Research map / Large language models / Language, safety and interpretability
Interpretability at Scale: Identifying Causal Mechanisms in Alpaca
Zhengxuan Wu, Atticus Geiger, Thomas Icard, Christopher Potts, Noah Goodman
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use random rotation instead of learned rotation (ablation) | Boundless DAS · Left and Right Boundary, layer 10 pos 75 | IIA: 0.83 -> 0.53 | Hurt |
| evaluate only on correct examples | Boundless DAS · Left and Right Boundary | IIA_max: 0.86 -> 0.88 | Improved |
| evaluate only on incorrect examples | Boundless DAS · Left and Right Boundary | IIA_max: 0.86 -> 0.71 | Hurt |
| test alignment on unseen price bracket | Left Boundary alignment · New Bracket | IIA_max: 0.94 -> 0.95 | No clear effect |
| add random prefix contexts to instruction | Left and Right Boundary alignment · Irrelevant Contexts | IIA_max: 0.86 -> 0.83 | Hurt |
| change output template from yes/no to True/False | Left and Right Boundary alignment · Sibling Instructions | IIA_max: 0.86 -> 0.83 | Hurt |
| exclude top right corner from correlation | Sibling Instructions alignment · Sibling Instructions | correlation: 0.87 -> 0.92 | Improved |
| align two boolean variables simultaneously instead of one | Boundless DAS · Left vs Left and Right Boundary | IIA_max: 0.90 -> 0.86 | Mixed |
| replace hypothesis with mid-point distance causal model | Boundless DAS · Price Tagging | IIA_max: 0.86 -> 0.70 | Hurt |
| replace hypothesis with bracket identity causal model | Boundless DAS · Price Tagging | IIA_max: 0.86 -> 0.72 | Hurt |
| learn boundary (mask) instead of fixed-size subspace | Boundless DAS · Boundary learning dynamics | IIA | No clear effect |
Closest papers by meaning
- The Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability? (NeurIPS 2025 · similarity 0.86)
- MIB: A Mechanistic Interpretability Benchmark (ICML 2025 · similarity 0.86)
- Neurons to Words: A Novel Method for Automated Neural Network Interpretability and Alignment (AAAI 2025 · similarity 0.86)
- Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable? (ICLR 2025 · similarity 0.86)
- Bounded Rationality for LLMs: Satisficing Alignment at Inference-Time (ICML 2025 · similarity 0.86)
- HyperDAS: Towards Automating Mechanistic Interpretability with Hypernetworks (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
