Unverified50% confidenceBenchmarkExact time
The red-team evaluation tested adversarial instructions hidden in the case facts on five core models, and the models' compliance rate with such instructions reached as high as 73.3%–96.4%, far exceeding the decisions' sensitivity to replacement of the legal basis.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn Track B, the flat RAG baseline can detect 76%–97% of true contradictions, but produces false positives on 16%–43% of surface-similar safe drafts (the specific ratio depends on the backend).69% similarUnverifiedDuring testing, cases emerged where a model claimed 95% confidence in a move but scored 0 points according to the evaluation criteria.68% similarUnverifiedTraditional red-teaming assumes that system behavior is comparable across different scales, whereas the ecological perspective denies this assumption.68% similarUnverifiedA rule was set in the experiment: any judgment with a confidence level below 60% would be directly handed over to human processing.68% 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 scenarios68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1013279API
curl https://kongchang.com/api/v1/knowledge/claims/1013279MCP
get_claim(id=1013279)