Trending storyDeveloping

Google DeepMind proposes Evo2 to co-evolve agent harness and weights

1 article1 sourcesince Oct 13Last article 1h ago · Oct 13, 9:57 AM ET

Overview

AISummary of 1 article

Google DeepMind proposes Evo2, a framework in which one policy acts as both Proposer and Reflector to co-evolve an agent's harness and its weights.

The paper reports 48.3% on SWE-bench Verified and 82.1% on Polyglot, against 41.0% and 62.6% for the strongest baselines.

The paper also reports that harness changes alone raise a GPT-5-mini agent from 24.4% to 48.3% on SWE-bench Verified. These results are the paper's own reported figures and have not been independently verified.

Written by AI from the articles below · updated Oct 13, 10:18 AM ET

Check the sources:

Latest developmentOct 13, 9:57 AM ET

Google DeepMind's Evo2 co-evolves an agent's harness and weights

Article timeline

The articles in this story. Times are ET.

Oct 13

9:57 AM
elvis@omarsar0X
Google DeepMind's Evo2 co-evolves an agent's harness and weights
AIGoogle DeepMind proposes Evo2, a framework in which one policy acts as both Proposer and Reflector to co-evolve its harness and weights. The paper reports 48.3% on SWE-bench Verified and 82.1% on Polyglot, against 41.0% and 62.6% for the strongest baselines. The harness changes alone raise a GPT-5-mini agent from 24.4% to 48.3% on SWE-bench Verified.

Heat trend

Not enough continuous observations to show a trend yet.