Computation
The circuits and causal pathways that read internal representations, transform them, and write the results toward output, and the methods that establish those pathways causally.
Knowing what a model represents is not knowing what it computes. This programme studies the mechanisms: attention heads that route information between positions, MLPs that transform it, and the compositions of both: circuits that implement identifiable algorithms across layers.
The methodological core is causal intervention. Activation patching, causal tracing, attribution graphs, and model editing all answer the same underlying question: if this component's contribution is changed, does the behaviour change as predicted? Correlation-level evidence (attention weights, probe accuracy, plausible stories) does not settle mechanism, and this programme treats the gap between the two as its central discipline.
A specific concern of the institute sits here: the relationship between a model's verbalised reasoning and its actual internal computation. Chain-of-thought is an output, produced by the same machinery it purports to describe. Where the two diverge, only circuit-level evidence can say so.
Attention & MLP circuits
Head composition, induction, name-moving, and the algorithmic roles of individual components.
Attribution graphs
Tracing which upstream features cause which downstream ones through replacement models.
Activation patching
Swapping activations between runs to localise where behaviour-relevant computation happens.
Causal tracing & circuit discovery
Automated and manual identification of the subgraph that suffices for a behaviour.
Information flow across layers
The residual stream as a communication bus: who reads, who writes, and when.
Hidden vs verbalised computation
Where chain-of-thought matches the mechanism, and where it demonstrably does not.
Model editing & causal intervention
Editing stored associations as an existence proof of localised computation.
- How much of a given behaviour is explained by the circuits we can currently find, and what carries the remainder?
- Do discovered circuits generalise beyond the task distribution they were found on?
- Can attribution methods scale from single behaviours to the routine auditing of a frontier model?
- When chain-of-thought and internal computation diverge, is the divergence systematic and predictable?
Methods in use or proposed: Activation patching · path patching · attribution graphs · ablation studies · model editing · TransformerLens-based replication.
A short orientation list of primary sources, not a survey. The Interpretability Map holds the full dependency graph.
- Elhage, N., et al. (2021). A Mathematical Framework for Transformer Circuits. Transformer Circuits Thread. transformer-circuits.pub/2021/framework/index.html
- Olsson, C., et al. (2022). In-context Learning and Induction Heads. Transformer Circuits Thread. transformer-circuits.pub/2022/in-context-learning-and-induction-heads/index.html
- Wang, K., et al. (2022). Interpretability in the Wild: A Circuit for Indirect Object Identification in GPT-2 Small. arxiv.org/abs/2211.00593
- Meng, K., Bau, D., Andonian, A., & Belinkov, Y. (2022). Locating and Editing Factual Associations in GPT. NeurIPS 2022. arxiv.org/abs/2202.05262
- Conmy, A., et al. (2023). Towards Automated Circuit Discovery for Mechanistic Interpretability. NeurIPS 2023. arxiv.org/abs/2304.14997
- Heimersheim, S., & Nanda, N. (2024). How to Use and Interpret Activation Patching. arxiv.org/abs/2404.15255
- Lindsey, J., et al. (2025). On the Biology of a Large Language Model. Transformer Circuits Thread. transformer-circuits.pub/2025/attribution-graphs/biology.html