If you've ever used a coding agent to fix a bug or build a feature, you may have noticed something odd: the diff it produces is often huge. Where a human might change 10 lines, the agent changes 80 — with extra helper functions, defensive try/except blocks, logging statements, and comments explaining what the code obviously does. The tests pass, but the pull request looks bloated. The authors of TRIM call this "CodeSlop", and this paper is about where it comes from and how to get rid of it.

Their key insight is that CodeSlop isn't really a problem with how the agent writes code — it's a problem with how it searches for a solution. When a coding agent is grinding toward a passing test, it makes lots of speculative edits along the way: it adds a helper it thinks it might need, tries a defensive check, guesses at an edge case, changes its mind, tries something else. When a test finally passes, the agent stops. But all those speculative additions from earlier in the trajectory are still sitting in the final diff, even though many of them aren't actually necessary for the fix.

TRIM's approach is refreshingly direct: after the agent finishes, take the final diff and try to remove pieces of it. For each addition, ask, "Do the tests still pass without this?" If yes, drop it. If no, keep it. It's essentially delta debugging applied to agent output — the same technique used to minimize failing test cases, but pointed at the reverse problem of minimizing succeeding ones.

The results, per the abstract, show meaningful reductions in diff size while preserving correctness. A few things this unlocks:

• Reviewable PRs. A 15-line diff is something a human can actually read and reason about; an 80-line diff often gets rubber-stamped or rejected outright.
• Less technical debt. Speculative helpers and unused abstractions become future maintenance burden even if they don't break anything today.
• A cleaner signal about what the agent actually understood. When you strip away the search residue, you see the real fix — and whether it makes sense.

The broader lesson is that agent outputs shouldn't be treated as final products. There's a post-processing step available — verify-then-shrink — that most agent frameworks skip because they stop as soon as the green checkmark appears. TRIM argues that's exactly the wrong stopping point.