Mara Chain — learn from rejected agent configurations

TL;DR

Mara Chain retains rejected prompt, skill, harness and code candidates and refines them using evidence from earlier attempts. The authors report better task performance with fewer rollouts across skill, harness and retrieval-pipeline optimization. (source)

Authors & Org

Yubin Lyu, Fu Li, Jiawei Fei, Yang Zhao, Weixing Mei and Yinan Wu. First submitted 2026-09-25. Affiliations are not stated in the captured abstract page. (source)

Method

Instead of discarding a candidate that fails the acceptance criterion, the procedure carries its failure evidence into further revisions. Each refinement chain has a fixed maximum depth; Pareto-filtered Top-N selection bounds the candidate pool. The optimized objects are system artifacts rather than model weights. (source)

Results

• On AppWorld, the authors report up to 20.5% relative performance improvement over GEPA, ACE and SkillOpt-Lite, reaching the target score with 65.5% fewer rollouts than GEPA. The relative improvement is not a percentage-point gain. (source)
• On TerminalBench 2.1, pass rate improves by 20.2 percentage points over AHE and 22.5 percentage points over Meta-Harness. (source)
• On MuSiQue, test nDCG@10 and Recall@10 improve by 0.104 and 0.131, respectively, over a hand-written retrieval pipeline. (source)

Significance

For Agents (LLM Agents), failure records can be inputs to optimization rather than only reasons to reject an artifact. Interpretation: this complements persistent feedback in Opera — persistent feedback for coding agents, but Mara Chain changes candidate configurations while Opera tracks corrections during an agent’s work. (source) (Opera)

Open Questions

The abstract does not give absolute pass rates, rollout lengths, token costs or detailed selection settings. Fewer rollouts do not by themselves establish a proportional reduction in total compute. Full paper and implementation were not reviewed. (source)

Cite

arXiv:2609.35855 · (source)