Removing the NEEDLE in the Haystack: Backdoor Removal in LLMs via Weight Orthogonalisation

Benchmark Model Rank Results
code-generation-on-humanevalNEEDLE (Qwen3-4B-Instruct-2507, Targeted refusal, Sleeper)#88Pass@1: 68.29
sentence-completion-on-hellaswagNEEDLE (Qwen3-4B-Instruct-2507, Targeted refusal, BadNet)#102Accuracy: 71.27