# Source record Retrieved and inspected 11 November 2026. ## Full paper read - **Title:** GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning. - **Version used:** official ICLR 2026 proceedings PDF, 87 pages. - **[Official abstract](https://proceedings.iclr.cc/paper_files/paper/2026/hash/0e9e708b6f48e14fd0ac29e167413f76-Abstract-Conference.html)**. - **[Official PDF](https://proceedings.iclr.cc/paper_files/paper/2026/file/0e9e708b6f48e14fd0ac29e167413f76-Paper-Conference.pdf)**. - **[Local PDF](GEPA-ICLR-2026.pdf)**. - **[Local page-marked extraction](GEPA-ICLR-2026.txt)**. - **[arXiv record](https://arxiv.org/abs/2507.19457)**: first submitted 25 July 2025; v2 dated 14 February 2026. - **[arXiv v2 HTML](https://arxiv.org/html/2507.19457v2)**: cross-check only; appendix labels differ from the conference PDF. - **[Authors' research artifact](https://github.com/gepa-ai/gepa-artifact)**: README inspected; full experimental logs not downloaded or reproduced. The PDF's text was extracted with installed `pypdf`. Extraction includes 40,446 whitespace-delimited words including labels and page markers; it is not a publisher word count. Typography and some characters are imperfect in the extraction. Text order in dense tables is imperfect, so numerical checks used both reading order and layout extraction. The image-only prompt on page 82 was separately extracted and visually inspected. No OCR was needed for reading that image. The reading register in [01-GEPA-reading-notes.md](../01-GEPA-reading-notes.md) covers the main paper, references, and Appendices A–P. Every extracted page was read; inspection of text/plot captions is not independent validation of plotted measurements or the released experimental logs. `manifest.json` records the exact downloaded PDF hash, extraction hash, page count, and fingerprints of relevant current local evidence. These fingerprints identify the files inspected for this proposal; they do not establish the historical run's unrecorded harness version. ## Targeted primary literature screening Primary abstracts and selected method descriptions were checked for Self-Refine, Reflexion, TextGrad, MIPROv2, SPO, ProRefine, ACE, Feedback Descent, Prompt Duel Optimizer, Constitutional AI, and the October 2026 topic-disjoint reflective optimization preprint. Links and reading-depth limits are recorded in [06-evidence-and-related-work.md](../06-evidence-and-related-work.md); citation metadata is in [references.bib](../references.bib). The initial Feedback Descent OpenReview page was blocked by browser verification. Its arXiv primary record and indexed OpenReview method descriptions supplied the screening evidence instead. No claim is made that all versions, supplementary material, or code for these other papers were read. ## Scope of the dossier Statements about GEPA's measured performance are attributed to its authors. Statements about Cruise's current implementation are based on inspected source and exported documentation. The proposed method, hypotheses, task suite, budgets, and analysis are original planning recommendations and remain untested. The supplied PDF remains the authors' work, not part of an original manuscript written for this project.