Unverified50% confidenceFactExact time
On non-math tasks, the correlation between perplexity and per-question accuracy was weak (r between -0.15 and 0.08), indistinguishable from zero.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn the P-words task, the model achieves an R² of 0.9 and a normalized root mean square error (NRMSE) as low as 0.05.71% similarUnverifiedUnder the any-of rule (passing if any run passes), at p=0.10 the error rate decreases geometrically with N: 0.0010 for N=3 and 1.000e-10 for N=10.68% similarUnverified困惑度是测试集上每个token平均负对数似然的指数，值越低代表模型对语料的预测越准确67% similarUnverified对比学习中的推远操作只要求负样本相似度低于正样本，并不要求达到-1的极端值66% 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.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1018313API
curl https://kongchang.com/api/v1/knowledge/claims/1018313MCP
get_claim(id=1018313)