Gathered, Not Admitted: How Attention Brings a Latent Variable into Verbalizable Form
The paper argues that attention, not a workspace gate, makes a latent variable readable to the model.
Mazaheri tests open-weight models with Jacobian lenses and reports no gate at the predicted point. The readable form appears through attention-mediated gathering in a mid-depth window, with demand for flexible reuse raising the variable’s visibility. The paper says that window is demand-specific and shows similar fractional depth across two model families, while warning that the readout is not a calibrated measure of actual use. HF Daily Papers' note
Mazaheri tests open-weight models with Jacobian lenses and reports no gate at the predicted point. The readable form appears through attention-mediated gathering in a mid-depth window, with demand for flexible reuse raising the variable’s visibility. The paper says that window is demand-specific and shows similar fractional depth across two model families, while warning that the readout is not a calibrated measure of actual use. HF Daily Papers' note
score 4