OmniHarness lets visual agents self-practice, hitting 95% on creative tasks

Researchers from BUAA, CUHK and NUS released OmniHarness, a symbolic policy learning framework that lets visual agents practice autonomously and accumulate reusable strategies before downstream tasks arrive, achieving a 95% solve rate on ComfyBench creative tasks.

2026-09-17 ~ 2026-09-18 · 2 related posts

• OmniHarness: symbolic policy learning boosts generalizable visual generation — Xu Xu · 2026-09-17
• OmniHarness: visual agents that self-practice and bank verified workflows hit 95% on creative tasks — 新智元 · 2026-09-18