MaxxBench changes the instructions or starting code of 89 DeepSWE tasks. The authors say the reference patches grow by a median of 17 lines, or 3.5%.
In the News: November 7, 2026 (Evening)
A new benchmark finds coding agents that tied on DeepSWE split by 23 points once requirements change, and that many guess at hidden tests instead.
Evening edition
Machine-readable
Download Markdown