--- schema: 1 type: video title: >- Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA description: >- After one round of reflection on just three examples, GEPA doubled the gains that the RL algorithm GRPO reached after 25,000 rollouts. Lakshya A. Agrawal, creator of GEPA and a PhD student at UC Berkeley's Sky Computing Lab, explains why. RL squeezes a whole rollout down to a single score. GEPA instead has a model read the full trace, including chains of thought, tool calls and error messages, and write a better prompt. A Pareto pool of candidates keeps the search out of local optima. The same idea works on anything you can write as text and score, which is where Optimize Anything comes in. Agrawal shows it taking a four-line program to a six-step agent that lifts Gemini Flash on ARC-AGI from 32.5% to 89.5%. It pushed an AMD NPU coding agent from 4% to 30%, and improved a GPT-5 mini coding agent on Go issues from 24% to 93%, using learned skills that also carried over to Claude Sonnet. He also covers Databricks tuning an open model to beat Claude Opus at 90x lower cost, learning evals from production traces, and co-optimizing prompts and weights. Speaker info: X/Twitter: @LakshyAAAgrawal (https://x.com/LakshyAAAgrawal) LinkedIn: https://www.linkedin.com/in/lakshyaaagrawal/ Website: https://lakshyaaagrawal.github.io/ Related links: GEPA on GitHub: https://github.com/gepa-ai/gepa Timestamps: 0:00 Intro: reflective optimization 0:30 How we teach AI new tasks 1:00 The sample-efficiency bottleneck 2:05 What RL throws away 3:00 Reflecting in text space 4:25 GEPA 4:50 GEPA vs. GRPO: 3 examples vs. 25,000 rollouts 5:40 What GEPA learns 6:55 GPT-4.1 mini beats GPT-4.1 7:15 A new AMD NPU: 4% to 30% 8:25 How GEPA works 9:00 Why a Pareto pool beats a simple loop 10:10 Results across benchmarks 10:35 Beyond prompts: Optimize Anything 12:05 The API 12:55 A 3D unicorn 13:35 Discovering agent harnesses: ARC-AGI from 32.5% to 89.5% 14:40 MATH-500 15:10 Optimizing agent skills: 24% to 93% 16:40 Three optimization modes 17:30 In production: OCR and a 90x cheaper agent 18:40 Why better models need better prompts 19:10 Learning evals from production traces 19:55 Fast-slow learning: prompts and weights 20:20 Who uses GEPA 20:45 Get started --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/OA-Mc60Rboo-youtube-c7b7a577aa52a7d42876 Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/beating-rl-with-reflection-gepa-and-optimize-anythi-528973dac9ab markdown: >- https://liquidgold.cc/channel/aidotengineer/video/beating-rl-with-reflection-gepa-and-optimize-anythi-528973dac9ab.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:20:31.726Z' language: null children: [] videoId: OA-Mc60Rboo durationSeconds: 1288 publishedAt: '2026-09-26T17:00:06.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 12 endSeconds: 24 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=12' heldReason: rights - ordinal: 1 text: null startSeconds: 45 endSeconds: 68 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=45' heldReason: rights - ordinal: 2 text: null startSeconds: 68 endSeconds: 78 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=68' heldReason: rights - ordinal: 3 text: null startSeconds: 134 endSeconds: 146 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=134' heldReason: rights - ordinal: 4 text: null startSeconds: 182 endSeconds: 192 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=182' heldReason: rights - ordinal: 5 text: null startSeconds: 192 endSeconds: 267 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=192' heldReason: rights - ordinal: 6 text: null startSeconds: 267 endSeconds: 279 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=267' heldReason: rights - ordinal: 7 text: null startSeconds: 302 endSeconds: 344 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=302' heldReason: rights - ordinal: 8 text: null startSeconds: 355 endSeconds: 421 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=355' heldReason: rights - ordinal: 9 text: null startSeconds: 421 endSeconds: 433 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=421' heldReason: rights - ordinal: 10 text: null startSeconds: 457 endSeconds: 514 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=457' heldReason: rights - ordinal: 11 text: null startSeconds: 514 endSeconds: 526 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=514' heldReason: rights - ordinal: 12 text: null startSeconds: 560 endSeconds: 615 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=560' heldReason: rights - ordinal: 13 text: null startSeconds: 615 endSeconds: 628 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=615' heldReason: rights - ordinal: 14 text: null startSeconds: 638 endSeconds: 649 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=638' heldReason: rights - ordinal: 15 text: null startSeconds: 649 endSeconds: 734 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=649' heldReason: rights - ordinal: 16 text: null startSeconds: 734 endSeconds: 745 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=734' heldReason: rights - ordinal: 17 text: null startSeconds: 833 endSeconds: 888 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=833' heldReason: rights - ordinal: 18 text: null startSeconds: 888 endSeconds: 899 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=888' heldReason: rights - ordinal: 19 text: null startSeconds: 920 endSeconds: 1000 