New Paper Formalizes 'CodeSlop' — and Cuts It 17.9–32.9% by Minimizing the Agent's Trajectory, Not the Diff

arXiv (Google DeepMind / Columbia authors)·medium signal

TRIM (arXiv:2607.18161, submitted July 20) from Mathai, Iyer, Nogikh, Maniatis, Ivancic, Yang and Ray defines CodeSlop as functionally unnecessary residual edits — speculative changes, abandoned hypotheses, and temp code that coding agents accumulate while iterating. Instead of attacking redundancy in the final diff, TRIM minimizes the agent's search trajectory, cutting CodeSlop 17.9–32.9% across agent frameworks with negligible task regression and roughly half the validation cost of Delta Debugging. For anyone reviewing agent-authored PRs, this reframes bloat as an artifact of the search process rather than the generator — meaning the fix belongs in the harness, not the linter.