agents

Failure Recovery via Belief Updates and World Modeling for Long-Horizon Agents. By Xiyang Wu and six others, from the University of Maryland, Google, and one other institution. Published on the thirteenth of October. No code repository is linked to it.

A long-horizon agent fails, and the cause of that failure lies several steps back. Local reflection looks at the step that just went wrong. Backtracking undoes a little of the recent work. Neither reaches far enough when the real mistake is buried earlier in the run, and agents often struggle to recover in exactly these cases. The symptom shows up late, while the cause sits out of view.

BELIEFWEAVER brings two things together to close that distance. One is evidence that persists across attempts, instead of being thrown away after each try. The other is prospective reasoning: thinking ahead about which repair direction is likely to pay off before committing to it. Memory looks backward over what has been tried, and the forward reasoning looks at what to try next.

At its centre is a relational belief state over frontiers, attempts, observations and checkpoints. Two kinds of evidence feed it. Execution evidence updates the priorities of related frontiers, across attempts and across rollback, so what one attempt learns is not lost when the agent rewinds. Diagnostic evidence gates execution.

Alongside the belief state sits an implicit world model. It evaluates repair hypotheses, predicts progress, and identifies relations among frontiers to steer exploration. Because these world model reasoning components recur from task to task, the authors distill them into a nine billion parameter student, using supervised fine-tuning followed by on-policy distillation.

Evaluation spans three terminal-based benchmarks. Against native Terminus-2 continuation, BELIEFWEAVER raises the aggregate final score by five point one percentage points, with positive trends in exploration efficiency. The distilled student does better still against its own reference point: it improves final score by nine point eight percentage points over the reported native baseline, and cuts estimated teacher API cost by fifty-four point eight percent compared with the matched teacher.

Two claims come out of this, and they support each other. Remembering what earlier attempts revealed, and reasoning forward about repairs, helps an agent recover from mistakes it made long ago. And the reasoning that does this recurs often enough that a smaller student can learn it, which loosens the dependence on an expensive teacher at inference time.