Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models

1 Introduction

• SAE neurons score significantly higher in monosemanticity than original VLM neurons, even at equal width, indicating greater concept separability.The figure contrasts polysemantic top-activating images for an original VLM neuron with more monosemantic images for an SAE neuron.Source: Figure 1 · p. 2, §1 · p. 2
• The MonoSemanticity score (MS) measures activation-weighted pairwise image similarity, and human judgments confirm that score differences track perceived conceptual focus.A large-scale user study provides a benchmark for evaluating alignment between MS and human interpretation.Source: §1 · p. 2
• Intervening on a single monosemantic SAE neuron in CLIP’s vision encoder can insert or suppress its concept in multimodal LLM outputs while leaving the LLM untouched.The paper illustrates this transfer using LLaVA.Source: §1 · p. 2
• Matryoshka SAE achieves overall superior MS, while wider and sparser latents lead to better scores.Source: §1 · p. 2

Sparse Autoencoders

• SAEs build on dictionary learning for model interpretability, with learned features used to interpret and steer LLMs and mechanisms including Batch TopK [8], JumpReLU [48], and Matryoshka-inspired designs.Matryoshka representation learning was also explored for SAEs [34].Source: §2 · pp. 2–3

Vision-Language Models

• Prior work examined VLM representations and interpretable neurons [15] as interest grew in understanding and steering models whose trustworthiness is a concern [4].VLMs align images and text [32] [56] or generate text conditioned on images [12] [19] [36].Source: §2 · p. 3

SAEs for VLMs

• SAEs have been applied to CLIP and other vision encoders [35] [49] [50], diffusion models [11] [30] [31] [51], cross-encoder concept discovery [53], and multimodal LLMs [63].Prior multimodal monosemanticity work focused on inter-modality differences [58].Source: §2 · p. 3
• This work introduces an empirically human-aligned score for individual-neuron monosemanticity and a rigorous SAE evaluation framework for VLMs, and demonstrates steering multimodal LLMs.The score provides a grounded quantitative measure; the work also compares SAEs’ interpretation and control with operating directly on raw features.Source: §2 · p. 3

3.1 Background and Formulation of SAEs

• An SAE encodes VLM layer embeddings into an overcomplete latent activation vector and decodes them into a reconstruction, training its linear layers with reconstruction and sparsity objectives.The formulation uses expansion factor ε to set latent width ω = d × ε; ReLU with L2 reconstruction and L1 sparsity is a basic variant, while TopK controls sparsity through K and Matryoshka trains nested dictionaries with multiple reconstruction objectives.Source: §3.1 · pp. 3–4

3.2 Monosemanticity Score

• The MonoSemanticity score measures each neuron’s monosemanticity as pairwise image-embedding similarity weighted by normalized neuron activations across a diverse set of unseen images.The score is computed for each SAE neuron, and higher similarity among strongly activating images indicates a clearer shared concept.Source: §3.2 · p. 4, Figure 3 · p. 5

3.3 Steering MLLMs with Vision SAEs

• SAE interventions steer MLLM responses by changing a target concept’s activations in vision-token embeddings, without modifying the underlying language model or textual input.The SAE is applied after the vision encoder; manipulated activations are decoded back into token embeddings before the model generates its response.Source: §3.3 · p. 5

4.1 Experimental Settings

• SAEs were evaluated on CLIP, SigLIP, AIMv2, and WebSSL encoders using activations from ImageNet images.CLIP activations came from CLS tokens at layers 11, 17, 22, and 23 or the final projection; steering SAEs used two random token embeddings per image from layer 22. Comparisons covered BatchTopK and Matryoshka variants across expansion factors ε ∈ {1, 2, 4, 8, 16, 64}, with K = 20 for BatchTopK.Source: §4.1 · pp. 5–6

4.2.1 Alignment of MS with human perception

• Users’ pairwise judgments matched MS in 82.8% overall, rising from 56.6% at δ∈(0.0,0.1) to 100.0% at δ∈(0.8,0.9).The study comprised 1,000 questions from 71 users, with three answers per question aggregated by majority vote.Source: §4.2.1 · p. 6, Figure 4 · p. 6

4.2.2 Monosemanticity of SAEs

• SAEs produced higher best-neuron MS than No SAE, with scores increasing with expansion and Matryoshka gains consistent across vision encoders.The MS increase was observed even at expansion factor ε = 1; Matryoshka SAEs improved concept separation but achieved about 2 or 3 points lower R2 at the same expansion factors.Source: §4.2.2 · pp. 6–7, Table 1 · p. 7

