Oracle Lens
A multi-token activation-reading method that generates bullet-point-form output describing activation contents, with an RL-finetuned variant.
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.
Oracle Lens pursues more faithful and more concise readouts through autoregressive reconstruction plus RL optimization.
All Facts (2)
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%UnverifiedOracle Lens pursues more faithful and more concise readouts through autoregressive reconstruction plus RL optimization.
50%