Unverified50% confidenceEventExact time
Among the three DPO attempts, one reached a reward margin of 1.0 during training, but its free-generation outputs showed no change at all.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedPenalizing automaton-rejected tokens during training can improve development-set Smatch++ by 3.6 to 11.0 points after 10% of training is completed, but this advantage gradually diminishes as training progresses67% similarUnverifiedWhen the margin reward is positive, it indicates that the tool indeed brings gains and the question is a valuable training sample; when the difference is zero or negative, the question's training value is limited.66% similarUnverifiedSMD's repulsion mechanism does not require the discriminator commonly used in adversarial training, nor does it require online score estimation; it is merely an additional constraint on the training objective, and the training overhead is almost zero.66% similarVerifiedAlignment algorithms such as DPO (Direct Preference Optimization) reduce training costs while improving the consistency of model performance on open-ended instructions.64% similarUnverifiedSingle-pass training brings only modest overall calibration improvement and no improvement whatsoever in the early prefix stage.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1032788API
curl https://kongchang.com/api/v1/knowledge/claims/1032788MCP
get_claim(id=1032788)