Unverified50% confidenceBenchmarkExact time
Under 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).
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/3/2026
First Seen
Valid until: 2/1/2027
Sources
Related Entities
Related Claims
UnverifiedOf the 11 LoRAs, only 3 produced meaningful effects, and 4 scored lower than the no-LoRA control group66% similarUnverifiedThe 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.65% similarUnverifiedAfter excluding the three collapsed scorers, the compliance rates on the four remaining corpora still differ by 1.7 to 2.8 times, and the non-refusal rates differ by 10 to 22 percentage points.62% similarUnverifiedWhen only one of the four detectors was poisoned, the naive accuracy of the unverified LLM investigator dropped to 0.50-0.72; when three were poisoned, accuracy fell to 0.13 or below.62% similarUnverifiedIf a difficult problem fails across all samples (pass@k ≈ 0), the reward variance is zero, the gradient signal nearly vanishes, and the parameter update approaches zero.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130538API
curl https://kongchang.com/api/v1/knowledge/claims/1130538MCP
get_claim(id=1130538)