Unverified50% confidenceFactExact time
Standard DPO and its unlearning variants are typically built around noise prediction error and do not transfer well to FSD models.
1
Sources
50%
Confidence
Long-term
Relevance
10/9/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedThe generalization error of a model equals the squared bias plus variance plus irreducible noise.69% similarUnverifiedWhen unlearning objectives designed for multi-step models are applied to few-step distilled (FSD) models, the unlearning effectiveness is significantly diminished.68% similarUnverifiedEvaluating diffusion models should rely on downstream metrics such as FID, FD-DINOv2, and object accuracy rather than looking solely at flow-matching loss, because at high-noise stages the loss is dominated by the irreducible variance of the velocity target.68% similarUnverifiedUnder off-policy sampling conditions, when the preference margin is aligned in direction with the utility gradient, DPO's update shifts from a corrective signal to a reinforcement-style update similar to reinforcement learning.67% similarVerifiedNoise information in the context degrades model reasoning accuracy in a nonlinear manner66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/990738API
curl https://kongchang.com/api/v1/knowledge/claims/990738MCP
get_claim(id=990738)