Unverified50% confidenceFactExact time
The training objective of Diffusion-LM includes three loss terms: the diffusion denoising loss, an embedding matching term, and a rounding loss.
1
Sources
50%
Confidence
Long-term
Relevance
9/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedDuring training, MRL simultaneously optimizes loss functions at multiple dimensional levels, with the total loss being a weighted sum of losses at each granularity.74% similarUnverified低精度引入的数值扰动会在RL的多轮迭代中通过策略梯度的乘积形式被放大，可能导致训练崩溃或模型性能退化66% similarVerified仅用成功轨迹训练的模型往往存在分布外泛化的严重缺陷，遇到工具返回错误时便丧失纠偏能力66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/843369API
curl https://kongchang.com/api/v1/knowledge/claims/843369MCP
get_claim(id=843369)