Unverified50% confidenceBenchmarkExact time
Adding 1 layer of lookahead search on top of the policy network improved the win rate from 0.625 to 0.944 (160 paired matches, 95% confidence interval +0.24 to +0.40).
1
Sources
50%
Confidence
Long-term
Relevance
9/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedDistilling the search gains back into the policy network via a distributional value target combined with the DAgger method yielded a win rate improvement of +0.045 (1,600 matches, p = 0.007).77% similarUnverifiedIn the fact-updating task, the success rate of the embedding retrieval method fluctuated between 0.30 and 0.95, with a maximum gap of 0.45 across different random seeds.69% similarUnverified模板匹配的置信度阈值通常取0.8～0.95来判断是否找到目标66% similarUnverified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%66% similarUnverifiedFor categorical policies, as the output distribution becomes increasingly sharp, the expected score-gradient norm approaches zero, resulting in the weakest updates for tokens that are confident but incorrect.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/953839API
curl https://kongchang.com/api/v1/knowledge/claims/953839MCP
get_claim(id=953839)