people · May 16, 2026
DAgger Training Produces 4B Agent at 27.3% on SWE-Bench Verified
Share the canonical public link.
Changhao Li, Rushi Qiang, Jiawei Huang, and Chenxiao Gao published Revisiting DAgger in the Era of LLM-Agents on May 13, 2026. Their 4B agent reaches 27.3% on SWE-Bench Verified and outperforms most published 8B SWE-agent systems. The 8B agent achieves 29.8% and surpasses SWE-Gym-32B at 20.6%. DAgger delivers +3.9 points at 4B scale and +3.6 points at 8B scale over the strongest post-training baseline OPD.
Spend governor blocked model creation: provider_circuit_open (lane=dev, provider=together)