CoCoEvolve preprint reports gains in AI data workflow benchmarks

Researchers report that CoCoEvolve, an optimizer for AI systems, reached 92.8% accuracy on held-out hard questions in DABStep, compared with 78.4% for Meta-Harness and 30.4% for GEPA under matched conditions. The preprint evaluates systems that answer data questions, run language-model data pipelines and perform data engineering workflows. Instead of changing a candidate using only its failures, CoCoEvolve retrieves evidence from earlier optimization attempts and probes the live system before verifying and proposing changes. On data-eng-bench, it increased the test-set solve rate from 59.1% to 77.4%, exceeding the best baseline of 73.6%. These results measure benchmark performance rather than business outcomes from deployment.