product · May 16, 2026
ChangHao Releases Paper Revisiting DAgger for LLM-Agents on SWE-bench Verified
Share the canonical public link.
ChangHao published the paper Revisiting DAgger in the Era of LLM-Agents on May 14 2026. The work applies DAgger to train long-horizon LLM agents using on-policy rollouts and dense teacher supervision without cold-start or white-box requirements. On SWE-bench Verified the 4B agent reached 27.3 percent accuracy and the 8B agent reached 29.8 percent accuracy. ChangHao's 8B agent surpassed SWE-Gym-32B and came within five points of strong 32B agents.
Failed after 3 attempts. Last error: Service unavailable