Skip to content

Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Sparse Autoencoders (SAEs) are applied to Vision-Language Models (e.g., CLIP) to decompose polysemantic neurons into monosemantic features. The authors introduce the MonoSemanticity Score (MS), a quantitative metric validated against human perception, and show that SAE interventions on CLIP's vision encoder can steer multimodal LLM outputs (LLaVA) to insert or suppress specific concepts without modifying the language model.

Problem

VLM neurons (e.g., in CLIP) are polysemantic — a single neuron fires for semantically unrelated concepts due to the superposition hypothesis. No neuron-level quantitative metric existed to measure monosemanticity in vision representations, and VLM interpretability via SAEs lacked rigorous evaluation or demonstrated downstream steerability.

Method

SAEs (BatchTopK and Matryoshka BatchTopK variants) are trained on pre-extracted activation vectors from CLIP ViT-L/14 (and SigLIP, AIMv2, WebSSL) at various layers, with expansion factors ε ∈ {1, 2, 4, 8, 16, 64} and sparsity K = 20 by default.

MonoSemanticity Score (MS): For each SAE neuron k, given a dataset of N images, embeddings are extracted with a frozen DINOv2 ViT-B encoder to build a pairwise cosine similarity matrix S. Neuron activations are min-max normalized per neuron to produce relevance weights R^k_{nm} = ã^k_n · ã^k_m. MS^k is the activation-weighted average pairwise image similarity: MS^k = Σ_{n<m} R^k_{nm} s_{nm} / Σ_{n<m} R^k_{nm}. Higher MS indicates more topically coherent (monosemantic) activating images.

MLLM Steering: A pretrained SAE (trained on layer 22 token embeddings of CLIP) is injected after CLIP's vision encoder in LLaVA-1.5-7b. For a target neuron k, its activation is clamped to a scalar α across all visual tokens; the manipulated activations are decoded back to token embeddings, which propagate through the unchanged LLM.

Human validation uses 1,000 pairwise questions on Mechanical Turk (71 users, majority voting) to assess MS alignment with human monosemanticity judgments.

Key Contributions

  • MonoSemanticity Score (MS): A neuron-level, activation-weighted pairwise image similarity metric for quantifying monosemanticity in vision SAEs, validated against human perception.
  • Large-scale user study benchmark: 1,000 annotated pairwise neuron comparisons released for future benchmarking.
  • Systematic SAE evaluation: Quantitative comparison of BatchTopK vs. Matryoshka BatchTopK SAEs across 5 CLIP layers, 6 expansion factors, 4 sparsity levels, and 4 vision encoders.
  • MLLM concept steering: Demonstrated insertion and suppression of concepts in LLaVA by manipulating single SAE neurons in CLIP's encoder, outperforming Difference-in-Means (DiffMean) baselines.

Results

  • Human alignment: MS agrees with human monosemanticity judgments in 82.8% of cases overall; alignment rises monotonically from 56.6% (δ ∈ [0.0, 0.1]) to 100% (δ ∈ [0.8, 0.9]).
  • SAE vs. no SAE: Original CLIP neurons score MS ≈ 0.01; SAEs at ε=1 already reach MS 0.57–0.73 (BatchTopK) and 0.82–0.86 (Matryoshka); at ε=64, both variants reach MS = 1.00 for best neurons across all evaluated layers.
  • Matryoshka superiority: Matryoshka BatchTopK achieves higher MS than plain BatchTopK at equivalent expansion factors (e.g., layer 23, ε=4: 0.89 vs. 0.83), though at ~2–3 points lower R².
  • Sparsity effect: Lower K yields higher MS (K=1: R²=31.3%; K=50: R²=74.9%); K=20 selected as balance (R²=66.8%).
  • Cross-encoder generality: MS improves universally with SAEs across WebSSL, CLIP, SigLIP, AIMv2 (all No SAE: 0.01; Matryoshka ε=4: 0.85–0.92).
  • Concept insertion (vs. DiffMean): Both criteria satisfied — 42.4% (SAE) vs. 35.8% (DiffMean); prompt adherence 85.8% vs. 66.2%.
  • Concept suppression (vs. DiffMean): Both criteria satisfied — 52.5% (SAE) vs. 33.3% (DiffMean); unrelated content preserved 81.4% vs. 38.7%.
  • CLIP-space sanity check: Steering improves image-text similarity by ~22% relative gain within the reference [0.185, 0.283] range.

Limitations

  • MS is computed over ImageNet, so monosemanticity reflects dataset-level concept coverage; neurons may be polysemantic on out-of-distribution inputs.
  • High MS neurons do not always produce precise MLLM steering effects (e.g., a "golden retriever" neuron triggers generic dog outputs), due to MLLM fine-grained alignment gaps with the vision encoder.
  • A fraction of SAE neurons acting as feature detectors exhibit no clear steering effect.
  • Evaluation is limited to common dictionary-learning SAE variants; MS is model-agnostic but not tested on alternative decomposition approaches.
  • MS is currently defined only for vision; extension to text representations is left as future work.

Relevance to Vision-Language Models

This paper directly advances interpretability of CLIP-class VLMs by providing the first rigorous, human-validated neuron-level monosemanticity metric for visual representations, filling a gap left by LLM-focused SAE work. The steering result — modifying a single SAE neuron in CLIP to control LLaVA outputs without touching the LLM — is practically significant for mechanistic control of multimodal pipelines. For VLM researchers, this establishes SAEs as an off-the-shelf probe that generalizes across training objectives (contrastive, autoregressive, self-supervised), and the released benchmark enables future comparison of VLM representation disentanglement methods.