Unverified50% confidenceFactExact time
The neural expansion framework shifts the focus from parameter space to the geometric structure of input space, proposing that the training process leads to a gradual expansion of "direction-selective ray coverage" in input space, thereby forming stable prediction regions around training samples.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAny framework that uses Gaussian diffusion can gain heavy-tailed exploration capability by adding one scale input, without replacing the training objective or the network architecture.71% similarUnverifiedA model's core parameter weights are fixed after training is complete; even when new information is obtained through external tools, the model still relies on the worldview formed during training when making integrated judgments.70% similarUnverifiedThe core idea of SpatialASM is spatial simulation at training time, without relying on additional tools at inference time70% similarUnverifiedIn this design, the width of the bottleneck is an explicitly settable hyperparameter, rather than being implicitly formed during training as in attention.69% similarUnverifiedSharing the input embedding and output projection layer in the early stages of training benefits stability, while separating them in the later stages benefits expressiveness, constituting a trade-off.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1065864API
curl https://kongchang.com/api/v1/knowledge/claims/1065864MCP
get_claim(id=1065864)