Mechanistic Interpretability for Clinical LLMs: Do Sparse Autoencoders Make Explanations Useful?
Central question: Can sparse autoencoders offer the kind of actionable interpretability for large language models that would make clinicians more comfortable relying on these models?
What this might mean
If SAEs meaningfully illuminate which model components contribute to diagnostic or triage outputs, clinical AI might progress beyond “black box” catchphrases. The real test, though, is whether these insights are understandable or actionable for frontline clinicians, or merely more engineering-level abstractions. The promise of better error detection and safer deployment is enticing but calls for demonstration in actual workflows, not only on interpretability benchmarks.