Research map / Large language models / Language, safety and interpretability
Interpretable and Steerable Concept Bottleneck Sparse Autoencoders
Akshay Kulkarni, Tsui-Wei Weng, Vivek Narayanaswamy, Shusen Liu, Wesam A. Sakla, Kowshik Thopalli
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 |
|---|---|---|---|
| prune low-utility SAE neurons and add concept bottleneck (CB-SAE) | SAE · LLaVA-1.5-7B | CLIP-Dissect interpretability: 0.154 -> 0.244 | Improved |
| CB-SAE | SAE · LLaVA-1.5-7B | steerability (unit vector): 0.198 -> 0.261 | Improved |
| CB-SAE | SAE · LLaVA-MORE | steerability (unit vector): 0.179 -> 0.192 | Improved |
| CB-SAE | SAE · LLaVA-MORE | CLIP-Dissect interpretability: 0.194 -> 0.291 | Improved |
| CB-SAE | SAE · UnCLIP | steerability (white image): 0.654 -> 0.664 | No clear effect |
| CB-SAE | SAE · UnCLIP | monosemanticity: 0.540 -> 0.594 | Improved |
| add steerability cyclic loss Lst (ablation) | CB-SAE · LLaVA-1.5-7B | steerability: +2.9% steerability | Improved |
| prune SAE with interpretability+steerability sum instead of one score (ablation) | CB-SAE · LLaVA-1.5-7B | interpretability and steerability | Mixed |
| retain fewer SAE neurons after pruning (ablation) | CB-SAE · LLaVA-1.5-7B | interpretability and steerability | Mixed |
| use CB neurons vs retained SAE neurons (ablation) | CB-SAE · LLaVA-1.5-7B | interpretability: 0.238 -> 0.323 | Improved |
| use CB neurons vs retained SAE neurons (ablation) | CB-SAE · LLaVA-1.5-7B | steerability (unit vector): 0.263 -> 0.231 | Hurt |
| keep discarded SAE neurons (ablation) | SAE · LLaVA-1.5-7B | CLIP-Dissect interpretability: 0.238 (retained) -> 0.084 (discarded) | Hurt |
Closest papers by meaning
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.92)
- Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment (ICML 2025 · similarity 0.91)
- Mechanistic Interpretability with Sparse Autoencoder Neural Operators (NeurIPS 2026 · similarity 0.91)
- Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control (ICLR 2025 · similarity 0.91)
- Cascaded Sparse Autoencoders LearnMulti-Level Visual Concepts in Multimodal LLMs (NeurIPS 2026 · similarity 0.90)
- From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit (NeurIPS 2025 · similarity 0.90)
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.
