Add paper: Evaluating Code Slop in Long-Horizon Coding Agents

[shubhamrgandhi]

Adds our paper to the Surveys {{SECTION}} Empirical Studies section.

Evaluating Code Slop in Long-Horizon Coding Agents
Shubham Gandhi, Shyam Agarwal, Nachiket Kotalwar, Atharva Naik (Carnegie Mellon University)
COLM 2026 Workshop on Agent Behavior. Paper: https://openreview.net/forum?id=VLgFkLRUfV

Coding-agent evaluations usually check whether a task was solved, not what code was left behind for humans to review and maintain. We call unnecessary code volume, control-flow complexity, and maintainability debt beyond task requirements "code slop" and measure it as trajectory-induced deltas in three static-analysis metrics. Across 2,268 runs (42 SWE-EVO tasks, three coding models, three seeds, six cleanup settings), aggregate functional utility stays within a narrow band (70–76%) while code-volume deltas range from −14.3 to −79.3 across cleanup policies. A single final cleanup pass yields little slop reduction, periodic in-trajectory cleanup helps but is costly, and a monitor-triggered policy gives the largest reduction on all three slop metrics at about 25% lower cost than the strongest periodic schedule.

I am one of the authors. The entry follows the format of the surrounding entries; happy to adjust placement or wording.

Placed at the top of Surveys & Empirical Studies as the newest entry and tagged Empirical Study. The paper has no arXiv version, so the Paper badge links to OpenReview and drops the arXiv logo.