The authors studied trajectories from six single-agent and multi-agent scaffolds across SWE-bench-Verified, SWE-bench-Pro and SWE-bench-Lite.
Guided code reasoning improves agents’ repository issue resolution
DeskCheck systematically simulates execution, audits patches and designs corner-case tests, outperforming two existing agent scaffolds on SWE-bench-Pro.
Independent
Changshu Liu · Ji young Kim · Alireza Ghazanfari · Saanvi Nallu · Reyhaneh Jabbarvand
Research Digest··3 min read
95 percent of them, with greater reasoning associated with higher success.
§