What is mechanistic interpretability, and how does it differ from post-hoc interpretability methods?

• A. Mechanistic interpretability is about external explanations and does not examine internal computation.
• B. Mechanistic interpretability focuses on the broad outcomes rather than internal mechanisms.
• C. Mechanistic interpretability seeks to understand the internal components and exact causal mechanisms by which a model produces outputs; post-hoc methods approximate explanations without confirming the internal structure and can be brittle.
• D. Mechanistic interpretability aims to mimic human reasoning without inspecting neurons.

Answer: (C)

Mechanistic interpretability aims to uncover the actual machinery inside a model—the specific internal components and the exact causal pathways that transform inputs into outputs. It involves tracing how information flows through neurons or submodules, identifying which parts implement particular operations, and showing how altering a component changes the result, thus proving a causal link between internal structure and behavior. Post-hoc interpretability, by contrast, provides explanations after training that approximate the model’s reasoning without confirming the true internal circuitry. Techniques like feature attributions or saliency maps can produce plausible explanations but may not reflect the real internal mechanisms and can be brittle if the model or inputs change. Because mechanistic interpretability seeks to map and validate the actual internal computations, it best captures the distinction described.