J-lens
An activation reading method based on the Jacobian matrix that constructs a readout direction from the Jacobian of the output with respect to the activations of a given layer, capturing the influence of activation changes on the output distribution and ultimately compressing it into a single-token representation.
Timeline (last 90 days)
WorkspaceBench also provides a dedicated hallucination evaluation, and uses "precision/recall relative to J-lens" as an auxiliary metric.
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.
J-lens, R-lens, and similar linear-probe readout methods are ultimately all compressed into expression on a single token, and therefore cannot represent concepts that require multiple words to state clearly.
In terms of readout method, J/R/Logit-lens takes top-10 tokens every 4 layers from layers 20 to 60 of the 64-layer Qwen-3.6-27B.
Single-token readers cannot cover important workspace content because they are bag-of-words.
All Facts (5)
WorkspaceBench also provides a dedicated hallucination evaluation, and uses "precision/recall relative to J-lens" as an auxiliary metric.
50%UnverifiedThe 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%UnverifiedJ-lens, R-lens, and similar linear-probe readout methods are ultimately all compressed into expression on a single token, and therefore cannot represent concepts that require multiple words to state clearly.
50%UnverifiedIn terms of readout method, J/R/Logit-lens takes top-10 tokens every 4 layers from layers 20 to 60 of the 64-layer Qwen-3.6-27B.
50%UnverifiedSingle-token readers cannot cover important workspace content because they are bag-of-words.
50%