Causality is Key for Interpretability Claims to Generalise

TL;DR

This paper emphasizes the importance of causal inference in interpretability studies of large language models, proposing a framework that clarifies what claims can be supported and how to ensure their generalizability.

Contribution

It introduces a causal hierarchy-based diagnostic framework for interpretability, guiding practitioners in selecting appropriate methods and evaluations for valid, generalizable claims.

Findings

Causal inference clarifies what interpretability claims can justify.

Interventions support causal effects on model behavior.

Counterfactual claims require controlled supervision for verification.

Abstract

Interpretability research on large language models (LLMs) has yielded important insights into model behaviour, yet recurring pitfalls persist: findings that do not generalise, and causal interpretations that outrun the evidence. Our position is that causal inference specifies what constitutes a valid mapping from model activations to invariant high-level structures, the data or assumptions needed to achieve it, and the inferences it can support. Specifically, Pearl's causal hierarchy clarifies what an interpretability study can justify. Observations establish associations between model behaviour and internal components. Interventions (e.g., ablations or activation patching) support claims how these edits affect a behavioural metric (e.g., average change in token probabilities) over a set of prompts. However, counterfactual claims -- i.e., asking what the model output would have been for the same prompt under an unobserved intervention -- remain largely unverifiable without controlled supervision. We show how causal representation learning (CRL) operationalises this hierarchy, specifying which variables are recoverable from activations and under what assumptions. Together, these motivate a diagnostic framework that helps practitioners select methods and evaluations matching claims to evidence such that findings generalise.