Tuned lens
An improved method over the logit lens that trains a separate affine transformation for each layer to map representations into a shared space near the final layer before decoding, making readouts from each layer more stable.
Timeline (last 90 days)
The activation-reading tools evaluated are divided into single-token readers and multi-token readers: the single-token side includes Logit lens, J-lens, R-lens, and Tuned lens; the multi-token side includes Natural Language Autoencoders, Sparse Autoencoders, Patchscopes, Template lens, and Oracle Lens.
Tuned lens separately trains an affine transformation for each layer, maps that layer's representation into a shared space close to the final layer, and then decodes it, thereby alleviating the problem of mismatch between the intermediate-layer distribution and the output layer.
Single-token readers cannot cover important workspace content because they are bag-of-words.
All Facts (3)
The activation-reading tools evaluated are divided into single-token readers and multi-token readers: the single-token side includes Logit lens, J-lens, R-lens, and Tuned lens; the multi-token side includes Natural Language Autoencoders, Sparse Autoencoders, Patchscopes, Template lens, and Oracle Lens.
50%UnverifiedTuned lens separately trains an affine transformation for each layer, maps that layer's representation into a shared space close to the final layer, and then decodes it, thereby alleviating the problem of mismatch between the intermediate-layer distribution and the output layer.
50%UnverifiedSingle-token readers cannot cover important workspace content because they are bag-of-words.
50%