Mapping and Steering Causal Circuits in Language Models

This direction is merged into another direction: activation patching. Its last version stays readable and citable here.

Version 1 of 123 papers23 conceptsfrozen 2026-10-04

Large Language Models (LLMs) are often treated as monolithic "black boxes," where billions of parameters interact in inscrutable ways to produce text. However, a growing body of research suggests that these models are actually composed of sparse, modular sub-networks—or circuits—dedicated to specific tasks like multi-digit addition, grammatical agreement, or even human emotion inference. By identifying these circuits, researchers can move beyond mere observation and perform "surgical" interventions to fix model errors or steer behavior without retraining the entire system.

The Intuition of Activation Patching

The primary tool for discovering these circuits is activation patching (also known as causal mediation analysis). To understand why a model makes a specific decision, researchers compare two nearly identical scenarios: a "clean" input that produces the correct answer and a "corrupted" input that leads to an error.

By systematically swapping internal activations—such as the output of a specific attention head or a feed-forward layer—from the clean run into the corrupted run, researchers can measure which components "restore" the correct output. As Ethan Sun and Warren Shepard demonstrate in Algorithmic Phase Transitions in Language Models: A Mechanistic Case Study of Arithmetic, this allows us to isolate the exact causal path the model takes to solve a problem. If patching a single attention head fixes the corrupted output, that head is a critical node in the task's circuit.

Sparsity and Functional Localization

A recurring finding across this research is that complex behaviors are often driven by a tiny fraction of the model's architecture. Wei Zhang and colleagues found in Interpreting and Improving Large Language Models in Arithmetic Calculation that less than 5% of attention heads are responsible for arithmetic. Similarly, Wei Chen and Wei Chen 0005 identified that only about 4% of heads drive "sycophancy"—the tendency of models to agree with user misconceptions—in From Yes-Men to Truth-Tellers: Addressing Sycophancy in Large Language Models with Pinpoint Tuning.

This localization isn't just limited to facts or logic. Ala N. Tak and Amin Banayeeanzade discovered that emotion inference is functionally localized within specific middle layers, as detailed in Mechanistic Interpretability of Emotion Inference in Large Language Models. These findings suggest that LLMs develop specialized "subroutines" rather than relying on a global, undifferentiated reasoning process.

Universal and Language-Specific Circuits

Interestingly, these circuits often transcend individual languages. In On the Similarity of Circuits across Languages: a Case Study on the Subject-verb Agreement Task, Javier Ferrando and Marta R. Costa-jussà found that models use nearly identical internal pathways to handle grammar in different languages.

However, when languages have unique requirements—like specific tense suffixes—the model recruits language-specific subroutines in its feed-forward networks. Ruochen Zhang and Ellie Pavlick explore this in The Same but Different: Structural Similarities and Differences in Multilingual Language Modeling, showing that models maintain a "universal core" of logic while branching out for linguistic nuances.

Surgical Intervention and Steering

Once a circuit is mapped, we can intervene directly on its components. This is more efficient than standard fine-tuning because it avoids the "tax" of degrading the model's general knowledge.

Portability via Function Vectors

Perhaps the most striking evidence of modularity is the discovery of Function Vectors (FVs). Eric Todd and Aaron Mueller showed in Function Vectors in Large Language Models that specific tasks (like translation) are represented as compact vectors within the model's layers. These vectors can be extracted from one context and "pasted" into another to trigger the task zero-shot, proving that the model's internal representations are not just descriptive, but causally potent instructions.

By treating the model as a collection of steerable circuits rather than an opaque block of weights, researchers are developing a "mechanistic" manual for AI—one that allows us to diagnose, repair, and enhance models with unprecedented precision.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_mapping_and_steering_causal_circuits_in__b821bb59,
  title = {Mapping and Steering Causal Circuits in Language Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mapping-and-steering-causal-circuits-in-language-models/rdr_b821bb59651a41edac3f009112de2bab},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_mapping_and_steering_causal_circuits_in__b821bb59_v1,
  title = {Mapping and Steering Causal Circuits in Language Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mapping-and-steering-causal-circuits-in-language-models/rdr_b821bb59651a41edac3f009112de2bab/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}