<?xml version="1.0" encoding="UTF-8" ?> <?xml-stylesheet type="text/xsl" href="rss.xsl"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/"> <channel> <title>GEPA</title><description>A framework for optimizing textual system components using LLM-based reflection and Pareto-efficient evolutionary search.</description><link>https://gepa-ai.github.io/gepa/</link><atom:link href="https://gepa-ai.github.io/gepa/blog/feed.xml" rel="self" type="application/rss+xml" /> <docs>https://github.com/gepa-ai/gepa</docs><language>en</language> <pubDate>Mon, 09 Nov 2026 07:48:16 -0000</pubDate> <lastBuildDate>Mon, 09 Nov 2026 07:48:16 -0000</lastBuildDate> <ttl>1440</ttl> <generator>MkDocs RSS plugin - v1.20.0</generator> <image> <url>None</url> <title>GEPA</title> <link>https://gepa-ai.github.io/gepa/</link> </image> <item> <title>Learned Error Diagnosis Improves GEPA&#39;s Reflection</title> <dc:creator>Andrei Cojocaru</dc:creator> <dc:creator>Lakshya A Agrawal</dc:creator> <dc:creator>Mert Cemri</dc:creator> <dc:creator>Joseph E. Gonzalez</dc:creator> <dc:creator>Alexandros G. Dimakis</dc:creator> <dc:creator>Ion Stoica</dc:creator> <dc:creator>Matei Zaharia</dc:creator> <description>AdaMAST learns a failure taxonomy from a program&#39;s own traces, and an LLM judge that never sees the score diagnoses every execution against it. Feeding those findings to GEPA&#39;s reflection model lifts held-out scores by up to 9.7 percentage points under the same online optimization budget.</description> <link>https://gepa-ai.github.io/gepa/blog/2026/10/06/named-failure-modes/</link> <pubDate>Tue, 06 Oct 2026 00:00:00 +0000</pubDate> <source url="https://gepa-ai.github.io/gepa/blog/feed.xml">GEPA</source><guid isPermaLink="true">https://gepa-ai.github.io/gepa/blog/2026/10/06/named-failure-modes/</guid> <enclosure url="https://gepa-ai.github.io/gepa/assets/images/social/blog/2026/10/06/named-failure-modes.png" type="image/png" length="0" /> </item> <item> <title>Optimize Cheap, Deploy Strong: A Recipe for Cost-Efficient GEPA</title> <dc:creator>Tal Oved</dc:creator> <dc:creator>Roi Pony</dc:creator> <dc:creator>Oshri Naparstek</dc:creator> <dc:creator>Udi Barzelay</dc:creator> <dc:creator>Lakshya A Agrawal</dc:creator> <description>GEPA&#39;s cost is dominated by the task model that scores every candidate. Move that role to the cheapest model, keep a strong reflector for the rare edits, and deploy the evolved prompt zero-shot on a stronger model. On HotpotQA that is +9.0 points at about 14x less cost. We call it positive transfer: a prompt tuned on a cheap model can beat one the strong model tuned for itself.</description> <link>https://gepa-ai.github.io/gepa/blog/2026/09/29/optimize-cheap-deploy-strong/</link> <pubDate>Tue, 29 Sep 2026 00:00:00 +0000</pubDate> <source url="https://gepa-ai.github.io/gepa/blog/feed.xml">GEPA</source><guid isPermaLink="true">https://gepa-ai.github.io/gepa/blog/2026/09/29/optimize-cheap-deploy-strong/</guid> <enclosure url="https://gepa-ai.github.io/gepa/assets/images/social/blog/2026/09/29/optimize-cheap-deploy-strong.png" type="image/png" length="0" /> </item> <item> <title>Bridging the Subjectivity Gap: How Automated Prompt Optimization Helps Teams Build Expert-Aligned AI Functions</title> <dc:creator>Seth Kimmel, Sutro Team</dc:creator> <description>A case study by Sutro showing how automated prompt optimization aligns repeated AI decisions with organization-specific judgment using only 30 annotations.</description> <link>https://gepa-ai.github.io/gepa/blog/2026/09/17/bridging-the-subjectivity-gap/</link> <pubDate>Thu, 17 Sep 2026 00:00:00 +0000</pubDate> <source url="https://gepa-ai.github.io/gepa/blog/feed.xml">GEPA</source><guid isPermaLink="true">https://gepa-ai.github.io/gepa/blog/2026/09/17/bridging-the-subjectivity-gap/</guid> <enclosure url="https://gepa-ai.github.io/gepa/assets/images/social/blog/2026/09/17/bridging-the-subjectivity-gap.png" type="image/png" length="0" /> </item> <item> <title>Batching the Reflective Optimization Loop: Parallel Proposals Make GEPA Faster and Better</title> <dc:creator>Jialin Zhang</dc:creator> <dc:creator>Lakshya A Agrawal</dc:creator> <dc:creator>Shangyin Tan</dc:creator> <dc:creator>Donghyun Lee</dc:creator> <dc:creator>Lutfi Eren Erdogan</dc:creator> <dc:creator>Dan Klein</dc:creator> <dc:creator>Koushik Sen</dc:creator> <dc:creator>Alexandros G. Dimakis</dc:creator> <dc:creator>Matei Zaharia</dc:creator> <description>GEPA now supports proposing and evaluating a batch of candidates on each optimization step instead of one candidate at a time. In our sweep on two tasks, most batched runs finished in half the wall-clock time or less, and the fastest in about a quarter to a third. Batched settings also achieved higher held-out test scores: from 68.9% to 72.1% on LiveBench-Math (with 2×2) and from 49.0% to 60.0% on HoVer (with 8×1).