Research map / Architectures and efficiency / Attention and transformers
Discovering Interpretable Algorithms by Decompiling Transformers to RASP
Xinting Huang, Aleksandra Bakalova, Satwik Bhattamishra, William Merrill, Michael Hahn
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 |
|---|---|---|---|
| apply causal pruning to the D-RASP translation | D-RASP program, 4-layer 4-head transformer · Most Frequent | program lines: 16,201,616 -> 3 | Improved |
| apply causal pruning to the D-RASP translation | D-RASP program, 1-layer 4-head transformer · Most Frequent | program lines: 56 -> 3 | Improved |
| apply decompilation to length-generalizing models | Decompilation pipeline · Formal languages | decompilation success: 9 of 14 languages | Mixed |
| apply decompilation to non-length-generalizing models | Decompilation pipeline · Formal languages | decompilation success: 0 of 3 languages | Hurt |
| apply decompilation to non-length-generalizing models | Decompilation pipeline · Algorithmic tasks | decompilation success: 0 of 3 tasks | Hurt |
| apply decompilation to length-generalizing models | Decompilation pipeline · Algorithmic tasks | decompilation success: 8 tasks | Improved |
| apply causal pruning to non-length-generalizing models | Causal pruning · Algorithmic tasks | program size reduction | No clear effect |
| linearize layer norm (ablation) | Transformer · D3 Dyck language | decompilation success | Hurt |
Closest papers by meaning
- What Algorithms can Transformers Learn? A Study in Length Generalization (ICLR 2024 · similarity 0.92)
- Length Generalization for Transformers via Compression (NeurIPS 2026 · similarity 0.88)
- Do Efficient Transformers Really Save Computation? (ICML 2024 · similarity 0.87)
- Knee-Deep in C-RASP: A Transformer Depth Hierarchy (NeurIPS 2025 · similarity 0.87)
- Universal Length Generalization with Turing Programs (ICML 2025 · similarity 0.86)
- The Expressive Power of Transformers with Chain of Thought (ICLR 2024 · 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.
