Unverified50% confidenceBenchmarkExact time
Experiments show that models processed with abliteration exhibit almost no degradation on standard capability evaluations, but the probability of refusing harmful requests drops significantly.
1
Sources
50%
Confidence
Long-term
Relevance
9/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAnthropic's research shows that excessive alignment may cause a model's refusal rate on harmless questions to rise from under 1% to over 10%.75% similarVerifiedIn a 2024 research paper, Anthropic quantified the phenomenon of model over-refusal, finding that safety-trained models had false positive refusal rates as high as 15–30% in certain scenarios75% similarUnverifiedAnthropic的研究显示，即使面对毫无根据的质疑，在某些测试中模型被用户简单反驳后放弃正确答案的比例超过50%75% similarUnverifiedIn open-ended generation tasks, the study observed occasional accuracy drops, which were more pronounced in non-English responses.74% similarUnverifiedResearch shows that a model's accuracy on self-criticism tasks is significantly lower than in scenarios where criticism is performed by another model.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/922736API
curl https://kongchang.com/api/v1/knowledge/claims/922736MCP
get_claim(id=922736)