other

Softmax bottleneck

Property of a language model whose next-token distribution comes from projecting a lower-dimensional hidden state through a softmax layer, which caps the rank of the logits it can produce at the embedding size.

Known aliases

low-rank output layer

Relationships

No evidence-backed relationships are recorded.

Current clusters

build1 publisher

The softmax bottleneck leaks a model's hidden size to anyone who can read its logits

A paper on arXiv put under $1,000 of queries through gpt-3.5-turbo and estimated its embedding size at about 4,096, then reused the same output basis as a fingerprint sensitive to weight changes.

Publishers:arxiv.org

Reality

Evidence66
Adoption: Insufficient
Hype gap+15
Incentives40
Confidence58