Unverified50% confidenceFactExact time
The theoretical analysis of WAS-PO shows that even if all reward weights are positive, the effective weight of the whitened aggregate objective on certain original rewards may still be negative.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe paper gives an exact two-reward condition under which the reward with the lighter weight is inverted in the equivalent objective.70% similarUnverifiedWhen the within-group reward distribution is severely skewed, the mean is pulled off by extreme values, causing distorted advantage estimates; in such cases, the median or a trimmed mean may be more robust.67% similarUnverifiedThere is an unverified risk with calibration rewards: they may simply flatten the confidence of all outputs toward a middle value, performing well on the ECE metric while losing discriminative information.66% similarUnverifiedUnder three rewards, WAS-PO's held-out accuracy is slightly lower than GDPO's (not significant); under four rewards it is absolutely lower by 0.03 (only two seeds were evaluated).65% similarUnverifiedIn the synthetic experiment, "always output RISK_LEVEL: 1" achieves an ordinary accuracy as high as 0.922, but an average reward of −0.000, showing that class-balanced rewards can prevent laziness.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130532API
curl https://kongchang.com/api/v1/knowledge/claims/1130532MCP
get_claim(id=1130532)