Unverified50% confidenceFactExact time
Greedy decoding selects the highest-scoring token at each step; sampling selects among candidate tokens according to probability.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
Unverified在歧义点的四类行为中，'先搜再问'成功率最高，平均通过率达93.4%，远高于'直接猜'的56.5%，'搜多次再猜'为51.5%63% similarUnverifiedThe winner determination logic is to first sort by average score from highest to lowest; if average scores are tied, completion rate is compared; if still tied, the one with fewer steps wins.61% similarUnverifiedThe best-of-8 selection strategy refers to generating 8 candidate outputs for the same prompt and then picking the best, which can significantly improve evaluation scores but roughly increases inference cost by 8 times; comparing a system using this strategy with a single-sample system is not entirely fair.61% similarUnverifiedIn the experiment, when the same model acted as the deceiver, its ranking was almost identical to when it acted as the defender: mistral was the most persuasive liar (72% success rate), while gemma2 was both the hardest to deceive and the worst at lying.61% similarUnverified最优阈值可通过最大化Youden指数（灵敏度+特异度-1）确定，在假阳性与假阴性代价对等时是最优阈值选择准则60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1067375API
curl https://kongchang.com/api/v1/knowledge/claims/1067375MCP
get_claim(id=1067375)