TRIM: Reducing AI-Generated CODESLOP via Agent Trajectory Minimization

Coding agents often produce patches that pass tests but contain excessive, redundant, or unnecessary code—a phenomenon the authors term "CodeSlop." The research asks whether it is possible to systematically identify and remove these residual search artifacts without compromising the patch's functionality, and whether the agent's own internal search history can be leveraged to do this more efficiently than standard black-box minimization techniques.

The authors define CodeSlop as functionally unnecessary edits that persist in a final patch. To address this, they introduce TRIM (Trajectory-guided Redundancy Identification and Minimization). Unlike traditional methods like Delta Debugging, which treat a patch as an unstructured collection of changes, TRIM uses the agent's "repair trajectory"—the sequence of edits and validation steps taken during the search process—to guide the minimization. TRIM performs a hierarchical counterfactual search, testing whether groups of edits from specific stages of the trajectory can be removed while maintaining the patch's passing status. By pruning large, exploratory groups of edits first, the algorithm significantly reduces the search space.

TRIM demonstrates that agent-generated patches contain substantial amounts of redundant code that can be safely removed. Across four different agent scaffolds (including SWE-AGENT and OpenHands), TRIM achieved a 17.9% to 32.9% reduction in CodeSlop. Furthermore, by exploiting the structure of the repair trajectory, TRIM proved to be highly efficient, requiring roughly half the validation cost of baseline minimization algorithms while maintaining the correctness of the final repair.