Logit Lens

methods / Causal Validation / Logit Lens

Techniquebeginner

Projects the residual stream after an intermediate layer directly through the model's final unembedding matrix, reading off what the model's 'current best guess' would be if generation stopped at that layer — a forward-pass-only diagnostic for tracking how a prediction builds up across depth.

Used in (11 observations)

FFN value vectors are vocabulary-space directions in upper layersmeasured

structure: Linear Direction · models: Adaptive Input Transformer LM (16-layer, WikiText-103) · paper: Transformer Feed-Forward Layers Are Key-Value Memories

Leela Chess Zero's correct moves appear early, then get overriddenmeasured

structure: Linear Subspace, Linear Direction · models: Leela Chess Zero (policy network) · paper: The Algorithm Is Not the Behavior: Learned Priors Override Look-Ahead in a Chess-Playing Neural Network

FFN value-vector directions causally suppress toxicitymeasured

structure: Linear Direction · models: GPT-2 Small · paper: Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space

Fine-tuned GPT-2-XL solves addition with sparse Fourier featuresmeasured

structure: Lissajous Curves · models: GPT-2-XL · paper: Pre-trained Large Language Models Use Fourier Features to Compute Addition

Sparse logit-lens directions in VLA FFN value vectors steer robot actionsmeasured

structure: Linear Direction · models: OpenVLA-7B, pi0 (PaliGemma VLA backbone, predecessor of pi0.5) · paper: Mechanistic Interpretability for Steering Vision-Language-Action Models

A per-token mean-difference direction fingerprints narrow fine-tuningmeasured

structure: Linear Direction · models: Qwen3-1.7B, Llama-3.1-8B-Instruct, Gemma 3 1B IT, Qwen2.5-7B-Instruct · paper: Narrow Finetuning Leaves Clearly Readable Traces in Activation Differences

Function vectors steer even where logit lens and probes cannot decodemeasured

structure: Linear Direction · models: Llama-3.1-8B, Llama-3.1-8B-Instruct, Gemma-2-9B, Gemma-2-9B-it, Mistral-7B-v0.3, Mistral-7B-Instruct-v0.3 · paper: Steerable but Not Decodable: Function Vectors Operate Beyond the Logit Lens

Editing MLP output-vector directions suppresses memorized sequencesmeasured

structure: Linear Direction · models: SmolLM-360M, OLMo-1B, OLMo-7B, Llama-2-7B · paper: Output Vector Editing for Memorization Mitigation in Large Language Models

TabPFN linearly encodes regression coefficients and arithmetic intermediatesmeasured

structure: Linear Subspace · models: TabPFN v2 (tabular in-context-learning foundation model) · paper: TabPFN Through The Looking Glass: An Interpretability Study of TabPFN and Its Internal Representations

A Sudoku transformer builds substructure-level linear constraint directionsmeasured

structure: Linear Direction, Constraint-Algebra Basis Hypothesis · models: Sudoku Transformer (8 layers, 8 heads, d_model=576) · paper: Transformers Linearly Represent Highly Structured World Models

Baseline-corrected cosine similarity between intermediate hidden states of semantically-equivalent inputs (different languages, notations, code, images, audio) is elevated at middle layers relative to a non-equivalent-pair baseline, logit-lens decoding shows non-English/non-textual inputs pass through tokens of the model's dominant pretraining language, and causally patching an intermediate representation with its dominant-language equivalent predictably redirects downstream output across data typesmeasured

structure: Platonic Representation Hypothesis · models: Llama-2-7B, Llama-2-70B, Llama-3-8B, Llama-3-70B · paper: The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities