Unverified50% confidenceFactExact time
Recent research reveals that the refusal behavior of aligned large language models is surprisingly fragile, and simple activation-space or weight-space interventions are enough to break their refusal ability.
1
Sources
50%
Confidence
Long-term
Relevance
10/22/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe success rate of relying on large language model judgment to reject malicious prompt injections is unstable, and scope restrictions serve as a fallback mechanism when model judgment fails.72% similarUnverifiedForcing small language models to approximate deterministic problems probabilistically sacrifices accuracy and energy consumption, and much of the unreliability in edge small models is avoidable.72% similarUnverifiedPrevious research has mostly attributed over-refusal to static representation overlap, where benign prompts and harmful prompts are too close to each other in the model's internal representation space.71% similarUnverifiedWhen fine-tuned on samples of safe refusals, the model learns refusal as a tendency, which may lead it to over-refuse legitimate questions as well.70% similarUnverifiedAnthropic研究发现安全拒绝行为在激活空间中前3个主成分往往能解释超过80%的方差，这是abliteration能以少量操作奏效的根本原因69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1084574API
curl https://kongchang.com/api/v1/knowledge/claims/1084574MCP
get_claim(id=1084574)