Interpretability Goals › Taxonomy of Interpretability Goals
Language AI HandbookAll chapters Interpretability Goals › Taxonomy of Interpretability GoalsOne section · Sections in this chapter5 / 14 ↓
Reading tools Definitions & inline links
Choose your expertise level to adjust how many terms are explained. Beginners see more tooltips, experts see fewer to maintain reading flow. Hover over underlined terms for instant definitions.
Article links
Make inline references clickable
The techniques above serve different goals. A clear taxonomy helps you choose the right tool for the question at hand. Using an attention visualization to verify safety properties, or a probing classifier to explain a specific user-facing decision, is likely to produce misleading results because the method is not suited to the goal.
| Goal | Primary Technique | What It Answers |
|---|---|---|
| Debug a specific failure | Attribution methods, attention visualization | Why did this input produce this output? |
| Audit for bias | Probing, feature analysis | Does the model encode protected attributes? |
| Locate factual knowledge | Activation patching, logit lens | Where is this fact stored? |
| Verify safety properties | Mechanistic analysis, probing | Is this dangerous capability present? |
| Scientific understanding | Sparse autoencoders, mechanistic analysis | What algorithms does the model implement? |
| Steer behavior | Activation steering | Can I modify this behavior by modifying activations? |
No single method answers all questions. Real interpretability investigations typically combine multiple techniques, using each to address its most suitable question and cross-validating findings across methods to distinguish robust results from method-specific artifacts.