Unverified50% confidenceFactExact time
The mean-field likelihood estimation relied upon by existing RL methods for diffusion models is biased and has high variance.
1
Sources
50%
Confidence
Long-term
Relevance
9/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe average evaluation metrics of a new LLM may be better than those of an older model, but regressions can occur in niche critical scenarios.72% similarUnverifiedIn 2023, Gao et al. from OpenAI published an empirical study on reward model overoptimization, finding that as the number of RL optimization steps increases, model performance on real human evaluations follows an inverted-U curve.66% similarUnverifiedThe remaining 8.5% deviation is amplified when involving billions of samples, and minor systematic biases may cause simulation conclusions to diverge significantly from reality; therefore, such simulations are better suited for trend prediction and relative comparison rather than absolutely precise numerical forecasts.65% similarUnverified靶向数据工程中，针对性数据占总微调数据的 5%-15% 通常能取得良好效果，过高比例可能导致模型在其他维度退化65% similarVerifiedStandard diffusion models often require 20–50 steps to generate.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/927924API
curl https://kongchang.com/api/v1/knowledge/claims/927924MCP
get_claim(id=927924)