Logit Lens

Unsupervised interpretability technique that projects activations through unembedding matrix; provides comparison point for NLA approach.

Neighborhood — ranked by edge-count

paper

thinker

nostalgebraist
introduces
Author of LessWrong post deriding GPT-3 evaluations, advocate of ecological evaluation.

finding

Pearson-Vogel et al.: accurate self-description prompts increase introspective detection from 0.3% to 39.9%
supports
Cited to mechanistically support why the contemplative prompt changes what post-training-shaped final layers allow through

concept

Residual Stream
about
Proposed pathway flowing through layers at each position; calculates K/V values that feed horizontal information flow.
predictive integration
implements
The mid-to-late layer computational process that converts routed perturbation signals into explicit predictions

method

Natural Language Autoencoders (NLAs)
associated_with
Core unsupervised method for generating natural language explanations of LLM activations through a verbalizer-reconstructor pair trained with RL.

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Logit-based self-reportmethod0.747
Primary self-report measure: probability-weighted expected value over all ten digit-token logits, yielding a continuous rating that preserves full distributional signal
Logit Weight Analysismethod0.729
Computing each feature's linear effect on output token logits via path expansion through MLP output weights and unembedding matrix
Cognitive Lensframework0.723
Logit Bias Constraintmethod0.718
Used with GPT models to constrain responses to binary options (0/1) in belief coherence experiments.
Logit Weight Similaritymethod0.704
Correlating logit weight vectors between features from different models as a measure of downstream-effect universality
global logit shiftconcept0.693
The methodological confound identified by this paper: injection biases model toward 'YES' for any binary question regardless of content
logarithm transformationmethod0.685
Parameter-free loss transformation applied to each task loss to equalize scales
logistic fitting for shot thresholdsmethod0.661
Fit a sigmoid to accuracy vs. k to estimate k50 and phase width.