Unverified50% confidenceBenchmarkExact time
Distilling 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).
1
Sources
50%
Confidence
Long-term
Relevance
9/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAdding 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).77% similarUnverifiedAlmost all useful optimization value is concentrated in the reflection/control slot, whose leave-one-in gain is as high as +0.119, while the independent contributions of the other three slots are nearly zero.66% 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.64% similarUnverifiedThe Mann-Whitney U test shows that task_score under the naive survival reward falls below the random-policy baseline, with p=0.012.64% similarUnverifiedIn games with extremely sparse rewards such as Montezuma's Revenge, standard DQN scores near zero, and breakthroughs were only achieved after the introduction of advanced exploration strategies.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/953840API
curl https://kongchang.com/api/v1/knowledge/claims/953840MCP
get_claim(id=953840)