Research map / Architectures and efficiency / Attention and transformers
Sparse autoencoders reveal selective remapping of visual concepts during adaptation
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.
Trains PatchSAE, a patch-level sparse autoencoder on CLIP ViT-B/16, to interpret concepts and their spatial attributions, and uses it to show that MaPLe adaptation mainly remaps existing concepts to downstream classes rather than introducing new ones.
Problem
It is unclear which internal mechanisms change when foundation models like CLIP are adapted with prompt-based methods, and how concepts in the vision encoder relate to downstream classification gains.
Approach
Apply sparse autoencoders to all patch and CLS tokens of a frozen CLIP ViT residual stream to get spatially localized monosemantic concepts, then share this SAE with the prompt-adapted MaPLe model to compare class-level concept activations and ablate concepts via top-k masking.
Key results
- Top-k class-level SAE latent masking shows class-discriminative information; full SAE reconstruction recovers 64.82% vs 68.25% original ImageNet zero-shot accuracy, while all-zero mask gives 0.1%.
- Table of class-level latent groups shows most top-activating latents overlap between CLIP and MaPLe, with few high-to-low or low-to-high latents on most of 11 base-to-novel datasets.
- EuroSAT shows the largest remapping effect: MaPLe improves base accuracy from 42.61% to 73.07% (per class-level comparison), with 4.1 latents newly active on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MaPLe multimodal prompt adaptation | CLIP ViT-B/16 · 11 base-to-novel benchmarks, base split | accuracy: MaPLe outperforms CLIP on base classes in all 11 datasets in the table | Helped |
| add MaPLe multimodal prompt adaptation | CLIP ViT-B/16 · 11 base-to-novel benchmarks, novel split | accuracy | Mixed |
| mask all SAE latents, retaining the identity reconstruction mask (ablation) | CLIP ViT-B/16 · ImageNet-1K zero-shot | accuracy: 68.25% -> 64.82% | Hurt |
| mask all SAE latents to zero (ablation) | CLIP ViT-B/16 · ImageNet-1K zero-shot | accuracy: 68.25% -> 0.1% | Hurt |
| select per-class top-activating SAE latents rather than random or dataset-level latents (ablation) | CLIP classification with PatchSAE · ImageNet-1K zero-shot | accuracy | Helped |
| set λ_l1 to 0.00 rather than 8e-5 (ablation) | PatchSAE ViT-L/14 · ImageNet SAE training | MSE: 0.0026 -> 0.0000 | Mixed |
| increase λ_l1 from 8e-5 to 8e-4 (ablation) | PatchSAE ViT-L/14 · ImageNet SAE training | MSE: 0.0026 -> 0.0069 | Hurt |
| use expansion factor 32 rather than 64 (ablation) | PatchSAE ViT-L/14 · ImageNet SAE training | MSE: 0.0026 -> 0.0028 | Hurt |
| use expansion factor 128 rather than 64 (ablation) | PatchSAE ViT-L/14 · ImageNet SAE training | MSE: 0.0026 -> 0.0025 | Helped |
| remove ghost gradient (ablation) | PatchSAE ViT-L/14 · ImageNet SAE training | L0: 29.48 -> 19.35 | Hurt |
| initialize decoder bias with the mean rather than geometric median (ablation) | PatchSAE ViT-L/14 · ImageNet SAE training | MSE: 0.0026 -> 0.0027 | Hurt |
| train on CLS tokens only rather than all tokens (ablation) | PatchSAE ViT-B/16 · ImageNet SAE training | MSE: 0.0025 -> 0.0009 | Helped |
Surprising findings
- Adaptation gains on most of the 11 tasks can be explained by existing concepts already present in the non-adapted CLIP; adaptation mainly reassigns concepts to classes rather than introducing new ones.
- MaPLe makes better use of the same number of SAE latents than CLIP does in top-k masking experiments.
Limitations
- Analysis limited to CLIP ViT-B/16 and a single prompt-based adaptation method (MaPLe); other adaptation methods such as full fine-tuning are left to future work.
- Interpretability of latents is assessed qualitatively via reference images and statistics; many high-entropy latents are uninterpretable.
- Scope restricted to the CLIP vision encoder without joint multimodal analysis.
Experimental setup
Datasets: ImageNet-1K Caltech101 OxfordPets StanfordCars Flowers102 Food101 FGVC Aircraft SUN397 DTD EuroSAT UCF101 ImageNet-Sketch
Benchmarks: ImageNet-1K Caltech101 OxfordPets StanfordCars Flowers102 Food101 FGVC Aircraft SUN397 DTD EuroSAT UCF101
Metrics: accuracy improvement rate (Delta) MSE L0 L1 contrastive loss
Baselines: CLIP zero-shot MaPLe
Builds on: sparse autoencoders MaPLe CoOp CLIP
Closest papers by meaning
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.88)
- Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment (ICML 2025 · similarity 0.88)
- Interpreting CLIP with Hierarchical Sparse Autoencoders (ICML 2025 · similarity 0.88)
- Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models (NeurIPS 2025 · similarity 0.88)
- Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering (NeurIPS 2026 · similarity 0.88)
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · 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.