Optimizing deployed AI systems increasingly means editing prompts, skills, harnesses, and code rather than retraining model weights. Most current methods follow a propose–evaluate–select loop: candidate configurations are tested, and only those passing an acceptance threshold are kept. According to the paper Mara Chain: Rethinking Failure as a Stepping Stone for AI System Auto-Evolution, this habit is costly — discarded candidates often carry information that later proposals need, so optimizers keep rediscovering the same failure modes.
Mara Chain changes the loop. Instead of throwing away a rejected candidate, it retains the candidate along with its rollout traces, structured analyses, and residual failures, then iteratively refines it on the same minibatch. Each refinement chain has a fixed depth, and the candidate pool is bounded using Pareto-filtered Top-N selection, keeping the search tractable.
The authors report gains across three distinct optimization targets. On AppWorld skill optimization, Mara Chain outperforms GEPA, ACE, and SkillOpt-Lite by up to 20.5% relative performance, reaching the target score with 65.5% fewer rollouts than GEPA (3,280 vs 9,514 rollouts; 11.9h vs 35.8h). On TerminalBench 2.1 harness optimization, it reaches a 71.9% pass rate, up 20.2 percentage points over AHE and 22.5 points over Meta-Harness, while using fewer output tokens than the unoptimized baseline. On MuSiQue retrieval-pipeline optimization, it improves test nDCG@10 by 0.104 and Recall@10 by 0.131 over a hand-written pipeline, evaluated on a test split disjoint from the tuning data.
The team says the effect holds across GLM-5, DeepSeek, and Qwen backbones, with average gains of 40.3, 32.7, and 23.4 percentage points respectively — evidence, they argue, that the system itself improves rather than one specific model. The framework is fully white-box and requires no weight updates.
Code is available at github.com/ant-research/AntOmniEvo, and the paper at arXiv:2609.35855. A companion "lobster gym" visualizer renders each candidate as a lobster leveling up, with a full lineage tree making every fix traceable.