Unverified50% confidenceBenchmarkExact time
At a training scale approaching 100 billion tokens, the difference in final smoothed training loss between the polynomial version and the native implementation falls between -0.107 and +0.079.
1
Sources
50%
Confidence
Long-term
Relevance
10/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe pre-training validation loss across the three generations of models shows a steady downward trend: 2.87 → 2.78 → 2.59.77% similarUnverifiedThe trained model achieves a per-token negative log-likelihood (NLL) of 0.93, compared to 12.60 for a randomly initialized twin model of the same architecture.74% similarUnverifiedThe power-law exponent between test loss and compute is approximately 0.05, meaning that every 10x increase in compute yields approximately a 12% improvement in model performance.73% similarUnverifiedThroughout the entire training process, the flow-matching loss only slowly decreased from 0.805 to 0.754.72% similarUnverified即使只有5-10%的训练样本存在错位，也会导致VLM模型在下游任务上出现显著的性能退化70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/971820API
curl https://kongchang.com/api/v1/knowledge/claims/971820MCP
get_claim(id=971820)