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=920' heldReason: rights - ordinal: 20 text: null startSeconds: 920 endSeconds: 1000 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=920' heldReason: rights - ordinal: 21 text: null startSeconds: 1012 endSeconds: 1056 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=1012' heldReason: rights - ordinal: 22 text: null startSeconds: 1056 endSeconds: 1067 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=1056' heldReason: rights - ordinal: 23 text: null startSeconds: 1067 endSeconds: 1121 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=1067' heldReason: rights - ordinal: 24 text: null startSeconds: 1121 endSeconds: 1131 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=1121' heldReason: rights - ordinal: 25 text: null startSeconds: 1165 endSeconds: 1196 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=1165' heldReason: rights - ordinal: 26 text: null startSeconds: 1196 endSeconds: 1207 watch: 'https://liquidgold.cc/watch/OA-Mc60Rboo?t=1196' heldReason: rights --- # Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA After one round of reflection on just three examples, GEPA doubled the gains that the RL algorithm GRPO reached after 25,000 rollouts. Lakshya A. Agrawal, creator of GEPA and a PhD student at UC Berkeley's Sky Computing Lab, explains why. RL squeezes a whole rollout down to a single score. GEPA instead has a model read the full trace, including chains of thought, tool calls and error messages, and write a better prompt. A Pareto pool of candidates keeps the search out of local optima. The same idea works on anything you can write as text and score, which is where Optimize Anything comes in. Agrawal shows it taking a four-line program to a six-step agent that lifts Gemini Flash on ARC-AGI from 32.5% to 89.5%. It pushed an AMD NPU coding agent from 4% to 30%, and improved a GPT-5 mini coding agent on Go issues from 24% to 93%, using learned skills that also carried over to Claude Sonnet. He also covers Databricks tuning an open model to beat Claude Opus at 90x lower cost, learning evals from production traces, and co-optimizing prompts and weights. Speaker info: X/Twitter: @LakshyAAAgrawal (https://x.com/LakshyAAAgrawal) LinkedIn: https://www.linkedin.com/in/lakshyaaagrawal/ Website: https://lakshyaaagrawal.github.io/ Related links: GEPA on GitHub: https://github.com/gepa-ai/gepa Timestamps: 0:00 Intro: reflective optimization 0:30 How we teach AI new tasks 1:00 The sample-efficiency bottleneck 2:05 What RL throws away 3:00 Reflecting in text space 4:25 GEPA 4:50 GEPA vs. GRPO: 3 examples vs. 25,000 rollouts 5:40 What GEPA learns 6:55 GPT-4.1 mini beats GPT-4.1 7:15 A new AMD NPU: 4% to 30% 8:25 How GEPA works 9:00 Why a Pareto pool beats a simple loop 10:10 Results across benchmarks 10:35 Beyond prompts: Optimize Anything 12:05 The API 12:55 A 3D unicorn 13:35 Discovering agent harnesses: ARC-AGI from 32.5% to 89.5% 14:40 MATH-500 15:10 Optimizing agent skills: 24% to 93% 16:40 Three optimization modes 17:30 In production: OCR and a 90x cheaper agent 18:40 Why better models need better prompts 19:10 Learning evals from production traces 19:55 Fast-slow learning: prompts and weights 20:20 Who uses GEPA 20:45 Get started --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/OA-Mc60Rboo-youtube-c7b7a577aa52a7d42876 Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/beating-rl-with-reflection-gepa-and-optimize-anythi-528973dac9ab · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:12 to 1:08) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:08 to 2:14) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (2:14 to 3:02) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (3:02 to 4:27) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (4:27 to 4:50) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (4:50 to 5:44) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (5:44 to 7:01) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (7:01 to 7:13) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (7:13 to 8:34) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (8:34 to 9:09) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 10: Document excerpt (9:09 to 10:15) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 11: Document excerpt (10:15 to 10:49) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 12: Document excerpt (10:49 to 12:14) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 13: Document excerpt (12:14 to 12:58) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 14: Document excerpt (12:58 to 13:42) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 15: Document excerpt (13:42 to 14:48) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 16: Document excerpt (14:48 to 15:10) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 17: Document excerpt (15:10 to 16:40) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 18: Document excerpt (16:40 to 17:36) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 19: Document excerpt (17:36 to 18:41) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 20: Document excerpt (18:41 to 19:13) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 21: Document excerpt (19:13 to 19:56) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 22: Document excerpt (19:56 to 20:41) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 23: Document excerpt (20:41 to 21:28) > Held: transcript rights do not permit this chunk's text.