💻 computer science

Interpretability Transfer from Language to Vision via Sparse Autoencoders

This paper introduces VISTA, a framework that transfers interpretability from language to vision by aligning visual tokens with a pre-existing, labeled textual sparse autoencoder space in vision-language models, thereby enabling precise visual concept localization and effective cross-modal interventions without training dedicated vision SAEs.

Original authors: Alexey Kravets, Da Li, Chuan Li, Da Chen, Vinay P. Namboodiri

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alexey Kravets, Da Li, Chuan Li, Da Chen, Vinay P. Namboodiri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart robot that can see pictures and talk about them. This robot has a "brain" (a Large Language Model) that is incredibly good at understanding words, but it's a bit of a mystery how it actually "sees" the world. Usually, when we try to peek inside this robot's brain to see what it's thinking about a picture, it's like trying to read a book written in a language nobody speaks yet.

This paper introduces a new method called VISTA (Visual Interpretability via SAE Transfer Alignment) to solve this problem. Here is how it works, explained simply:

1. The Problem: The "Dictionary" Mismatch

Think of the robot's brain as having a massive, pre-written dictionary of concepts (like "dog," "running," "happy"). This dictionary was built by reading millions of books and is very organized. However, when the robot looks at a picture, the signals it gets from its eyes (the visual data) don't match the words in this dictionary. They are like two different languages.

To understand what the robot sees, scientists usually try to build a new dictionary specifically for pictures. But this is hard because labeling visual concepts is confusing. Is a "dog" feature just a dog, or a specific breed, or a dog running? It's ambiguous and takes a lot of work to label.

2. The Solution: VISTA (The Universal Adapter)

Instead of building a new dictionary for pictures, VISTA acts like a universal translator adapter.

The Setup: The robot has a frozen "brain" (the language model) with its existing, labeled dictionary. It also has a "camera" (a vision encoder) that sees the world.
The Trick: VISTA trains a small, simple connector (a projector) between the camera and the brain.
The Goal: The goal is to force the camera's signals to fit perfectly into the brain's existing dictionary. The researchers do this by adding a special rule during training: "The picture signals you send to the brain must be reconstructible using the brain's own word-dictionary."

It's like forcing a square peg (the picture) to fit into a round hole (the word dictionary) by reshaping the peg, rather than trying to build a whole new square hole.

3. The Result: Seeing is Understanding

Because VISTA forces the picture signals to align with the word dictionary, we can now "read" what the robot sees using the words it already knows.

The "Neuronpedia" Connection: The paper uses a public database called Neuronpedia, which acts like a legend for the robot's brain. It tells us that a specific signal means "cat" or "cookie."
The Breakthrough: With VISTA, when the robot looks at a picture of a cat, the signal it sends to the brain lights up the exact same "cat" concept in the dictionary as if someone had just typed the word "cat." This happens without needing to re-label anything or train a new dictionary.

4. The Camera Matters: DINOv2 vs. CLIP

The paper also tested different "cameras" (vision encoders) to see which one works best with this translator.

CLIP (The Global Thinker): Imagine a camera that looks at a whole picture and says, "This is a scene with a cat." It's good at the big picture but bad at details. If you ask it to point to exactly where the cat is, it might get confused and point to the background.
DINOv2 (The Local Detective): This camera is like a detective that looks at small patches of the image. It knows exactly which tiny square of pixels contains the cat's ear and which contains the tail.
The Finding: VISTA works best with DINOv2. Because DINOv2 keeps the location of objects precise, VISTA can perform "surgery" on the robot's vision.

5. The Magic Trick: Visual Steering

Because the picture signals are now perfectly aligned with the word dictionary, the researchers can perform "visual steering." This is like editing the robot's perception of reality using math.

The Experiment: They took an image of a "girl eating a chocolate cookie."
The Action: They found the specific signals for "chocolate" and subtracted them, or added signals for "bread."
The Result: The robot's description changed from "eating a chocolate cookie" to "eating a sandwich" or "eating bread," but only for that specific part of the image. The rest of the scene (the girl, the chair) stayed exactly the same.

Without VISTA (and specifically with the DINOv2 camera), the robot would either fail to change the object or would start hallucinating weird things (like a person appearing out of nowhere) because it couldn't pinpoint exactly where to make the change.

Summary

In short, this paper shows that we don't need to teach a robot a new language to understand its vision. Instead, we can teach its "eyes" to speak the same language as its "brain." By using a specific type of camera (DINOv2) and a clever training trick, we can make the robot's vision so clear and precise that we can not only understand what it sees but also edit its perception of specific objects in a picture, all while keeping the rest of the world intact.

Technical Summary: Interpretability Transfer from Language to Vision via Sparse Autoencoders

1. Problem Statement

The interpretability of Large Multimodal Models (LMMs) remains a significant challenge. While Sparse Autoencoders (SAEs) have successfully decomposed polysemantic language features into monosemantic, interpretable concepts within Large Language Models (LLMs), applying this success to the visual domain is hindered by two primary limitations:

Labeling Ambiguity: Training domain-specific SAEs on visual encoders yields monosemantic visual features, but labeling these features is difficult. Distinguishing whether a feature represents a specific object, a breed, or an action is ambiguous. Furthermore, vision-only SAEs do not reveal how visual features are interpreted once they enter the LLM's residual stream.
Computational Cost: Fine-tuning existing LLM-based SAEs to incorporate visual information is prohibitively expensive. For instance, adapting a model like Gemma Scope requires massive computational resources and extensive re-labeling efforts (e.g., hundreds of thousands of LLM requests) to account for shifted latent spaces.

