Unverified50% confidenceFactExact time
Action chunk value learning lets the policy predict a sequence of actions within a time window (action chunk) all at once, and estimate the value of this chunk as a whole, rather than outputting single-step actions frame by frame.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedReinforcement learning attempts to learn a general state-to-action mapping policy, whereas sequence search directly seeks a complete sequence of actions in the action space to reach the goal state from the initial state.75% similarUnverifiedIn multi-hop reasoning scenarios, each hop's retrieval should incrementally supplement based on the conclusions of the previous hop, and the decision model can explicitly track the known information set to avoid redundant recall.71% similarUnverified将推理轨迹与行动步骤交替结合，比单纯的行动序列或单纯的推理链取得更好效果70% similarUnverifiedPerforming task-specific reinforcement learning on a small 4B-parameter model improved its mapping reasoning accuracy by 15.7 percentage points.70% similarUnverifiedValerant leverages a pre-trained action-conditioned world model to predict subsequent visual frames based on given actions.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1041792API
curl https://kongchang.com/api/v1/knowledge/claims/1041792MCP
get_claim(id=1041792)