No SAE

• SAE latent neurons achieved higher maximum MS scores than original neurons across all tested image encoders, with larger expansion factors helping discover more monosemantic units.The reported comparison concerns the highest observed MS scores in the last layers of the encoders.Source: No SAE · p. 7
• About 90% of SAE neurons scored above original neurons even at ε = 1; scores rose overall through ε = 4, then declined at larger expansion factors.Although the relative fraction of neurons at each MS level decreased for very wide latents, their absolute number increased.Source: No SAE · p. 7
• Stricter sparsity improved monosemanticity, but R2 ranged from 31.3% at K = 1 to 74.9% at K = 50; K = 20 balanced these with R2 of 66.8%.The authors selected K = 20 to balance interpretability and reconstruction quality.Source: No SAE · pp. 7–8

4.3 Steering Multimodal LLMs

• For insertion, intended-concept introduction was 48.7% vs. 53.1% and base-prompt adherence 85.5% vs. 66.2% for SAE vs. DiffMean.The evaluation used 100 SAE neurons, 10 diverse text prompts per neuron, and an LLM-as-a-judge setup.Source: Table 2 · p. 8, §4.3 · p. 8
• For suppression, target-concept removal was 64.4% vs. 64.0% and unrelated-content preservation 81.4% vs. 38.7% for SAE vs. DiffMean.The suppression evaluation checked whether descriptions omitted the target concept while correctly describing unrelated images.Source: Table 2 · p. 8, §4.3 · p. 8
• Steering increased image-text similarity by 22% within reference bounds of 0.185 ± 0.028 for random pairs and 0.283 ± 0.034 for correct image-class pairs.The sanity checks tested steering on white images and random ImageNet images.Source: Table 3 · p. 9, §4.3 · p. 9
• Increasing a pencil-associated SAE neuron's activation shifted LLaVA's love-poem output from following the prompt toward pencil attributes and eventually pencils alone.The intervention clamped the selected neuron's activation at progressively higher α values.Source: Figure 6 · p. 8, §4.3 · p. 9
• Decreasing knife- and laptop-associated neuron activations made LLaVA progressively omit those concepts while continuing to describe other image objects.The authors present this as a possible way to filter undesired content before it reaches the language model.Source: §4.3 · p. 9, Figure 7 · p. 9

5 Conclusion

• The MonoSemanticity score evaluates neuron-level monosemanticity in VLM SAEs; sparsity, wider latents, and Matryoshka SAEs were associated with stronger monosemanticity.The conclusion identifies sparsity and wider latents as primary contributors and highlights Matryoshka SAEs' superior performance.Source: §5 · p. 10
• A large-scale human study verified that the MonoSemanticity score aligns with human perception.Source: §5 · p. 10
• Potential future work includes adapting the metric to text representations and studying interactions between specialized low-level and broad high-level concepts.Source: §5 · p. 10
• The evaluation focused on SAE architectures that scale to large VLMs, although the model-agnostic metric could also assess comparable approaches.Source: §5 · p. 10
• High-scoring neurons do not always steer precisely: an ImageNet golden-retriever feature may elicit any dog-related output, and some feature detectors show no clear steering effect [1].The authors suggest this may reflect MLLMs' limited fine-grained understanding or imperfect alignment with the vision encoder.Source: §5 · p. 10

A Broader Impact

• By incorporating human judgment, the MonoSemanticity score offers a way to evaluate dictionary-learning methods such as SAEs and support trust in systems that use them.Source: A Broader Impact · p. 15
• SAEs can encourage or discourage model behaviors and help models recognize or ignore concepts, including potentially dangerous ones, to support desired outputs and alignment with human values.Source: A Broader Impact · p. 15

B More details on steering

• Intervening on an SAE neuron steers reconstructed CLIP image tokens toward its concept, influencing the LLM’s generated output.SAEs are trained to reconstruct vision-encoder token embeddings and attached after the encoder at inference; increasing a neuron’s activation can add elements associated with images that highly activate it.Source: B More details on steering · p. 15
• An SAE neuron can make LLaVA-like models respond as if they see an absent concept, while interventions also alter scientific titles at different strengths.Figure A1 illustrates steering toward a panda absent from the input image; Figure A2 categorizes intervention magnitudes as 0, medium, and high for neurons associated with polka dots, shipwreck, and rainbow.Source: Figure A1 · p. 15, B More details on steering · p. 15, Figure A2 · p. 16
• Steering was evaluated with an LLM-as-judge setup across ten prompt types, including poems, scientific titles, to-do lists, and dialogue.Other prompts requested a rap song, math word problem, Wikipedia-style paragraph, new holiday, newspaper headline and opening paragraph, or party conversation starter.Source: B More details on steering · p. 16

