Unverified50% confidenceFactExact time
Existing @k optimization methods are all bespoke specialized algorithms, and there is no general framework that can transfer mature @1 policy optimization methods to the @k setting.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/3/2026
First Seen
Valid until: 2/1/2027
Sources
Related Entities
Related Claims
UnverifiedThe researchers propose @k Adaptive Pressure for Policy Optimization (KAPPO), whose workflow includes a one-time reference calibration, after which existing @1 methods are adapted to @k through prompt-adaptive regularization or reward transformation.69% similarUnverifiedKL regularization refers to adding a KL divergence penalty term against the reference policy to the optimization objective, usually in the form -β·KL(π||π_ref), to constrain the new policy from deviating too far from the initial model.63% similarUnverifiedMoE通过Top-K路由策略每次前向传播仅激活K个专家（通常为2-4个），实现稀疏激活以扩大模型容量而不线性增加计算量60% similarVerifiedThe MoE architecture uses a router to dynamically select the Top-K experts for computation for each token, allowing the total parameter count to scale while keeping the number of activated parameters per inference constant.56% similarVerifiedThe MoE architecture uses a gating network to dynamically select the K most relevant experts (typically Top-K, K=2) for each input token, activating only approximately 10%–30% of total parameters during inference.55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130867API
curl https://kongchang.com/api/v1/knowledge/claims/1130867MCP
get_claim(id=1130867)