MUDAPIBench: Gradient Difference Is the Best Unlearning Method for Stopping Code LLMs From Emitting Deprecated APIs
arXiv·low signal
MUDAPIBench has 7,000+ model-specific instances from 145 deprecated-to-current API mappings across eight Python libraries, keeping only cases where the model actually emits the deprecated call. Eight unlearning methods were tested on three code LLMs for forgetting, correct replacement, preservation of unrelated APIs and general coding ability. Gradient Difference gave the best overall trade-off.