Mara Chain: Rethinking Failure as a Stepping Stone for AI System Auto-Evolution

Excited to share Mara Chain — rethinking failure as a stepping stone for AI system auto-evolution! 🧬

Standard propose–evaluate–select optimization discards every candidate that fails the acceptance bar — so later proposals keep revisiting the same failure modes. Mara Chain instead retains each rejected candidate together with its rollout traces, structured analyses, and residual failures, and iteratively refines it on the same minibatch. Failure becomes a stepping stone, not a dead end. Chain depth is fixed and the pool is bounded by Pareto-filtered Top-N selection.

📊 Highlights (same model, same budget):
• AppWorld (skill optimization): up to +20.5% relative over the best prior optimizer; reaches the target score with 65.5% fewer rollouts than GEPA (3,280 vs 9,514 rollouts; 11.9h vs 35.8h); best final score on all four metrics — 89.9/83.9/76.7/59.0 vs 48.2/35.7/33.3/19.4 unoptimized
• TerminalBench 2.1 (harness optimization): 71.9% pass rate, +20.2pp over AHE and +22.5pp over Meta-Harness — with fewer output tokens than the unoptimized baseline
• MuSiQue (retrieval-pipeline optimization): test nDCG@10 +0.104 / Recall@10 +0.131 over a hand-written pipeline, on a test set disjoint from the tuning split — real generalization, not memorization
• Verified on GLM-5 / DeepSeek / Qwen backbones (avg +40.3/+32.7/+23.4pp): the system improves, not a specific model

One framework optimizes skills, agent harnesses, and retrieval pipelines — no weight updates needed, fully white-box.

🔗 Code: https://github.com/ant-research/AntOmniEvo (quickstart in README)
📄 Paper: https://arxiv.org/abs/2609.35855

Bonus 🦞: it ships with a "lobster gym" visualizer — every candidate is a lobster leveling up, weak ones train in the basement, strong ones enter the golden hall. Full lineage tree, every fix traceable.

Happy to answer any questions — feedback and stars welcome!