The aim, as its proponents put it, is to do to a network what a programmer does to a compiled binary: recover readable source code (Elhage et al., 2021). Instead of asking which part of the input mattered for a prediction, one asks what computation the internal components implement (attention heads, neurons, directions in the residual stream) and how they chain together into circuits.
Millière & Buckner (2024b) frame the program within philosophy of science (2024b, secs. 2.2, 2.5). A mechanistic explanation identifies the parts, activities and organization that produce a phenomenon; Woodward's interventionism supplies the causal criterion: X causes Y if intervening on X changes Y with everything else held fixed. Like systems neuroscience, mechanistic interpretability targets Marr (1982)'s algorithmic level (see levels of analysis).
The signature tool is activation patching: run the model on "The capital of France is…" and cache its activations; run it on "The capital of Germany is…" and swap in a single component's activation from the first run. If the answer flips back to "Paris," that component carries causally relevant information (Millière & Buckner, 2024b, sec. 2.5).
Landmark results: induction heads, which complete patterns of the form [A][B]…[A] → [B] and may account for part of in-context learning (Elhage et al., 2021; Buckner, 2024, pp. 281–282); the algorithm a small network learns for modular addition, tied to grokking; and the board representations in Othello-GPT, which suggest a world model (Millière & Buckner, 2024b, secs. 2.5.2–2.5.3). The limits: most of this work uses toy models, the largest models are usually closed, and superposition prevents reading the network neuron by neuron.
The findings feed old disputes. For Buckner (2024), the fact that copying plus abstract similarity yields something rule-like favors empiricism (2024, p. 282); Millière and Buckner also note that the mechanism resembles variable binding in classical systems (Millière & Buckner, 2024b, sec. 2.5.1). The quarrel between connectionism and symbolic AI thus reappears inside the network.