Unverified50% confidenceFactExact time
Martini's goal is: given only limited access to the oracle reward r* through a small feedback set D, to approximate the tilted distribution q* of r* using a combination function G of multiple cheap surrogate rewards.
1
Sources
50%
Confidence
Long-term
Relevance
11/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedPolicy gradient methods (such as PPO) rely on a discount factor γ to exponentially decay future rewards, with a reward t steps away from the current moment retaining only a weight of γ^t.63% similarUnverifiedThe hybrid reward idea is phased scheduling: tolerate sparsity when information is insufficient, introduce calibrated dense signals in trap-prone regions, and return to true feedback as the endgame approaches.62% similarUnverified研究者提出长度归一化奖励方法,即在计算奖励分数时除以回答长度,使模型无法通过单纯增加长度获取更高奖励60% similarUnverifiedStrategies such as Monte Carlo Tree Search that rely on executed rewards to rank nodes can only make decisions based on very few and noisy scores under a limited compute budget.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1127427API
curl https://kongchang.com/api/v1/knowledge/claims/1127427MCP
get_claim(id=1127427)