Mara Chain: Rethinking Failure as a Stepping Stone for AI System Auto-Evolution
Mara Chain refines rejected AI candidates, beating GEPA on AppWorld with 65.5% fewer rollouts.
Mara Chain retains rejected prompt, skill, harness, and pipeline candidates and iteratively refines them using evidence from earlier attempts, with fixed chain depth and Pareto-filtered Top-N selection. On AppWorld it outperforms GEPA, ACE, and SkillOpt-Lite by up to 20.5% relative performance and hits the target with 65.5% fewer rollouts than GEPA. It also lifts TerminalBench 2.1 pass rates by 20.2 and 22.5 points over AHE and Meta-Harness, and improves MuSiQue nDCG@10 and Recall@10 by 0.104 and 0.131 versus a hand-written pipeline.
- Rejected candidates are refined rather than discarded.
- Fixed depth and Pareto Top-N selection bound the pool.
- AppWorld target is reached with 65.5% fewer rollouts than GEPA.
- TerminalBench 2.1 pass rate rises by up to 22.5 points.
From the original article
1 min read at huggingface.co
Optimizing deployed AI systems increasingly amounts to editing prompts, skills, harnesses, and code rather than model weights. Existing approaches commonly optimize these artifacts through propose-evaluate-select procedures, where candidate configurations are evaluated and only those meeting an acceptance criterion are selected.
By Yubin Lyu, Fu Li, Jiawei Fei · 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.