Mara Chain: Rethinking Failure as a Stepping Stone for AI System Auto-Evolution
Rejected AI-system candidates can carry enough signal to make the next optimization attempt better.
Mara Chain keeps failed prompt, skill, harness, or code configurations and refines them instead of dropping them after evaluation. The paper says this avoids repeating the same failure modes across later proposals. In tests on AppWorld, TerminalBench 2.1, and MuSiQue, the method improved task performance while using fewer rollouts than several comparison systems. HF Daily Papers' note
Mara Chain keeps failed prompt, skill, harness, or code configurations and refines them instead of dropping them after evaluation. The paper says this avoids repeating the same failure modes across later proposals. In tests on AppWorld, TerminalBench 2.1, and MuSiQue, the method improved task performance while using fewer rollouts than several comparison systems. HF Daily Papers' note
score 5