Termination Lag: Coding Agents Solve Early but Stop Late
In 38–82% of successful runs, coding agents produced a correct solution on their first source-code edit.
Researchers tested every distinct repository patch across five models, two agent frameworks and three SWE-bench datasets. On SWE-bench Verified and Multilingual, steps after the first correct patch accounted for a median 35–37% of successful runs for Qwen3-Coder and 54–60% for Laguna-XS.2. Most of that additional work went into verification and repository exploration. In 1–8% of cases, continued activity undid an earlier correct patch, leaving the final solution incorrect.
HF Daily Papers' note
Researchers tested every distinct repository patch across five models, two agent frameworks and three SWE-bench datasets. On SWE-bench Verified and Multilingual, steps after the first correct patch accounted for a median 35–37% of successful runs for Qwen3-Coder and 54–60% for Laguna-XS.2. Most of that additional work went into verification and repository exploration. In 1–8% of cases, continued activity undid an earlier correct patch, leaving the final solution incorrect.
HF Daily Papers' note
score 5