Chejian Xu and co-authors write about Evo2, a framework for weight-harness co-evolution in self-evolving agents. Existing methods typically improve either model weights or the agent harness separately, overlooking the relationship between the two. Evo2 proposes a dual-role pipeline where a single policy serves as both a Proposer (writing harness edits) and a Reflector (diagnosing their effects). The training pipeline EvoTIDE recovers usable trajectories through edit extraction and prefix continuation, combining supervised fine-tuning with interleaved GRPO.
- Achieves 82.1% on Polyglot and 48.3% on SWE-bench Verified, outperforming the strongest baselines (62.6% and 41.0% respectively)
- Evolved harnesses transfer to different tasks without further evolution and rank first in cross-benchmark evaluations
- The framework extends to proprietary models, improving a GPT-5-mini agent's performance from 24.4% to 48.3%
- Achieves 82.1% on Polyglot and 48.3% on SWE-bench Verified, outperforming the strongest baselines (62.6% and 41.0% respectively)
- Evolved harnesses transfer to different tasks without further evolution and rank first in cross-benchmark evaluations
- The framework extends to proprietary models, improving a GPT-5-mini agent's performance from 24.4% to 48.3%
