Mara Chain Recycles Failures, Cuts GEPA Rollouts by 65.5%

A method called Mara Chain reaches GEPA's target score on AppWorld skill optimization with 65.5% fewer rollouts by refusing to throw its rejected candidates away. The paper, Mara Chain: Rethinking Failure as a Stepping Stone for AI System Auto-Evolution, argues that the standard propose-evaluate-select loop silently wastes the information inside its own failures.

The authors frame the work against the broader shift away from weight updates: "Optimizing deployed AI systems increasingly amounts to editing prompts, skills, harnesses, and code rather than model weights." In that regime, their analysis finds, "discarded candidates often contain information critical for subsequent optimization. Discarding them causes later proposals to revisit the same failure modes."

Mara Chain's response is to keep each rejected candidate in a chain, refine it using evidence from prior attempts, and bound the candidate pool with Pareto-filtered Top-N selection and a fixed chain depth. The reported wins span three benchmarks: up to 20.5% relative gain over GEPA, ACE and SkillOpt-Lite on AppWorld skill optimization; pass-rate improvements of 20.2 and 22.5 percentage points over AHE and Meta-Harness on TerminalBench 2.1; and nDCG@10 and Recall@10 gains of 0.104 and 0.131 over a hand-written MuSiQue retrieval pipeline.

The results come from a single preprint on three chosen benchmarks. The abstract names no base model, publishes no number for the fixed depth of a refinement chain, and does not report the storage or compute cost of keeping the failure pool around.