Unverified50% confidenceFactExact time
Subspace locking refers to the phenomenon in which the model encodes the capabilities of different domains into relatively fixed and separated subspaces in the representation space during training; once formed early on, it is difficult for subsequent training to reallocate or share them.
1
Sources
50%
Confidence
Long-term
Relevance
11/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedPretrained classifiers are typically designed for fixed category systems, and when facing entirely new domains or highly specialized tasks may require fine-tuning or even retraining, and cannot replace large models in complex semantic understanding and generation.73% similarVerifiedExternal scheduling at inference time and internally learned capabilities during training differ in generalization performance and controllability.73% similarUnverifiedThe 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.72% similarUnverifiedThe zero-gating design means the convolutional path is nearly inactive early in training and is gradually released as training progresses, avoiding conflicts between local priors and global attention in early training.72% similarUnverifiedIn distributed training, a small number of nodes whose kernel parameters are inconsistent with the rest can drag down the throughput of the entire training job, creating bottleneck nodes.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1165105API
curl https://kongchang.com/api/v1/knowledge/claims/1165105MCP
get_claim(id=1165105)