Research map / Large language models / Language, safety and interpretability
Interpretable Embeddings with Sparse Autoencoders: A Data Analysis Toolkit
Nicholas Jiang, Xiaoqing Sun, Lisa Dunlap, Lewis Smith, Neel Nanda
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 SAE embeddings instead of LLM-based summarize/cluster baselines for dataset diffing | SAE embeddings · Model-output diffing (six comparisons) | average verified frequency difference per hypothesis | Improved |
| use SAEs instead of pure LLMs for hypothesis generation | SAE embeddings · Dataset diffing | token cost: 2-8x lower cost | Improved |
| use SAE embeddings instead of an LLM for finding injected correlations | SAE embeddings · Synthetic correlations in Pile (10 shuffles) | recovery of synthetic correlations: SAE found all 4; LLM found 1/10 to 9/10 | Improved |
| use SAE method instead of LLM and correlated topic model baselines on real data | SAE embeddings · CivilComments and Pile | verified NPMI of discovered pairs | Improved |
| filter SAE latents to reasoning-related ones for targeted clustering | SAE embeddings · GSM8K solutions | clustering along reasoning style | Improved |
| use SAE embeddings instead of dense embeddings on property-based retrieval | SAE embeddings · Six-dataset property retrieval benchmark | MAP | Improved |
| skip LLM reranking of retrieved latents (ablation) | SAE retrieval · Property-based retrieval | MAP | Hurt |
| relabel latents with the Pile and LMSYS-1M (ablation) | SAE retrieval · Property-based retrieval | MAP | Mixed |
| combine strongest baseline (OpenAI+LLM) with SAE scores | SAE retrieval · Property-based retrieval | MAP | Improved |
| increase reader model size from 8B to 70B (ablation) | SAE latents · GPT-5 responses (chat-like data) | median F1 of latents | Improved |
| increase reader model size from 8B to 70B (ablation) | SAE latents · arXiv and Pile | F1 of latents | No clear effect |
| use SAE retrieval on data far from the SAE training distribution | SAE retrieval · Biology abstracts and short stories | MAP | Mixed |
Closest papers by meaning
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.91)
- SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability (ICML 2025 · similarity 0.89)
- Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe (NeurIPS 2024 · similarity 0.89)
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.89)
- Demystifying Embedding Spaces using Large Language Models (ICLR 2024 · similarity 0.88)
- Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words (ICLR 2025 · similarity 0.88)
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.