C User study

• For each neuron pair, three annotators chose which grid of 16 top-activation images looked more similar and focused; the majority choice was compared with the higher-MS neuron.Alignment was defined as agreement between the human and MS choices, and the overall score was the fraction of evaluated neuron pairs on which they agreed.Source: C User study · p. 17
• The study collected 1,000 pair rankings from 71 Mechanical Turk annotators and sampled neurons from CLIP ViT-L and three SAE settings spanning CLIP ViT-L and SigLIP SoViT-400m.The SAE samples were BatchTopK on CLIP ViT-L, Matryoshka on CLIP ViT-L, and BatchTopK on SigLIP SoViT-400m, each with ε = 4 and K = 20.Source: C User study · p. 17
• Across MS-distance bins 0.0–0.1 to 0.4–0.5, AS (E = CLIP ViT-B) was 0.55, 0.81, 0.93, 0.96, 0.93; AS (E = DINOv2 ViT-B) was 0.53, 0.77, 0.92, 0.96, 0.93.The respective pair counts were 292, 249, 186, 178, and 95; groups were designed to be sufficiently large and comparable, though not perfectly equal.Source: C User study · pp. 17–18, Table A1 · p. 18

D Benchmark

• The benchmark releases user-study data to support evaluation of neuron monosemanticity and future alternatives to MS.Researchers can compare methods directly to MS using the same underlying information, including complete ImageNet validation-set neuron activations.Source: D Benchmark · p. 22
• The benchmark includes 1,000 neuron pairs with user preferences and MS values for DINOv2 ViT-B and CLIP ViT-B.It also provides each study neuron's 16 most activating ImageNet training images and activations across all 50,000 ImageNet validation images.Source: D Benchmark · p. 22

E.1 Unnormalized plots

• As SAE width increases, neuron scores cover a wider range and more neurons exceed a given monosemanticity threshold.Figure A6 shows scores in decreasing order for No SAE and SAE expansion factors ×1, ×2, ×4, ×16, and ×64.Source: §E.1 · p. 22, Figure A6 · p. 22

E.2 Detailed statistics and more models

• Average MS remains similar as expansion increases, but rises substantially from the original layer to ε = 1; best scores improve consistently with greater expansion.Tables A2 and A5 report average, best, and worst MS statistics for CLIP ViT-L evaluated with DINOv2 ViT-B.Source: §E.2 · p. 23, Table A2 · p. 23, Table A5 · p. 24
• With CLIP ViT-B as evaluator, CLIP ViT-L SAE neurons still outperform original neurons less distinctly, while SigLIP SoViT-400m results resemble those for CLIP ViT-L.The SigLIP SAEs are evaluated with CLIP ViT-B as the image encoder.Source: §E.2 · p. 23
• SAE neurons are more monosemantic than original neurons across three model–evaluator setups: CLIP ViT-L with DINOv2, CLIP ViT-L with CLIP ViT-B, and SigLIP SoViT-400m with CLIP ViT-B.The single-neuron comparisons are shown for the last layers, with and without an SAE.Source: Figure A7 · p. 25, §E.2 · p. 25
• With CLIP ViT-B evaluation, higher expansion and lower sparsity help identify more monosemantic units, following patterns similar to DINOv2 ViT-B evaluation.Figures A8 and A9 compare MS across neurons at different expansion factors and sparsity levels.Source: §E.2 · p. 25, Figure A8 · p. 25, Figure A9 · p. 26

E.3 Matryoshka hierarchies

• Average LCA depth and average and maximum MS correlate with Matryoshka group level, aligning the learned hierarchy with taxonomy and locating the most specialized neurons at lower levels.Neuron depth is estimated from the average LCA depth among pairs of each neuron's top-16 activating images; group sizes correspond to the first five iNaturalist taxonomy levels [55].Source: §E.3 · p. 26, Table A8 · p. 26

F Reconstruction of SAEs

• The reconstruction appendix reports R2 and L0 for SAE variants and different K values, noting that test-time sparsity may differ from K = 20.BatchTopK is replaced at test time by ReLU(x − γ), where per-neuron thresholds are estimated from minimum positive activations across batches; the appendix reports comparisons in Tables A9–A11.Source: F Reconstruction of SAEs · p. 27, Table A9 · p. 27, Table A10 · p. 27

G Uniqueness of concepts

• Only 16,000 of 75,368,503 neuron pairs (> 0.03%) had J > 0, and just 20 had J > 0.5, indicating high uniqueness of learned concepts.The analysis uses top-16 activating training images for 12,288 neurons, excluding 10 neurons with fewer than 16 activating images, in a last-layer Matryoshka SAE with expansion factor 16.Source: G Uniqueness of concepts · p. 28

• Figure A10 shows highly activating images for the Pencil neuron used for steering in Figure 6.Source: H Additional qualitative results · p. 28
• Across DINOv2 ViT-B and CLIP ViT-B, higher MS scores correspond to more similar highest-activating images, illustrating the metric’s correlation with monosemanticity.Source: H Additional qualitative results · p. 28, Figure A11 · p. 29, Figure A12 · p. 30
• Figure A11 shows DINOv2 ViT-B neuron examples ordered from high to low MS, with higher-scoring neurons’ highest-activating images more similar.Source: Figure A11 · p. 29
• Figure A12 shows CLIP ViT-B neuron examples ordered from high to low MS, with higher-scoring neurons’ highest-activating images more similar.Source: Figure A12 · p. 30