Jialin Zhang

CS @ UC Berkeley | Undergraduate researcher @berkeley_ai & @BerkeleySky

[Alex Dimakis @AlexGDimakis]

Check out Parallel GEPA. One surprising finding in evolutionary prompt optimization: multiple parallel proposals reduce overfitting.

[Shangyin Tan @ShangyinT]

I recommend checking out our new work that scales GEPA along a few interesting dimensions!! Led by our amazing undergrad collaborator @benzhang04

[Lakshya A Agrawal @LakshyAAAgrawal]

GEPA's core reflective optimization loop has been sequential until now. @benzhang04's release introduces PxN sampling, enabling each GEPA step to select multiple Pareto-frontier candidates and propose multiple mutations for each, fully batched to maximize candidate yield in a single latency wave.

This makes GEPA run 3–4× faster, and surprisingly, it also generalized better in our experiments (up to +11pp vs. sequential): exploring the search space more broadly transfers better beyond the validation set.

With P and N, GEPA now gives you 4 dials over the compute spent per step: how many candidates are mutated (P) and how many mutations each (N), both new here, plus reflection data (ComBEE minibatch size) and reflection compute (choice of LM, thinking effort as well as agentic proposers like Claude-Code).

GEPA now also supports custom batch_evaluators, that can interface with any custom evaluation/concurrency framework of your choice.

Try it in GEPA today!

[Jialin Zhang @benzhang04]

9/ Many thanks to our collaborators and advisors: @LakshyAAAgrawal @ShangyinT @lukedhlee @eren_lutfi78249 Dan Klein @koushik77 @AlexGDimakis @matei_zaharia. We really appreciate the helpful feedback from @lihanc02 and @dbreunig.
We’re also grateful to the GEPA open-source community, whose ideas and feedback continually provide new insights and helped motivate this improvement.

[Jialin Zhang @benzhang04]

1/ We just made GEPA’s optimization loop much faster, along with improved generalization!

Instead of proposing and evaluating one candidate at a time, each GEPA step now proposes a batch of candidates and evaluates them all concurrently. Parallel proposals significantly cut wall-clock time (3–4× faster) while, surprisingly, also generalizing better (up to +11 pts). 🧵

[Jialin Zhang @benzhang04]

8/ Check out the full blog here: https://t.co/cEeclnCOKn

[Shangyin Tan @ShangyinT]

Many LLM-based optimization algorithms have been proposed recently. We find that GEPA, AutoResearch, and Meta-Harness each win on different tasks and no single LLM optimizer wins everywhere.

How do we leverage the complementary performance profiles of these optimization algorithms?

Today, we release optimize_anything omni: a meta-optimizer that leverages the best of each optimizer. At a matched budget, omni beats every standalone optimizer. (thread)

[Dawn Song @dawnsongtweets]

Really excited for the Agentic AI Summit 2025 at @UCBerkeley—2K+ in-person attendees and ~10K online! Building on our 25K+ LLM Agents MOOC community, this is the premier global forum for advancing #AgenticAI.
👀 Livestream starts at 9:15 AM PT on August 2—tune in! https://t.co/eUeJw4qf5A