GEPA Optimization Upgrade: Parallel Sampling Delivers 3-4x Speedup and Better Generalization
iamrobotbear · x · 2026-08-06
Developer @benzhang04 released a major update to the GEPA optimization loop, introducing PxN sampling.
- Speed: Moving away from sequential reflective optimization, each step now proposes and evaluates batches of candidates concurrently, achieving a 3-4x wall-clock speedup.
- Generalization: Surprisingly, exploring the search space more broadly transfers better beyond the validation set, yielding up to an 11 percentage point improvement.
- Compute Control: Users gain four dials to manage compute per step, including the number of candidates mutated (P), mutations per candidate (N), reflection data size, and reflection compute (choice of LM or agentic proposers like Claude-Code).
More from coding & agent
- CrowdStrike and AWS Launch $100k Prompt Injection AI Challenge — AccBalanced · 2026-08-06
- Open-Source Project Clones Zack D. Films 3D Animated Short Workflow — matchaman11 · 2026-08-06
- Muse Code Introduces Persistent Background Agents for Parallel Task Execution — alexandr_wang · 2026-08-06
- Latent Labs Launches Latent-Y, First Lab-Validated Drug Design Agent — chaitjo · 2026-08-06
- OpenAI's Codex CLI is Open Source, Surpasses 100k GitHub Stars — dkundel · 2026-08-06
- Using Claude with Local ComfyUI MCP for Conversational Video Generation — EndPsychological8822 · 2026-08-06