Unverified50% confidenceFactExact time
On a single GPU, the researchers compared an uninterrupted training run with a run terminated midway and then resumed; after both completed 8 optimizer updates, the trainable parameters and training state were bit-for-bit identical, and the reload probabilities on 30 check questions were exactly the same.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/18/2026
First Seen
Valid until: 1/16/2027
Sources
Related Entities
Related Claims
UnverifiedBetween NPU training and GPU training, the loss trajectory correlation coefficient reaches 0.99, which measures the similarity of training dynamics and does not mean throughput parity or that speed can catch up with GPUs.72% similarUnverifiedIn controlled single-run pre-training comparison experiments, the model trained on the Edu-QuRating-curated mixture achieved higher overall observed accuracy across nine benchmarks than the FineWeb-Edu baseline model, with performance gains concentrated on specific tasks.71% similarUnverified批次归一化在训练模式使用当前批次统计量，在推理模式使用训练过程中累积的移动平均值70% similarUnverifiedThe reason for using a correlation coefficient rather than simply comparing final loss values is that it can capture the dynamic consistency across the entire training run, which proves the numerical validity of the training process better than a single endpoint metric.70% similarUnverifiedGrokking refers to the phenomenon where, after a model has already achieved near-perfect fit on the training set, test set performance remains at a random level for a long time before generalization ability suddenly improves after a large number of additional training steps.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1064022API
curl https://kongchang.com/api/v1/knowledge/claims/1064022MCP
get_claim(id=1064022)