AI bug-fixing loops break more code than they repair

Language models running as autonomous bug-fixers break more correct code than they repair, according to a paper posted September 9.

The risk sits with anyone running an autonomous repair loop against code that's already correct. The paper ties the damage to a steering vector, an internal signal for editing propensity that fires whether or not a bug exists.

Run the loop long enough and some settle into what the authors call a pseudo-bug-fixing cycle: an edit added, removed, then added back, never converging. Turn the vector up and the model edits more; turn it down, it edits less.

Two more papers back this up. A manual review of LLM program repair found repair hallucinations in 72.7% of cases. A separate study spanning 5 models, 4 benchmarks, and more than 6,000 faulty program instances found fault detection rates near zero on the hardest cases. The tests the models wrote to check their own fixes missed the failing behavior they were supposed to catch.

For anyone running repair loops against production code, a smarter confidence check won't help. Cap it at three attempts, then surface the diff for a human. If the codebase already passes its tests, flip the default and require the loop to prove a failing test exists before it touches anything.

The steering vector points to something worth building on. If editing propensity is a controllable direction, an agent harness could expose it as a dial. A "conservative mode" setting would then mean something mechanical, not a prompt asking the model to be careful.