SERA trains recursive agents to score their own subtasks
Original title: Self-Evaluating Recursive Agents
On two text benchmarks, SERA beat recursive-agent baselines by 5.38 and 13.14 points on average.
The paper presents SERA, a training method for recursive language-model agents that decompose tasks and delegate them to child copies of the same policy. Before delegation, the parent writes a weighted rubric for each subtask; a ranking objective against verified outcomes trains that rubric, and a leaf-coverage signal credits decomposition. The authors say learning to judge subtasks improves solving because the same weights evaluate and execute, and that the learned rubric can also select branches during inference-time tree search.
Unconfirmed: Benchmark gains are reported by the paper and are not independently confirmed in this item.
- Gains are reported on TextCraft-Synth and TextWorld-Sync.
- Rubric-guided tree search added 2.43 points on TextWorld-Sync.
- The judge is used to train rubric generation, not to score every solve step.
- Authors say the agent's own rubric scores beat direct use of the judge.
From the original article
1 min read at huggingface.co
Recursive language-model agents decompose tasks and delegate subtasks to child instances of the same policy, forming a tree of work. Training them, however, is hard: the final outcome is verifiable, but the self-invented intermediate subtasks are numerous and carry no ground truth. Existing methods score each node with a verifier or judge, which is costly at scale and blind to decomposition quality.
By TianYi Lyu, Xiaozhe Li, Yang Li · 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.