logit lens

**Logit lens** is the **analysis technique that projects intermediate hidden states through the final unembedding to estimate token preferences at each layer** - it offers a quick view of how predictions evolve across model depth. **What Is Logit lens?** - **Definition**: Applies output projection to hidden activations before final layer to inspect provisional logits. - **Interpretation**: Shows which candidate tokens are being formed at intermediate computation stages. - **Speed**: Provides lightweight diagnostics without full retraining or heavy instrumentation. - **Limitation**: Raw projections can be biased because intermediate states are not optimized for direct decoding. **Why Logit lens Matters** - **Layer Insight**: Helps visualize when key information appears during forward pass. - **Debug Utility**: Useful for spotting layer regions where target signal is lost or distorted. - **Education**: Provides intuitive interpretability entry point for new researchers. - **Hypothesis Generation**: Supports rapid exploration before deeper causal analysis. - **Caution**: Results need careful interpretation due to calibration mismatch. **How It Is Used in Practice** - **Comparative Use**: Compare logit-lens trajectories between successful and failing prompts. - **Token Focus**: Track rank and probability shifts for specific expected tokens. - **Validation**: Confirm lens-based hypotheses with patching or ablation experiments. Logit lens is **a fast diagnostic lens for intermediate token prediction dynamics** - logit lens is valuable for exploration when its projection bias is accounted for in interpretation.