Unverified50% confidenceFactExact time
The Jacobian lens compresses the matrix of partial derivatives of an intermediate-layer hidden state with respect to the final output logits into low-dimensional coordinates indexed by vocabulary, decoupling the communication signal from model width.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAdvantage-sign-aware Renyi preprocessing is used to counteract saturation at the logit level, enabling gradient signals to propagate effectively in high-confidence regions and penalizing confident but incorrect predictions.67% similarUnverifiedExperiments found that in both models, the final stage of the encoder became the critical boundary for grounding: bypassing the last module caused output divergence on almost every utterance, while bypassing intermediate modules had almost no effect.66% similarUnverifiedRNN通过隐藏状态在时间步上串行传递信息，导致梯度消失/爆炸和无法并行计算两个核心瓶颈66% similarUnverifiedThe Token Entanglement Hypothesis holds that the root cause of the unconscious number phenomenon lies in tokens of different semantic categories in the model's output layer (Unembedding Matrix) being close to each other in high-dimensional vector space, causing activation to leak into adjacent categories.66% similarUnverifiedThe text encoder sits at the very front of the generation pipeline, and its output errors are progressively amplified through the diffusion model's denoising steps and cross-attention mechanisms (cascading errors), making precision loss from encoder quantization more critical than that of the backbone network.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1098422API
curl https://kongchang.com/api/v1/knowledge/claims/1098422MCP
get_claim(id=1098422)