Unverified50% confidenceConceptExact time
Process reward scores intermediate steps within the multi-step trajectories an agent executes, providing a denser signal that alleviates the credit assignment problem in long-horizon tasks; outcome reward is given only at the end of a task based on the final state, with a sparse signal but a clear objective.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedReward engineering is essentially a highly iterative process: design the reward, observe behavior, discover loopholes, correct the reward, and repeat.78% similarVerifiedReward shaping refers to adding extra, denser guiding reward signals on top of the original sparse reward to help agents learn faster and more correctly.76% similarUnverifiedCheckpoint partial rewards densify sparse signals by setting milestone-style rewards at intermediate nodes in long tasks, helping gradient propagation.75% similarUnverified当奖励函数未对控制代价加以惩罚时，梯度会持续将策略推向动作边界，产生 Bang-Bang 行为，这被称为奖励作弊（Reward Hacking）70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1101021API
curl https://kongchang.com/api/v1/knowledge/claims/1101021MCP
get_claim(id=1101021)