Unverified50% confidenceFactExact time
The study observes that weight norms exhibit a monotonically compressing trajectory during the grokking transition.
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe theoretical guarantees of Decoupled Descent rest on strong assumptions including stylized Gaussian mixture data, full-batch gradient descent, and high-dimensional asymptotic limits, and there is still a long way to go before it can be applied to large-scale models.64% similarUnverified在长时间自主运行场景中，模型可能通过上下文积累逐渐偏离安全边界，这被研究者称为'对齐漂移'（alignment drift）64% similarUnverifiedThe researcher observed three divergence patterns: differences being absorbed and then converging, differences persistently leading to different outcomes, and divergence—reconvergence—divergence again.63% similarUnverifiedAcceleration solutions such as consistency models and flow matching are narrowing the computational overhead gap of multi-step inference in diffusion models.63% similarUnverifiedThe double descent phenomenon occurs not only along the model size dimension, but can also appear with changes in training epochs and data volume.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/916335API
curl https://kongchang.com/api/v1/knowledge/claims/916335MCP
get_claim(id=916335)