Logit Lens

The core mechanism bypasses the final layer norm and unembedding step. For each layer l, the residual stream activation h_l is multiplied directly by the unembedding matrix W_U to produce logits. This reveals what the model would predict if it stopped early, exposing the iterative refinement process. The technique requires no additional training or learned parameters, making it a zero-cost diagnostic tool.

Applying the logit lens across all layers produces a trajectory of predictions. Early layers typically output nonsensical or low-confidence tokens, reflecting local syntactic processing. Middle layers often converge on the correct semantic category. Final layers refine the exact token identity. This trajectory visualizes the progressive sharpening of the output distribution as information flows through the residual stream.

The logit lens applies the unembedding matrix without any learned transformation, which can produce noisy or distorted logits due to representational drift across layers. The tuned lens improves on this by learning an affine transformation for each layer's hidden state before unembedding. This distinction is critical: the logit lens reveals raw representational content, while the tuned lens provides a more accurate but less direct projection.

Researchers often identify a specific layer where the logit lens output most closely matches the model's final prediction. This 'logit lens layer' is where the model has effectively settled on its answer, with subsequent layers performing minor refinements or confidence calibration. In GPT-2 style models, this often occurs around layers 10-14 out of 24, indicating that substantial computation is dedicated to output polishing rather than core reasoning.

The logit lens enables causal experiments by patching intermediate activations. An analyst can take the residual stream from a corrupted input at layer l and patch it into a clean forward pass, then use the logit lens to observe how the prediction shifts. This isolates the causal contribution of specific layers to the final output without requiring full ablation studies, accelerating circuit discovery workflows.

By examining logit lens outputs during factual recall tasks, researchers can identify early divergence points where the model begins predicting incorrect tokens. This divergence often occurs in middle layers, well before the final output, providing a potential signal for real-time hallucination detection. The technique reveals that factual errors are not sudden final-layer mistakes but the culmination of an incorrect trajectory initiated much earlier.