rsi.watch

Zack Yu et al. (11 authors) · 2026-10-27

GEPA-Curator: Curated Learning for Self Improving Agents

“GEPA-Curator matches or exceeds the mean performance of both baselines, with gains of up to 20.7% over GEPA and 21.2% over ACE.”
Who
Zack Yu et al. (11 authors)
Date
2026-10-27
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ imprecise. The reported spread is over repeated evaluations of each optimized prompt, and the cost section counts one optimization run per method and benchmark. The paper reports no spread across repeated optimization runs. “averaged over five evaluations of each optimized prompt”
how we rate evidence
Replication
No rerun found. The paper gives no link to its own code. We found no rerun by another group. Searched up to 2026-10-29. all claims
Checked
2026-10-29, against the source page
Links
Source Paper

Also on the page

  • An actor agent, gpt-oss-120b, runs tasks, and a Reflector revises its prompt from the resulting trajectories. GEPA-Curator adds a Curator, run on the same optimizer model (Gemini 3.1 Pro or Claude Opus 5), that picks which training tasks enter each minibatch of 16 and can attach one earlier successful trajectory of a failed task as a reference. It builds on "a simplified, greedy variant of GEPA that removes Pareto-based candidate selection".
  • On four GoBrowse websites with Claude Opus 5 as optimizer, GEPA improves held-out test success over the seed prompt by 6.4% and GEPA-Curator by 11.4%. On ALFWorld, AppWorld and τ-bench with Claude Opus 5, GEPA-Curator averages 8.1% over GEPA and 12.6% over ACE, and matches GEPA on τ-bench.
  • GoBrowse tasks are split 3:1:1 into train, validation and held-out test sets. The paper says "The evaluation numbers are averaged over 5 repeated runs"; its cost section counts "one run per method and benchmark". An ablation on GitLab splits its tasks "into 520 for training, 200 for val-idation"; GitLab has 720 tasks, so the ablation also used the GitLab tasks that are held-out test tasks in the main split.
  • In an appendix on four QA benchmarks from the GEPA paper, GEPA-Curator "matches or outperforms GEPA on three of four benchmarks" and is lower on HotpotQA. Applied to Meta-Harness, which edits harness code, with "one optimization experiment per method and setting", the authors report "modest improvements in held-out performance"; there, "USPTO accuracy is slightly lower", and on a math benchmark ProofBench accuracy goes from 35.56 to 33.89.

In the self-improving AI catalog

  • GEPA-Curator: what it changes, whether the change lasts, and how it was tested

How it got here

Earlier work in this catalog that GEPA-Curator cites, oldest first. Under each: where that work sits on the map, then where GEPA-Curator sits (changes in bold), and what GEPA-Curator's own paper says about it. Its whole family on the citation map

Cites Anthropic, recursive self-improvement 2026

← Ledger Report an error