💬 NLP

Circuit Fingerprints: How Answer Tokens Encode Their Geometrical Path

The paper proposes the "Circuit Fingerprint" hypothesis, which posits that answer tokens geometrically encode the directions of the circuits that produce them, allowing for circuit discovery and activation steering through geometric alignment rather than gradient-based or causal intervention methods.

Original authors: Andres Saurez, Neha Sengar, Dongsoo Har

Published 2026-02-11

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Andres Saurez, Neha Sengar, Dongsoo Har

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand how a massive, complex city works. There are two main ways you could study it:

The Detective Approach (Circuit Discovery): You follow people around, watch where they walk, and see which intersections they use to get from point A to point B. You are trying to map out the "roads" (the circuits) that people use to complete specific tasks.
The Architect Approach (Activation Steering): You don't care about the history; you just want to change the city. You want to know, "If I want everyone to feel happy, which streets should I decorate with flowers?" You are trying to "steer" the city's vibe.

For a long time, AI researchers thought these were two totally different jobs. But this paper, "Circuit Fingerprints," argues that they are actually two sides of the same coin.

The Big Idea: The "Breadcrumb" Theory

The researchers discovered something fascinating: The destination itself contains a map of the journey.

Think of it like this: If you find a pile of sand in the middle of a desert, you can look at the shape and texture of that sand to figure out exactly which wind currents and dunes it traveled through to get there. The sand carries a "fingerprint" of its path.

In an AI model, the "answer" (like the word "Paris") isn't just a random word that pops out. The way the model's internal "brain" represents the word "Paris" actually contains a geometric signature of all the mathematical "roads" the model traveled to reach that conclusion.

The "Read-Write" Duality

The paper proposes that because the answer holds the map, we can do two things with it:

READING (The Detective): We can look at the "fingerprint" of an answer to work backward and find the exact "roads" (the neurons and connections) the model used to get there. This allows us to find the "circuit" without having to do the exhausting work of manually testing every single connection one by one.
WRITING (The Architect): Because we now know exactly which "roads" lead to a certain feeling or fact, we can "write" on those roads. If we want the AI to be more joyful, we don't just ask it nicely (prompting); we actually go into its internal "map" and nudge the activations along the "joyful" roads.

Why does this matter? (The Results)

The researchers tested this on several AI models and found two big wins:

It’s Efficient: They could find the "roads" the AI uses to solve puzzles (like identifying names or agreeing with verbs) almost as well as the expensive, slow, traditional methods, but much more simply.
It’s More Powerful: When they tried to make the AI act "emotional," their "steering" method worked much better than just telling the AI, "Please act happy." Their method got the emotion right about 70% of the time, while just asking the AI via text only worked about 53% of the time. Most importantly, the AI stayed smart and factual while it was being emotional.

The Takeaway

This paper suggests that an AI isn't just a black box of random numbers. Instead, it is a geometric landscape. Every thought the AI has leaves a footprint, and if we learn to read those footprints, we don't just understand how the AI thinks—we gain the steering wheel to guide it.

Technical Summary: Circuit Fingerprints

Title: Circuit Fingerprints: How Answer Tokens Encode Their Geometrical Path
Authors: Andres Saurez, Neha Sengar, Dongsoo Har (KAIST)

1. Problem Statement

In mechanistic interpretability, two major research threads have traditionally operated in isolation:

Circuit Discovery: Locating the specific sub-networks (attention heads and MLPs) responsible for a task, typically via causal intervention (activation patching) or gradient-based methods.
Activation Steering: Controlling model behavior by adding learned vectors to the residual stream to shift the model's output style or content.

The authors address the fundamental question: Are these two approaches studying the same underlying structure? They hypothesize that the distinction between "reading" (interpreting) and "writing" (controlling) is artificial, and that both are dual operations on a single geometric structure encoded within the model's activation space.

2. Methodology: The Circuit Fingerprint Hypothesis

The core hypothesis is that answer tokens encode the computational pathways that produced them. Even when processed in isolation, the representation of an answer token (e.g., "Paris") carries a "geometric signature" of the circuit required to generate it.

A. Reading (Circuit Discovery via Geometric Attribution):
Instead of using expensive backpropagation or iterative patching, the authors propose a geometric method:

Target Direction Extraction: They take two contrastive answer tokens ( $a^+$ and $a^-$ ) and compute their difference in the residual stream at the final layer: $\Delta r(L) = r(a^+) - r(a^-)$ . This defines the direction that distinguishes the two answers.
Component Scoring: For every component $c$ (head or MLP), they transform the target direction into the component's "native space" and measure the alignment between the component's differential output ( $\Delta o_c$ ) and this transformed target.
Edge-Level Attribution: To capture indirect influence, they use a Shapley decomposition to distribute credit among the Query (Q), Key (K), and Value (V) channels of attention heads. This allows them to propagate importance backward through the model, calculating how much an upstream component $i$ contributes to a downstream component $j$ .

B. Writing (Geometric Steering):
The authors use the same directions extracted during discovery to perform interventions:

Subspace Construction: They use SVD on answer token prototypes to create an orthonormal basis.
Intervention: They shift the model's activations along these learned directions. For stylistic steering (e.g., emotion), they use the magnitude of the source feature to inject the target feature into the residual stream.

3. Key Contributions

Unification of Interpretability: Proves that circuit discovery (reading) and activation steering (writing) are dual operations on the same geometric object.
Gradient-Free Discovery: Demonstrates that circuit membership can be recovered through geometric alignment alone, matching the performance of gradient-based methods (like EAP-IG) without the computational cost of backpropagation.
Instruction-Based Feature Discovery: Shows that one can discover "persona circuits" (e.g., a "pirate" or "joyful" circuit) simply by extracting directions from instruction-modified prompts, requiring no task-specific datasets.

4. Results

Circuit Tracing Performance: On benchmarks like Indirect Object Identification (IOI), Subject-Verb Agreement (SVA), and Multiple-Choice QA (MCQA), the "Circuit Fingerprint" (CF) method achieved CMD (Circuit-Model Distance) and CPR (Circuit Performance Ratio) scores comparable to state-of-the-art gradient methods across four model families (GPT-2, Qwen, Llama, OPT).
Steering Efficacy: In emotion steering tasks, the geometric method achieved 69.8% emotion classification accuracy, significantly outperforming standard instruction prompting (53.1%), while maintaining high factual accuracy (~89.6%).
Functional Clustering: The Shapley decomposition revealed that attention heads cluster by functional roles (e.g., "Name Movers" are Q-dominated, while "S-Inhibition" heads are K-dominated), providing a mechanistic explanation for head behavior.

5. Significance

This paper shifts the paradigm of mechanistic interpretability from viewing circuits as purely causal graphs to viewing them as fundamental geometric structures in activation space. By showing that the "fingerprint" of an answer is its own computational path, the authors provide a more efficient toolkit for both understanding (interpreting) and manipulating (steering) large language models. This "read-write duality" suggests that the complexity of model control may be lower than previously thought, as the directions for understanding a model are the same directions for controlling it.