Sparse Readout Prism: Explaining Logit-Lens Scores in Features Instead of Tokens
SRP reframes logit-lens explanations around sparse readout features, not token labels.
The paper argues that lens outputs can change when the lens is fit on a different corpus, even for the same hidden states. SRP avoids that dependency by decomposing the readout weights directly and attributing token logits or logit differences to sparse features. In tests, its sparse approximation reconstructs 8.9 to 17.3 percentage points more of the measured logit differences than the strongest of six geometry-based baselines. The authors also report that dominant readout features stay stable even when token readings vary by fitting corpus. HF Daily Papers' note
The paper argues that lens outputs can change when the lens is fit on a different corpus, even for the same hidden states. SRP avoids that dependency by decomposing the readout weights directly and attributing token logits or logit differences to sparse features. In tests, its sparse approximation reconstructs 8.9 to 17.3 percentage points more of the measured logit differences than the strongest of six geometry-based baselines. The authors also report that dominant readout features stay stable even when token readings vary by fitting corpus. HF Daily Papers' note
score 4