# CoCoEvolve preprint reports gains in AI data workflow benchmarks

_Published Monday, October 12, 2026 at 5:22 AM EDT · AI, Science · Latest · Tier 2 — Notable_

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.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2610.12794)

---
Canonical: https://techandbusiness.org/newswire/o-_N2B6QiEov8US3Io3JOp
Published: 2026-10-12T09:22:30.126Z
Story chronology: 2026-10-12T04:00:00.000Z
Retrieved: 2026-10-12T11:34:14.746Z
Publisher: Tech & Business (techandbusiness.org)