</description> <link>https://gepa-ai.github.io/gepa/blog/2026/07/30/parallel-proposals/</link> <pubDate>Thu, 30 Jul 2026 00:00:00 +0000</pubDate> <source url="https://gepa-ai.github.io/gepa/blog/feed.xml">GEPA</source><guid isPermaLink="true">https://gepa-ai.github.io/gepa/blog/2026/07/30/parallel-proposals/</guid> <enclosure url="https://gepa-ai.github.io/gepa/assets/images/social/blog/2026/07/30/parallel-proposals.png" type="image/png" length="0" /> </item> <item> <title>optimize_anything Goes omni: Composing Optimizers into Meta-Optimizer Pipelines</title> <dc:creator>Shangyin Tan</dc:creator> <dc:creator>Lakshya A Agrawal</dc:creator> <dc:creator>Donghyun Lee</dc:creator> <dc:creator>Jialin Zhang</dc:creator> <dc:creator>Dan Klein</dc:creator> <dc:creator>Koushik Sen</dc:creator> <dc:creator>Alexandros G. Dimakis</dc:creator> <dc:creator>Matei Zaharia</dc:creator> <description>optimize_anything now supports customizable optimizers that can be composed into meta-optimizer pipelines. Why compose? Because no single optimizer wins everywhere. Our omni meta-optimizer combines the best of each and beats every standalone optimizer on Frontier-CS tasks under a matched budget.</description> <link>https://gepa-ai.github.io/gepa/blog/2026/07/22/optimize-anything-omni/</link> <pubDate>Wed, 22 Jul 2026 00:00:00 +0000</pubDate> <source url="https://gepa-ai.github.io/gepa/blog/feed.xml">GEPA</source><guid isPermaLink="true">https://gepa-ai.github.io/gepa/blog/2026/07/22/optimize-anything-omni/</guid> <enclosure url="https://gepa-ai.github.io/gepa/assets/images/social/blog/2026/07/22/optimize-anything-omni.png" type="image/png" length="0" /> </item> <item> <title>Learning, Fast and Slow: Towards LLMs That Adapt Continually</title> <dc:creator>Rishabh Tiwari</dc:creator> <dc:creator>Kusha Sareen</dc:creator> <dc:creator>Lakshya A Agrawal</dc:creator> <dc:creator>Joseph E. Gonzalez</dc:creator> <dc:creator>Matei Zaharia</dc:creator> <dc:creator>Kurt Keutzer</dc:creator> <dc:creator>Inderjit S Dhillon</dc:creator> <dc:creator>Rishabh Agarwal</dc:creator> <dc:creator>Devvrit Khatri</dc:creator> <description>Fast-Slow Training (FST) interleaves prompt optimization with reinforcement learning, treating the prompt as fast weights and parameters as slow weights, improving data efficiency, performance ceiling, plasticity, and continual learning.</description> <link>https://gepa-ai.github.io/gepa/blog/2026/05/11/learning-fast-and-slow/</link> <pubDate>Mon, 11 May 2026 00:00:00 +0000</pubDate> <source url="https://gepa-ai.github.io/gepa/blog/feed.xml">GEPA</source><guid isPermaLink="true">https://gepa-ai.github.io/gepa/blog/2026/05/11/learning-fast-and-slow/</guid> <enclosure url="https://gepa-ai.github.io/gepa/assets/images/social/blog/2026/05/11/learning-fast-and-slow.png" type="image/png" length="0" /> </item> <item> <title>Scaling GEPA with Combee: Parallel Prompt Learning for Self-Improving Agents</title> <dc:creator>Hanchen Li</dc:creator> <dc:creator>Runyuan He</dc:creator> <dc:creator>Qizheng Zhang</dc:creator> <dc:creator>Changxiu Ji</dc:creator> <dc:creator>Qiuyang Mang</dc:creator> <dc:creator>Xiaokun Chen</dc:creator> <dc:creator>Lakshya A Agrawal</dc:creator> <dc:creator>Wei-Liang Liao</dc:creator> <dc:creator>Eric Yang</dc:creator> <dc:creator>Alvin Cheung</dc:creator> <dc:creator>James Zou</dc:creator> <dc:creator>Kunle Olukotun</dc:creator> <dc:creator>Ion Stoica</dc:creator> <dc:creator>Joseph E. Gonzalez</dc:creator> <description>Exploring how to scale GEPA&#39;s prompt learning across many parallel agents using the Combee framework, achieving up to 17× speedup with no quality loss.