AI
Removing the NEEDLE in the Haystack: Backdoor Removal in LLMs via Weight Orthogonalisation
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 03/10/2026, 04:42:48 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 10:18:24 UTC.
Briefings
A 2026 study introduces NEEDLE, a training-free backdoor removal method for Large Language Models. By estimating backdoor directions via activation vectors and applying sequential weight orthogonalisation, the approach suppresses malicious behavior without altering refusal representations. Evaluated across multiple model families, NEEDLE achieved the lowest mean Attack Success Rate (ASR) at 0% on challenging code injection attacks, outperforming existing defenses in both attack success and safety metrics.
Key points
- NEEDLE is a training-free method that uses activation vectors to estimate backdoor directions and applies sequential weight orthogonalisation to suppress backdoor behavior.
- The method achieves a 0% Attack Success Rate on challenging code injection attacks, representing the lowest ASR among evaluated defenses.
Context and limitations
Evaluation was conducted across multiple model families and attack types, but specific model names and attack type details were not provided in the evidence.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/10/2026, 04:42:48 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiacqvm4z563abr64vtrazhlz6utfrt3nygvjddga4uqj233dxq4pq