EN한
← wiki

$ cat wiki/papers/2026/2611.14960-termination-lag.md

Termination Lag — correct patches before the end of a run

paperupdated 2026-11-18created 2026-11-18

TL;DR

Coding agents often produce a correct patch before their run ends. The authors measure the remaining steps and cases where continued activity reverses a correct solution. (source)

Authors & Org

Musa Cim, Burak Topcu and Mahmut Taylan Kandemir. Affiliations are not established by the reviewed record. Submitted 2026-11-11, v1. (source)

Method

The authors record and test every distinct repository patch across 5 models, 2 agent frameworks and 3 SWE-bench datasets. Termination lag is the number of steps between the first correct patch and the end of the run. (source)

Results

  • 38–82% of successful runs reach a correct solution on the initial source-code edit. (source)
  • On SWE-bench Verified and Multilingual, median post-solution steps account for 35–37% for Qwen3-Coder and 54–60% for Laguna-XS.2, among successful runs. (source)
  • Most additional steps involve verification and repository exploration. In 1–8% of cases, continued activity undoes an earlier correct patch and leaves an incorrect final solution. (source)

Significance

Interpretation: Eval Harness Configuration should measure trajectories as well as final results. For Agents (LLM Agents), first-correct time and final correctness expose costs and regressions hidden by a single pass rate. (source)

Open Questions

The abstract does not establish a deployable stopping rule: its analysis tests patches retrospectively. Whether an agent can identify the first correct patch without hidden tests, and which verification steps remain necessary, are open questions here. Full experiments and code were not reviewed. (source)

Referenced by

Sources