Removing the NEEDLE in the Haystack: Backdoor Removal in LLMs via Weight Orthogonalisation
Open paper
Benchmark
Model
Rank
Results
code-generation-on-humaneval
NEEDLE (Qwen3-4B-Instruct-2507, Targeted refusal, Sleeper)
#88
Pass@1: 68.29
sentence-completion-on-hellaswag
NEEDLE (Qwen3-4B-Instruct-2507, Targeted refusal, BadNet)
#102
Accuracy: 71.27
Rank counts only results with a code link.