Ant Group's Mara Chain: refine prompts/harness/code, matches GEPA with far fewer rollouts
jiqizhixin · x · 2026-11-14
Mara Chain: optimizing AI systems without touching weights
Mara Chain, from Ant Group and collaborators, optimizes the mutable layer of AI systems — prompts, skills, harnesses and code — while keeping model weights fixed.
Key ideas:
- Failed edits aren't discarded: a bounded refinement chain inherits scores, diagnostics, change records and execution traces, letting later edits complete partial repairs before retention is evaluated.
- Pareto filtering keeps candidates with complementary task-level strengths, followed by a top-N pool limit.
Results:
- On AppWorld, it reaches a 0.8 validation score in 3,280 rollouts vs GEPA's 9,514 under the paper's setup.
- On the 89-task TerminalBench 2.1 with GLM-5, the optimized harness hits 71.9% pass rate.
More from coding & agent
- Infinitty overhaul: voice-driven Mac AI terminal with 20-agent multiplayer editing — jasonkneen · 2026-11-14
- Together AI webinar dissects Ripple-65B coding model: dense vs MoE, ablations, RL runs — togethercompute · 2026-11-14
- Agent harness or deterministic routing? Reddit debates the right AI architecture for SaaS — Odd_Phone_5660 · 2026-11-14
- YC F26 startup Anywhere launches a map app that lets AI agents read and edit your places — ycombinator · 2026-11-14
- Trail of Bits open-sources patching-bugs and regression-hunt agent plugins, cuts regression false positives 25% — JiaweiLiu_ · 2026-11-14
- Claude wrote all ~40k lines: mech arena game Brontide opens new Cavern map — clockspammer4life · 2026-11-14