The tuned lens is a calibrated method for reading intermediate transformer representations through the model’s output vocabulary. Instead of assuming every layer is already expressed in the final output basis, it learns a small translator for each layer before applying the model’s final unembedding.
In transformer interpretability, tuned-lens analysis addresses a major weakness of the ordinary logit lens: intermediate residual states can contain useful predictive information while living in layer-specific coordinate systems that the final unembedding reads poorly. A learned translator can expose that information without retraining the base model.
This Cognitive Art owner explains tuned-lens translators, frozen base models, layerwise training, cross-entropy objectives, calibration, prediction trajectories, basis mismatch, probe capacity, leakage, model comparison, causal limits and the relationship to linear probing and the ordinary logit lens.
1. Why calibration is needed
The ordinary logit lens applies the final readout at every depth. Intermediate layers can use shifted or rotated coordinates, making fixed readout underestimate predictive information.
2. The translator is deliberately small
A constrained affine or linear transformation is learned for each layer while the base model stays frozen. The aim is correction, not a new deep decoder.
3. Training target
The translator is optimized so translated intermediate states produce vocabulary predictions closer to the model’s predictive target or final distribution.
4. Freeze the base model
Changing transformer weights during lens training would stop the experiment from measuring the original representation.
5. Layerwise prediction trajectories
Calibrated distributions can show when uncertainty collapses, alternatives disappear or the eventual answer becomes stable.
6. Calibration is not natural use
A learned translator proves a simple mapping can recover output-relevant information. It does not prove downstream transformer blocks literally implement that translator.
7. Probe capacity matters
More expressive translators decode more. Report architecture, parameter count, regularization and held-out performance.
8. Train-test separation
Evaluate on sequences not used to fit translators. Entity and document overlap can inflate apparent generalization.
9. Compare with the ordinary logit lens
The improvement from tuned over untuned readout estimates how much fixed-basis mismatch mattered.
10. Compare with linear probes
Task-specific probes target one declared variable; the tuned lens targets vocabulary prediction. They answer different reader jobs.
11. Prediction similarity across layers
Cross-entropy, KL divergence, rank correlation and target-token trajectories quantify how intermediate distributions approach final output.
12. Causal follow-up
Use Causal Tracing to test whether stages identified by the tuned lens are functionally involved in the predicted transition.
13. Failure signatures
Overpowered translators, leakage, treating calibrated predictions as native intermediate softmaxes, cherry-picking and ignoring tuned-versus-untuned disagreement are major failures.
14. Evaluation blueprint
Freeze the model; define layer and token policy; train translators on separate data; evaluate held-out cross-entropy; compare with untuned lens and generic probes; inspect calibration and causal follow-up.
World Return
The tuned lens is a calibrated readout instrument. Its value is showing what predictive structure becomes simply recoverable at each depth without pretending the learned translator is itself the model’s hidden algorithm.