![[2602.16698] Causality is Key for Interpretability Claims to Generalise](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.16698] Causality is Key for Interpretability Claims to Generalise
Summary
This paper discusses the importance of causality in interpretability research for large language models, highlighting pitfalls in generalization and causal claims.
Why It Matters
Understanding the causal relationships in model behavior is crucial for developing reliable interpretability methods in machine learning. This research addresses common issues in the field, providing a framework for practitioners to ensure their findings are valid and generalizable.
Key Takeaways
- Causal inference is essential for valid interpretability claims.
- Current interpretability studies often fail to generalize findings.
- A diagnostic framework is proposed to align claims with evidence.
Computer Science > Machine Learning arXiv:2602.16698 (cs) [Submitted on 18 Feb 2026] Title:Causality is Key for Interpretability Claims to Generalise Authors:Shruti Joshi, Aaron Mueller, David Klindt, Wieland Brendel, Patrik Reizinger, Dhanya Sridhar View a PDF of the paper titled Causality is Key for Interpretability Claims to Generalise, by Shruti Joshi and Aaron Mueller and David Klindt and Wieland Brendel and Patrik Reizinger and Dhanya Sridhar View PDF 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 (\eg, 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 cau...
