The authors used prolonged overfitting on a small pool of solvable software tasks as a diagnostic test.
Careful system design makes asynchronous coding-agent training substantially stronger
Arctic-SWE combines established reinforcement-learning controls with curated tasks to improve 4B and 9B coding models across three software-engineering benchmarks.
Top University
Boyi Liu · Yite Wang · Xinyu Lian · Fan Shu · Sida Li · Feng Yan · +3 more
Snowflake · University of Illinois at Urbana-Champaign · University of Houston · University of Chicago
Research Digest··3 min read
Liu et al.
§