Blog
- October 19, 2026Self-HPO: Self-Evolving Coding Agents via Harness-Policy Optimization
The policy reads its own failures, writes a better harness for itself, then learns inside it with RL. No external model required: Qwen3.5-35B-A3B goes from 61.0% to 69.5% on SWE-Bench Verified.
harnessesRLcoding agents - May 20, 2026 · NeurIPS 2026You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories
RLVR weight updates are nearly rank-1 and evolve almost linearly, so RELEX extrapolates future checkpoints from the first 15–20% of training and matches full RLVR.
RLVRefficiency