Unverified50% confidenceFactExact time
Research records from ICLR 2026 show that in real fact-checking tasks, the disagreement rate among leading models is as high as 63%.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen a 61% pass rate is observed on 64 samples, the 95% Wilson confidence interval is approximately 48% to 73%69% similarUnverifiedThe 2023 TruthfulQA benchmark showed that even the strongest models at the time did not exceed 60% accuracy on questions requiring precise factual judgment.69% similarUnverified一个61分的模型和一个59分的模型在多数实际应用中的差异可能远小于分数暗示的程度68% similarUnverifiedOn the internal benchmark CoBench, Opus 5.5 successfully diagnoses real R&D problems approximately 56% of the time, while fully replacing Anthropic researchers would require at least 85%68% similarUnverifiedMcKinsey research shows that more than 60% of business decision-makers admit they are unable to accurately understand the model results delivered by their data science teams.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1027937API
curl https://kongchang.com/api/v1/knowledge/claims/1027937MCP
get_claim(id=1027937)