BeliefWeaver Adds 5.1 pp to Agents on Long Tasks; 9B Student Adds 9.8

An agent that works in a terminal for an hour often doesn't break on the last step. The cause sits somewhere in the middle of the chain: a wrongly chosen dependency, a skipped check, a corrupted config. Local reflection and a one-step rollback don't find that kind of error. The authors of a paper from Hugging Face Daily Papers propose a framework called BeliefWeaver. On three terminal benchmarks, it raises the aggregate final score by 5.1 percentage points over native Terminus-2 continuation. A version distilled into a 9B model gained +9.8 pp over the reported native baseline and cut the estimated cost of teacher API calls by 54.8%.

Memory of failures instead of restarting from scratch

The key idea is that the experience of failed attempts does not vanish between runs. BeliefWeaver maintains a relational belief state that links four types of entities: frontiers (directions the agent can explore next), attempts, observations and checkpoints. When an attempt reveals something, the framework updates the priorities not only of the current frontier but also of related ones, both across attempts and during rollback. Diagnostic data works as a filter here: execution is gated on it, so the agent does not move forward blind.

The second component is an implicit world model (IWM). It scores candidate fixes, predicts whether a chosen direction will lead to progress, and identifies links between frontiers. The agent gets more than a list of "what was tried"; it gets a forecast of "where it makes sense to go." According to the authors, this also improves exploration efficiency: the trends on this metric are positive, although the abstract gives no figures.

Distillation into 9B: where the savings come from

A world model built on a large LLM is expensive, because every hypothesis evaluation is an API call. So the authors isolated the recurring reasoning components of the IWM and transferred them into a student with 9 billion parameters. They used supervised fine-tuning and on-policy distillation (OPD). OPD typically means training on the student's own trajectories, graded by the teacher, rather than only on ready-made examples. This narrows the gap between training and real deployment.

• BeliefWeaver's gain over native Terminus-2 continuation: +5.1 pp in aggregate final score.
• The 9B student's gain over the reported native baseline: +9.8 pp.
• Reduction in estimated teacher API cost relative to the matched teacher: 54.8%.
• Evaluation on three terminal benchmarks.
• Training components: SFT and OPD.

The 5.1 and 9.8 figures cannot be compared directly. The first is measured against native Terminus-2 continuation, the second against a reported native baseline. The 54.8% savings are described as an estimate. The abstract does not give the benchmark names, the student's base model or the absolute scores. Check these in the full paper before drawing conclusions about reproducibility.

What teams building agents can take away

Even without BeliefWeaver itself, the work offers several engineering techniques. First, store failure evidence in a structured form and across attempts, not only in the context of the current run. Second, save checkpoints and attach observations to them, so a rollback can return to the real cause rather than to the last step. Third, put a diagnostic gate in front of expensive or irreversible actions. Finally, if the cost of calling a strong model in the loop is critical, it makes sense to test distilling narrow reasoning components into a small model. A 9.8 pp result with almost half the teacher cost suggests the approach works, but you need to confirm it on your own tasks.