Automated Discovery of Functional Circuits in Neural Networks

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

Version 1 of 122 papers24 conceptsfrozen 2026-10-04

For years, mechanistic interpretability has felt like digital archaeology. Researchers like Arthur Conmy and Augustine N. Mavor-Parker have spent months manually "reverse-engineering" small models, painstakingly tracing how specific attention heads or neurons collaborate to perform tasks like numerical comparison.

However, as models scale to billions of parameters, this manual approach hits a wall. The emerging research direction focuses on Automated Circuit Discovery: developing algorithms that can "perform surgery" on a model to identify the minimal functional subnetworks—or circuits—responsible for specific behaviors without human intervention.

The Scaling Bottleneck of Manual Interpretability

The fundamental problem identified across this literature is that manual discovery does not scale. In Towards Automated Circuit Discovery for Mechanistic Interpretability, the authors note that relying on researcher intuition to pre-select components restricts our ability to find unexpected logic. As Hassan Sajjad and Fahim Dalvi argue, the sheer volume of neurons in modern LLMs makes exhaustive fine-grained analysis computationally prohibitive.

To solve this, researchers are moving toward "automated discovery" tools. For instance, Michael A. Lepori and Ellie Pavlick introduced NeuroSurgeon, a toolkit that uses optimization-based masking to isolate these subnetworks. The goal is to move from "black box" models to a collection of "human-readable algorithms" hidden within the weights.

How Automated Discovery Works: Pruning and Attribution

The core intuition behind automated discovery is causal pruning. If you can remove a connection between two components without degrading the model's performance on a specific task, that connection is not part of the task's "circuit."

  1. Iterative Pruning (ACDC): The Automatic Circuit DisCovery (ACDC) algorithm iteratively prunes unimportant connections, successfully rediscovering complex logic (like GPT-2's numerical comparison) that previously took humans months to map.
  2. Continuous Sparsification: Philipp Mondorf and Barbara Plank use continuous sparsification to find minimal functional subnetworks for string-edit tasks, showing that these circuits are often modular.
  3. Cross-Layer Attribution (CLA): In vision models, Achyuta Rajaram and Sarah Schwettmann developed CLA to trace functional connectivity between neurons, allowing them to identify the specific subgraphs responsible for visual concepts.

Circuits as Modular Building Blocks

A key insight from this research is that circuits are not just task-specific one-offs; they are general-purpose algorithmic building blocks.

Jack Merullo and Ellie Pavlick demonstrated a 78% overlap between circuits for two different tasks. This modularity allows for "circuit composition," where simple circuits can be combined (using set operations like Union) to handle complex behaviors that individual circuits cannot Philipp Mondorf et al..

This modularity even extends to how models represent data. Xinyan Hu and Jacob Steinhardt found that in-context learning for addition relies on low-dimensional activation subspaces structured with trigonometric functions—a geometric "circuit" for arithmetic.

From Understanding to Intervention

The ultimate value of automated circuit discovery lies in its application to model training and safety. Once a circuit is identified, it can be manipulated:

  • Circuit Distillation: Somin Wadhwa and Byron C Wallace proposed Circuit Distillation, which teaches a small "student" model to align its internal computational subgraphs with a "teacher" model, transferring the actual algorithm rather than just the output.
  • Mechanistic Unlearning: Phillip Guo and colleagues found that targeting specific "Fact Lookup" (FLU) circuits allows for more robust unlearning than traditional methods, which often only mask the model's outputs.
  • Model Steering: By identifying the "relay team" of attention heads responsible for variable binding, Brenden M. Lake and Cheng Tang showed we can "steer" model behavior through precise internal activation edits.

By automating the discovery of these internal mechanisms, the field is moving toward a future where we can debug, edit, and compress neural networks with the same precision we use for traditional software.

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_automated_discovery_of_functional_circui_605766da,
  title = {Automated Discovery of Functional Circuits in Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-discovery-of-functional-circuits-in-neural-networks/rdr_605766da2ded42f29a4e0443307b123e},
  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_automated_discovery_of_functional_circui_605766da_v1,
  title = {Automated Discovery of Functional Circuits in Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-discovery-of-functional-circuits-in-neural-networks/rdr_605766da2ded42f29a4e0443307b123e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}