Mara Chain: Rethinking Failure as a Stepping Stone for AI System Auto-Evolution
Yubin Lyu, Fu Li, Jiawei Fei, Yang Zhao, Weixing Mei, Yinan Wu
Ant Group · Beijing Intelligent Game and Decision Lab · Beijing Defense Innovation Institute
Published September 25, 2026 · on the feed October 9, 2026 · ▲ 8 · arXiv 2609.35855
The abstract
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. Yet our analysis shows that discarded candidates often contain information critical for subsequent optimization. Discarding them causes later proposals to revisit the same failure modes. We introduce Mara Chain, a refinement procedure that turns rejected candidates into stepping stones. Rather than discarding a rejected candidate, Mara Chain retains and iteratively refines it using evidence accumulated across preceding attempts. The procedure limits each refinement chain to a fixed depth and applies Pareto-filtered Top-N selection to bound the candidate pool. Across AppWorld skill optimization, TerminalBench 2.1 harness optimization, and MuSiQue retrieval-pipeline optimization, Mara Chain delivers greater task-performance gains with fewer rollouts. It outperforms GEPA, ACE, and SkillOpt-Lite by up to 20.5% in relative performance on AppWorld, reaching the target score with 65.5% fewer rollouts than GEPA. It improves the pass rate by 20.2 and 22.5 percentage points over AHE and Meta-Harness on TerminalBench 2.1, respectively, and improves MuSiQue test nDCG@10 and Recall@10 by 0.104 and 0.131 over a hand-written retrieval pipeline.
The Brief
2:19A short podcast summary of the paper. Written and voiced by AI.
Transcript
Mara Chain. Rethinking Failure as a Stepping Stone for AI System Auto-Evolution. By Yubin Lyu and five others, from Ant Group, the Beijing Intelligent Game and Decision Lab, and one other institution. Published on the twenty-fifth of September. Its code repository is worked out: about two hundred source files, with a readme and a licence.
Improving a deployed AI system now often means editing prompts, skills, harnesses and code, rather than touching model weights. The usual way to optimise those artifacts is a cycle of propose, evaluate and select. Candidate configurations are scored, and only the ones that meet an acceptance criterion survive. The authors' analysis finds a cost in that cycle. Discarded candidates often hold information that matters for later optimisation, and throwing them away leads later proposals to walk back into the same failure modes.
Mara Chain is a refinement procedure that treats rejected candidates as stepping stones instead of waste. When a candidate fails the acceptance test, it is kept, and then refined again and again using the evidence gathered over the attempts that came before it. A failure becomes the starting point of the next try rather than a dead end.
Two limits keep this from growing without bound. Each refinement chain is capped at a fixed depth. And the pool of candidates is bounded by Top-N selection filtered on the Pareto front, so only a limited set of the best trade-offs is carried forward.
The method is tested on three kinds of optimisation: skills on AppWorld, a harness on TerminalBench 2.1, and a retrieval pipeline on MuSiQue. Across them, Mara Chain delivers larger gains in task performance while using fewer rollouts. On AppWorld it beats GEPA, ACE and SkillOpt-Lite by up to twenty point five percent in relative performance, and reaches the target score with sixty-five point five percent fewer rollouts than GEPA.
On TerminalBench 2.1, the pass rate rises by twenty point two percentage points over AHE, and by twenty-two point five points over Meta-Harness. On MuSiQue, compared with a retrieval pipeline written by hand, test nDCG@10 improves by zero point one zero four and Recall@10 by zero point one three one.
The reframing is in the subtitle: failure as a stepping stone. A rejected configuration still records what was tried and why it fell short, and keeping that record in play means the search stops repeating its own mistakes.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free