Circuit Tracing

Transcoder based Attribution tracing

• CLT compresses amplification chains across multiple MLP layers into a single feature representation, resulting in shorter path lengths in the graphs

• CLT globally optimizes MLP outputs across layers (joint training), achieving lower MSE than PLT while explaining much more variance than thresholded neurons

Overall Pipeline

1. Build a replacement model by substituting the original model's MLP with a "Cross-Layer Transcoder (CLT)"

2. Visualize the computational flow of the replacement model as an "attribution graph" for specific prompts

3. Identify critical paths in the graph

4. Verify the causal role of individual features using "patching" techniques

CLT (cross-layer

Transcoder)

Unlike

Crosscoder which only shares latent dimensions, CLT shares encoders and trains different decoders for subsequent layers. While PLT (Per Layer Transcoder) was trained to mimic each layer's MLP input-to-output function to learn causality, CLT scaled this approach with n encoder-decoder pairs. CLT maintains the same encoder for each layer, but has decoders for all subsequent causal MLP outputs (including itself), capturing much more diverse cross-layer causality. Through this, with + encoders trained, it creates a half fully-connected graph, then combines various pruning techniques with correlation-based importance scores and metrics like TWERA and ERA to obtain the final

Attribution Graph. Causality is then verified through patching, though the graph construction itself is not based on patching-derived causality.

Decoder sparsity loss

Building on

Crosscoder's decoder sparsity loss, CLT uses tanh activation for decoders to achieve appropriate regularization: behaving linearly near 0 and saturating to 1 for larger values, which stabilizes training.

Local Replacement Model

• Replace all MLP subblocks in attention blocks with CLT

• Freeze the attention patterns and layer-norm denominators from the original model

• Correct node output differences (errors) for each token and layer using error nodes

Causal Role Verification

1. First, run the base model to record MLP outputs at each layer and store the decoder contributions of CLT features.

2. In the second forward pass, multiply all decoder contributions of selected features by a coefficient and replace with "recorded MLP output + ".

3. Analyze logit changes by considering suppression, removal, and amplification results from patching.

2. Global circuit discovery (Context independent Global graph)

Global Weights

1. TWERA (Target-Weighted Expected Residual Attribution)

2. ERA (Expected Residual Attribution)

Evaluation

• Path Length
• Measure how much the path between "embedding→logit" relies on short paths by examining the cumulative impact ratio of to calculate the average number of steps to reach the output

• Completeness
• The ratio of edge contribution sum from "feature·embedding→logit" in pruned graph to total edge contribution sum. The closer to 1, the smaller the error nodes (unexplained portions)

• Replacement Score
• The proportion of paths from "embedding→logit" that use only feature paths without errors

• Mechanistic Fidelity
• Cosine similarity between perturbations, measuring the similarity of downstream activation change vectors when patching the same features in the same way between original model vs. local replacement model
• Normalized MSE: Difference in magnitude of changes between the two models shows accumulated errors per layer

• Automated Interpretability
• Sort Eval: Show two features (top-k examples) to an LLM and measure the distinguishability index of feature visualization if they are decomposable
• Contrastive Eval: Present differing features from contrasting prompts (varying by one point) to LLM and measure the probability of correctly identifying the source prompt as an indicator of how well features capture differences

Limitation

Supernode is a grouping of multiple feature nodes into a "single semantic unit" to simplify visualization and causal effect analysis of nodes with similar roles. These are manually designated.

Circuit Tracing: Revealing Computational Graphs in Language Models

We describe an approach to tracing the “step-by-step” computation involved when a model responds to a single prompt.

https://transformer-circuits.pub/2025/attribution-graphs/methods.html

On the Biology of a Large Language Model

We investigate the internal mechanisms used by Claude 3.5 Haiku — Anthropic's lightweight production model — in a variety of contexts, using our circuit tracing methodology.

https://transformer-circuits.pub/2025/attribution-graphs/biology.html#dives-tracing

Neuronpedia Research

circuit-tracer

decoderesearch • Updated 2026 Jun 25 13:28

Two-step reasoning (e.g., Dallas→Texas→Austin) actually uses intermediate holes. Language-agnostic reasoning followed by language-specific feature combination. CLT shows better replacement score/sparsity tradeoff compared to PLT, while skip PLT generally offers fewer benefits.

The Circuits Research Landscape: Results and Perspectives - August 2025

A multi-organization interpretability project to replicate and extend circuit tracing research.

https://www.neuronpedia.org/graph/info

gemma-2-2b Attribution Graph

https://www.neuronpedia.org/gemma-2-2b/graph

Attribution Graphs for Dummies - 1. What are Attribution Graphs?

Part 2: https://youtu.be/hdi1a9MjwDs An introduction to attribution graphs from Anthropic's Circuit Tracing and Model Biology papers, featuring Jack Lindsey (Anthropic), Emmanuel Ameisen (Anthropic), Tom McGrath (Goodfire AI), and Neel Nanda (Google DeepMind). 0:00 Introduction 2:18 Attribution Graph Orientation 19:10 Analyzing an Attribution Graph from Scratch 40:25 Reflection: What have we Learned? Explore Attribution Graphs: https://neuronpedia.org/graph Blog Post: https://www.neuronpedia.org/graph/info circuit-tracer GitHub: https://github.com/safety-research/circuit-tracer Original Papers by Anthropic - Circuit Tracing: https://transformer-circuits.pub/2025/attribution-graphs/methods.html - Biology of an LLM: https://transformer-circuits.pub/2025/attribution-graphs/biology.html

https://www.youtube.com/watch?v=ruLcDtr_cGo

TWERA

Circuits Updates – May 2026

A short update on understanding features through downstream connections.

https://transformer-circuits.pub/2026/may-update/index.html