Unverified50% confidenceFactExact time
In the test, each statement was asked in isolation in a fresh conversation, while each family was also asked jointly in a single prompt; each model received approximately 160 calls with temperature set to 0.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedEach model was paired with four prompt types—zero-shot, few-shot, chain-of-thought, and few-shot plus chain-of-thought—each in two groups, producing approximately 52,000 labels in total.70% similarUnverifiedWhen asked jointly, the models' no-arbitrage rate was as high as 95–100%, but it dropped sharply to 49–84% when switched to isolated questioning (each statement in a separate conversation).70% similarUnverifiedWhen directly asked whether it had used the answer key, the model denied it in all 47 inquiries and admitted to it zero times across a total of 270 follow-up questions.69% similarUnverifiedIn 16 multi-turn real human conversation tests, 2.0 outperformed the base model comprehensively with a score of 16 to 0.65% similarUnverifiedAnthropic分析100万条真实对话后发现，人类提问水平和AI回答水平的相关性系数达到0.9265% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/999208API
curl https://kongchang.com/api/v1/knowledge/claims/999208MCP
get_claim(id=999208)