Unverified50% confidenceFactExact time
The reward modeling stage is usually based on pairwise comparison data, using the Bradley-Terry model to convert human preferences into probability form.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedReward model training can use learning-to-rank frameworks such as the Bradley-Terry model, training a neural network to predict the probability of human preference between any two responses, taking a (prompt, response) pair as input and outputting a scalar score.80% similarUnverifiedBradley-Terry模型是统计学中处理成对比较数据的经典概率模型，假设每个参与者有一个潜在实力参数，获胜概率由双方参数比值决定79% similarUnverifiedThe implicit reward margin magnitude refers to the difference between the reward scores implicitly assigned by the model to a pair of responses in a preference optimization model (such as the Bradley-Terry model), i.e., |r(A)-r(B)|.75% similarUnverifiedDPO uses the log-probability difference between the policy model and the reference model as an implicit reward, enabling the model to increase the generation probability of better responses while staying close to the reference model.73% similarUnverifiedReward guidance during diffusion model sampling is usually implemented by computing the gradient of the reward function with respect to the current sample and adding it, with a weight, to the denoising update, similar to classifier guidance or classifier-free guidance, but with the guidance signal coming from an external reward model.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1131593API
curl https://kongchang.com/api/v1/knowledge/claims/1131593MCP
get_claim(id=1131593)