Research map / Large language models / Language, safety and interpretability
Sparse Autoencoders for Hypothesis Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HypotheSAEs, which trains a k-sparse autoencoder on text embeddings, selects predictive SAE neurons with L1 regression, and interprets them with an LLM to produce natural-language hypotheses about text-target relationships.
Problem
Generating interpretable natural-language hypotheses that predict a target variable from text datasets, which is hard to scale with LLM prompting over few examples and hard to interpret with fine-tuned neurons.
Approach
Decouple hypothesis generation into unsupervised SAE feature learning on text embeddings, supervised L1 selection of predictive neurons, and LLM-based labeling of each neuron, with a theoretical triangle inequality linking neuron predictiveness and interpretation fidelity to hypothesis separation score.
Key results
- On synthetic tasks HypotheSAEs achieves F1 0.58 on Wiki-5 vs 0.48 for BERTopic, beating all baselines on 8 of 9 dataset-metric pairs.
- Across three real datasets, 45 of 60 candidate hypotheses are significant, versus at most 24 for baselines.
- HypotheSAEs is roughly 30-50x faster and 10-50x cheaper than NLParam and HypoGeniC on Congress; on 200K Yelp reviews it costs about $1 vs >$500 for baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAE features rather than selecting features directly from embeddings (ablation) | HypotheSAEs · Headlines, Yelp, Congress | Significant hypotheses: 30 / 60 -> 45 / 60 | Helped |
| Use SAE features rather than selecting features directly from embeddings (ablation) | HypotheSAEs · Headlines, Yelp, Congress | Average fidelity: 0.545 -> 0.836 | Helped |
| Use SAE features rather than selecting features directly from embeddings (ablation) | HypotheSAEs · Headlines, Yelp, Congress | Average predictiveness: 0.706 -> 0.717 | Helped |
| Use an SAE rather than PCA-reduced embeddings (ablation) | HypotheSAEs · Headlines, Yelp, Congress | Significant hypotheses: 23 / 60 -> 45 / 60 | Helped |
| Use an SAE rather than PCA-reduced embeddings (ablation) | HypotheSAEs · Headlines, Yelp, Congress | Average fidelity: 0.547 -> 0.836 | Helped |
| Use an SAE rather than a supervised 20-neuron bottleneck (ablation) | HypotheSAEs · Headlines, Yelp, Congress | Significant hypotheses: 27 / 60 -> 45 / 60 | Helped |
| Use an SAE rather than a supervised 20-neuron bottleneck (ablation) | HypotheSAEs · Headlines, Yelp, Congress | Average fidelity: 0.640 -> 0.836 | Helped |
| Use an SAE rather than a supervised 20-neuron bottleneck (ablation) | HypotheSAEs · Headlines, Yelp, Congress | Average predictiveness: 0.699 -> 0.717 | Helped |
| Recover reference topics with HypotheSAEs rather than BERTopic | Topic hypothesis generation · Wiki-5 | F1: 0.48 -> 0.58 | Helped |
| Recover reference topics with HypotheSAEs rather than BERTopic | Topic hypothesis generation · Wiki-15 | F1: 0.40 -> 0.47 | Helped |
| Use HypotheSAEs rather than BERTopic | Real-world hypothesis generation · Headlines | AUC: 0.62 -> 0.69 | Helped |
| Use HypotheSAEs rather than HypoGeniC | Real-world hypothesis generation · Yelp | R²: 0.78 -> 0.76 | Hurt |
Surprising findings
- Hypotheses from HypoGeniC on Yelp mostly restate the rating prediction task (e.g., 'expresses disappointment with the overall dining experience'), offering little specific insight.
- A simpler variant without the fidelity validation step achieves costs similar to BERTopic while still outperforming it.
Limitations
- Interpretation fidelity depends on LLM labeling; the approach relies on GPT-4o for concept generation and GPT-4o-mini for annotation, making results API-dependent.
- Evaluation on synthetic datasets relies on reference topics that may not reflect real-world hypotheses.
- Separation score is one measure of predictiveness; theoretical bound uses assumptions like Y in [0,1].
Experimental setup
Datasets: Wiki (WikiText-derived, Hierarchical topics) Bills (GovTrack) Headlines (Upworthy Research Archive) Yelp Open Dataset Congress speeches (109th Congress)
Benchmarks: Wiki-5 Wiki-15 Bills-5 Headlines Yelp Congress
Metrics: F1 similarity surface similarity AUC R^2 count of significant hypotheses separation score helpfulness and interpretability human ratings
Baselines: BERTopic NLParam HypoGeniC Embedding Lasso ablation PCA ablation Bottleneck ablation
Builds on: k-sparse autoencoder BERTopic HypoGeniC NLParam Lasso
Closest papers by meaning
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.86)
- Automatically Interpreting Millions of Features in Large Language Models (ICML 2025 · similarity 0.84)
- Sparsity-Guided Holistic Explanation for LLMs with Interpretable Inference-Time Intervention (AAAI 2024 · similarity 0.84)
- Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words (ICLR 2025 · similarity 0.84)
- Measuring and Guiding Monosemanticity (NeurIPS 2025 · similarity 0.84)
- Scaling and evaluating sparse autoencoders (ICLR 2025 · similarity 0.84)
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.