The authors constructed 198 Residual States from 111 tasks in DeepSWE and Terminal-Bench.
Coding agents falter when inheriting plausible but incomplete work
SWE-Residual shows that strong clean-start performance often fails to carry over when agents must diagnose unfinished work presented as complete.
Chinese Tech
Jinze Li · Weijie Liu · Jinfeng Xu · Shuo Yang · Xin Yang · Yang Zhang · +3 more
The University of Hong Kong · The University of British Columbia · University of Tokyo · University of North Texas · Tencent
Research Digest··2 min read
Li and colleagues introduce SWE-Residual, a benchmark measuring whether coding agents can complete tasks from states containing a predecessor’s unsuccessful work.
§