</description> <link>https://gepa-ai.github.io/gepa/blog/2026/04/09/gepa-at-scale-with-combee/</link> <pubDate>Thu, 09 Apr 2026 00:00:00 +0000</pubDate> <source url="https://gepa-ai.github.io/gepa/blog/feed.xml">GEPA</source><guid isPermaLink="true">https://gepa-ai.github.io/gepa/blog/2026/04/09/gepa-at-scale-with-combee/</guid> <enclosure url="https://gepa-ai.github.io/gepa/assets/images/social/blog/2026/04/09/gepa-at-scale-with-combee.png" type="image/png" length="0" /> </item> <item> <title>Confidence-Aware Prompt Optimization for LLM Classification</title> <dc:creator>Rodolfo Nobrega</dc:creator> <description>ConfidenceAdapter uses token-level log-probabilities to score prompt candidates on a continuous scale instead of binary correct/wrong, producing better disambiguation rules and higher-accuracy prompts for classification tasks.</description> <link>https://gepa-ai.github.io/gepa/blog/2026/03/17/confidence-adapter-benchmark/</link> <pubDate>Tue, 17 Mar 2026 00:00:00 +0000</pubDate> <source url="https://gepa-ai.github.io/gepa/blog/feed.xml">GEPA</source><guid isPermaLink="true">https://gepa-ai.github.io/gepa/blog/2026/03/17/confidence-adapter-benchmark/</guid> <enclosure url="https://gepa-ai.github.io/gepa/assets/images/social/blog/2026/03/17/confidence-adapter-benchmark.png" type="image/png" length="0" /> </item> <item> <title>Automatically Learning Skills for Coding Agents</title> <dc:creator>Shangyin Tan</dc:creator> <dc:creator>Lakshya A Agrawal</dc:creator> <dc:creator>Rohit Sandadi</dc:creator> <dc:creator>Dan Klein</dc:creator> <dc:creator>Koushik Sen</dc:creator> <dc:creator>Alexandros G. Dimakis</dc:creator> <dc:creator>Matei Zaharia</dc:creator> <description>Introducing gskill, a fully automated pipeline that uses GEPA and SWE-smith to learn repository-specific skills for coding agents.</description> <link>https://gepa-ai.github.io/gepa/blog/2026/02/18/automatically-learning-skills-for-coding-agents/</link> <pubDate>Wed, 18 Feb 2026 00:00:00 +0000</pubDate> <source url="https://gepa-ai.github.io/gepa/blog/feed.xml">GEPA</source><guid isPermaLink="true">https://gepa-ai.github.io/gepa/blog/2026/02/18/automatically-learning-skills-for-coding-agents/</guid> <enclosure url="https://gepa-ai.github.io/gepa/assets/images/social/blog/2026/02/18/automatically-learning-skills-for-coding-agents.png" type="image/png" length="0" /> </item> <item> <title>optimize_anything: A Universal API for Optimizing any Text Parameter</title> <dc:creator>Lakshya A Agrawal</dc:creator> <dc:creator>Donghyun Lee</dc:creator> <dc:creator>Wenjie Ma</dc:creator> <dc:creator>Karim Elmaaroufi</dc:creator> <dc:creator>Shangyin Tan</dc:creator> <dc:creator>Sanjit A. Seshia</dc:creator> <dc:creator>Koushik Sen</dc:creator> <dc:creator>Dan Klein</dc:creator> <dc:creator>Ion Stoica</dc:creator> <dc:creator>Joseph E. Gonzalez</dc:creator> <dc:creator>Omar Khattab</dc:creator> <dc:creator>Alexandros G. Dimakis</dc:creator> <dc:creator>Matei Zaharia</dc:creator> <description>GEPA&#39;s new API setting state-of-the-art results on optimizing any text parameter: code, prompts, agent architectures, and more. If you can measure it, you can optimize it.</description> <link>https://gepa-ai.github.io/gepa/blog/2026/02/18/introducing-optimize-anything/</link> <pubDate>Wed, 18 Feb 2026 00:00:00 +0000</pubDate> <source url="https://gepa-ai.github.io/gepa/blog/feed.xml">GEPA</source><guid isPermaLink="true">https://gepa-ai.github.io/gepa/blog/2026/02/18/introducing-optimize-anything/</guid> <enclosure url="https://gepa-ai.github.io/gepa/assets/images/social/blog/2026/02/18/introducing-optimize-anything.png" type="image/png" length="0" /> </item> <item> <title>Introducing the GEPA Blog</title> <dc:creator>Lakshya A Agrawal</dc:creator> <description>Research updates, applications, and community case studies from the GEPA ecosystem.</description> <link>https://gepa-ai.github.io/gepa/blog/2026/02/13/introducing-the-gepa-blog/</link> <pubDate>Fri, 13 Feb 2026 00:00:00 +0000</pubDate> <source url="https://gepa-ai.github.io/gepa/blog/feed.xml">GEPA</source><guid isPermaLink="true">https://gepa-ai.github.io/gepa/blog/2026/02/13/introducing-the-gepa-blog/</guid> <enclosure url="https://gepa-ai.github.io/gepa/assets/images/social/blog/2026/02/13/introducing-the-gepa-blog.png" type="image/png" length="0" /> </item> </channel> </rss>