BeliefWeaver
LONG-HORIZON AGENTS · ARXIV 2026

BeliefWeaver

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

Xiyang Wu*1Zhichao Liu*1Shengxin Zhang*2Zongxia Li1Ming Li1Tianyi Zhou3Dinesh Manocha1

1University of Maryland2Google3MBZUAI

* Equal contribution

Read the paper arXiv Code Coming soon Data Coming soon Model Coming soon Watch overview
THE IDEA

A failed attempt should change
what comes next.

Evidence from earlier attempts is retained, connected, and used to choose the next repair. ATTEMPTSPERSISTENT BELIEFNEXT FRONTIER FeedbackEvidenceCheckpoint EVIDENCE-SUPPORTED Retained across rollback
+5.1pp

Teacher final score
vs. native Terminus-2

+9.8pp

Distilled 9B final score
SFT + OPD vs. native Terminus-2

54.8%

Lower teacher API cost
SFT + OPD vs. matched teacher

Paper-reported results across three terminal benchmarks. Metrics and comparison details ↓

01 / OVERVIEW

Recovery that
builds on evidence.

An agent can fail because of a decision made many steps earlier. Trying again is useful only if the next attempt benefits from what the last one revealed.

BeliefWeaver maintains a persistent relational belief state over repair hypotheses, attempts, observations, and checkpoints. New evidence changes the priorities of related hypotheses—even across rollback. An implicit world model reasons about possible repairs and expected progress, while a separate diagnostic gate determines which repairs are ready to execute.

Recurring reasoning roles can then be distilled into a 9B student through supervised fine-tuning and on-policy distillation.

THE PAPER IN A MINUTE

From failure to a better next step.

Download video

Research overview · English synthetic narration · On-screen captions · Diagrams illustrate the method, not a live agent run.

Read the video transcript

The transcript is also available as English captions.

02 / HOW IT WORKS

Remember. Reconsider. Recover.

A belief score ranks a repair.
Diagnostic evidence gates its execution.

PERSISTENT EXPLORATION STATE
01 / PROPOSE & PROBE

Make the failure testable.

The implicit world model proposes repair hypotheses and diagnostic probes, estimates potential progress, and identifies relations between candidate repairs.

Paper Figure 1: the implicit world model proposes repairs; a persistent graph links checkpoints, frontiers, attempts and observations; belief updates rank hypotheses and an evidence gate controls execution.
THE COMPLETE FRAMEWORK Figure 1 from the paper, reproduced under CC BY 4.0. Click to view full size.

03 / MAIN RESULTS

Better recovery, across benchmarks.

Paper · Table 1 ↗

Compare the native agent, the full teacher framework, and the distilled students. Final score measures test-pass fraction or graded reward.

Final score (%, higher is better)

BaselinesFull teacherDistilled students

Aggregate final score across the three benchmark suites.

Values are means ± 95% bootstrap confidence-interval half-widths. All results are from the paper; the visualization does not report new runs.

View the full results, including exploration efficiency
Table 1 · Final score and normalized AUC metrics (×100)
MethodTB2 FinalTB3 FinalLHTB FinalOverall FinalStep-AUCTime-AUC

SC: Structured Continuation. WM: WM Grounding. Step-AUC and Time-AUC summarize progress over step and time budgets; higher is better.

04 / DISTILLING THE REASONING

A smaller model.
A cumulative recovery loop.

Transfer four recurring world-model roles to Qwen3.5-9B with SFT, then refine the student on states it visits through on-policy distillation.

Probe generationEvidence judgmentFrontier relationsTask-clause assessment

The controller, executor, and belief-update rules stay the same. SFT + OPD + Gen also uses a separate student adapter for hypothesis generation. Prospective evaluation remains teacher-based.

Teacher API cost

Estimated $ / task ↓
BeliefWeaver teacher$0.550 ± 0.145
SFT + OPD$0.248 ± 0.052

54.8% lower teacher API cost

SFT + OPD + Gen$0.100 ± 0.023

81.9% lower teacher API cost

Paper-reported savings vs. the matched teacher. Excludes executor, student-serving, and training costs. Accounting rates: $1 / $4 per million input / output tokens. Table 4 ↗

05 / EXPLORE FURTHER

Read it. Run it. Build on it.

01 ↗

The paper

Method, experimental setup, ablations, and limitations.

arXiv:2610.18082
02

Code, data & model

Method configurations, evaluation and training code, distillation data, and the 9B student.

Coming soon
03 ↓

The short video

A narrated visual introduction to the recovery loop and results.

Download MP4
Scope & limitations

Evaluated on terminal benchmarks with verifier feedback and checkpoint restoration. Incorrect world-model predictions can still misrank repairs; transfer to environments with sparse feedback or unreliable rollback remains an open question.

Citation

@misc{wu2026failure,
  title={Failure Recovery via Belief Updates and World Modeling for Long-Horizon Agents},
  author={Xiyang Wu and Zhichao Liu and Shengxin Zhang and Zongxia Li and Ming Li and Tianyi Zhou and Dinesh Manocha},
  year={2026},
  eprint={2610.18082},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2610.18082}
}