Unverified50% confidenceFactExact time
Increasing the learning rate to compensate for the smaller average step size failed to recover the rare-fact learning cost caused by quantization noise or by NVFP4 itself.
1
Sources
50%
Confidence
Long-term
Relevance
11/10/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn the control experiment where the NVFP4 error is doubled, noise injected only into the Adam second-moment estimate behaves like a reduced learning rate, and its cost gradually fades over training; noise injected into the update reduces the signal-to-noise ratio, carries about two-thirds of the cost, and persists until the end of training.79% similarUnverifiedIn the sampled-gradient setting, when the learning rate η decreases, the expected gain under fixed β vanishes faster than the fluctuation itself, causing the benefit of momentum to be swallowed by noise.74% similarVerifiedIf samples are too easy the model has nothing to learn; if they are too difficult, noisy learning signals may interfere with training; the most effective learning occurs in the region slightly above current capability but still within reach.73% similarUnverified数据规模的增长并不等同于有效学习信号的增长,在模型能力提升到一定阈值后简单堆砌样本的收益会急剧下降72% similarUnverifiedWhen out-of-distribution (OOD) failure occurs, learned selectors may persistently ignore key evidence, resulting in answers that appear fluent but omit key facts, increasing the difficulty of troubleshooting.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1154472API
curl https://kongchang.com/api/v1/knowledge/claims/1154472MCP
get_claim(id=1154472)