The central question addressed is: Can language sparse autoencoders be reused to interpret visual concepts without training new vision-specific SAEs?

2. Methodology: VISTA Framework

The authors propose VISTA (Visual Interpretability via SAE Transfer Alignment), a framework designed to transfer interpretability from language to vision within a LLaVA-style vision-language model. The core innovation is constraining the visual projector to map visual tokens directly into the LLM's pre-existing, labeled textual SAE space.

Architecture and Training Objective

Setup: The framework utilizes a frozen vision encoder (e.g., DINOv2, CLIP) and a frozen LLM. A trainable linear projector ( $P_\theta$ ) maps visual tokens ( $v$ ) into the LLM's embedding space ( $e_v$ ).
Loss Function: Unlike standard training which minimizes only Cross-Entropy ( $L_{CE}$ $L_{C E}$ ) on target text, VISTA introduces an auxiliary SAE Reconstruction Loss ( $L_{SAE}$ $L_{S A E}$ ).
- A pre-trained SAE, trained on the LLM's residual stream using only text data, is used as a constraint.
- The loss minimizes the Mean Squared Error between the projected visual tokens and their reconstruction by the text SAE: $L_{SAE} = \frac{1}{|M|} \sum_{l \in M} ||SAE_l(e^l_v) - e^l_v||^2_2$ .
- The total objective is $L_{total} = L_{CE} + L_{SAE}$ .
Mechanism: This forces visual tokens to lie on the manifold defined by the text SAEs. Consequently, visual inputs can be interpreted via pre-labeled linguistic SAE concepts with negligible additional computational cost, bypassing the need for vision-specific SAE training.

Fine-Grained Steering

Once aligned, visual manipulation becomes an arithmetic operation within the linguistic manifold. The framework enables localized steering by adding or subtracting decoder directions (concepts) of the text SAE to specific visual token positions ( $v_{source}$ ):

Removal: $v_{steered} = v_{source} - \beta \cdot v_{remove}$
Replacement: $v_{steered} = v_{source} + \alpha \cdot v_{add}$
This allows for the removal or replacement of specific objects in the model's perception while preserving the surrounding scene context.

3. Key Contributions

Interpretability Transfer: The introduction of a training objective that aligns visual tokens with LLM SAE features, facilitating cross-modal interpretability transfer without the computational burden of training or labeling vision-specific SAEs.
Encoder Analysis: A rigorous comparison of visual encoders enabled by VISTA, revealing that DINOv2 achieves superior and architecture-stable spatial localization. Unlike contrastive models like CLIP, which prioritize global summaries and suffer from spatial confusion, DINOv2's self-supervised local-to-global objective preserves fine-grained local information required for patch-level accuracy.
Causal Validation: Demonstration that arithmetic interventions on visual tokens using text-SAE directions produce coherent, localized changes in model output. This provides causal evidence that visual tokens inhabit the text SAE manifold.

4. Results and Findings

Alignment and Matching Rate

Matching Rate: VISTA constraints lead to a threefold increase in the matching rate (the accuracy with which activating textual concepts correspond to semantic elements in the image) compared to baselines without constraints.
Delayed Alignment: The paper challenges the notion of "delayed alignment" (where visual tokens naturally align with text features only in deeper layers). While this phenomenon was observed in Gemma models without constraints, it does not occur universally (e.g., in LLaMA-3.1-8B). VISTA constraints force immediate alignment across all architectures and layers.
Reconstruction and Sparsity: With SAE constraints, visual tokens achieve immediate convergence in reconstruction error and sparsity, matching the distribution of text tokens from the first layer, whereas unconstrained tokens often remain dense and high-error throughout the network.

Spatial Localization

DINOv2 vs. CLIP: DINOv2 demonstrates a Spatial Localization Accuracy (SLA) of 0.927 (with Gemma-2-2B), significantly outperforming CLIP (0.201) and I-JEPA (0.702).
CLIP Limitations: CLIP activations exhibit spatial confusion, often activating concepts far from the relevant object, making fine-grained steering unstable.
Robustness: DINOv2 maintains high localization accuracy across different LLM backbones (Gemma-2-2B, Gemma-2-9B, LLaMA-3.1-8B), whereas CLIP's performance varies significantly and remains lower.

Steering Performance

Quantitative Gains: Leveraging DINOv2's precision, VISTA achieves a 35% improvement in object removal tasks and a 47% improvement in object replacement tasks over the best vision-only baselines (VisionSAE and PatchSAE).
Qualitative Outcomes: Steering with DINOv2 allows for precise modifications (e.g., changing a "running dog" to a "running cat" or removing a cat from a couch) while preserving global scene context. In contrast, CLIP-based steering often results in hallucinations or global context alterations.
Causal Evidence: Controlled comparisons confirm that VISTA modifies the model's perception of specific visual patches rather than hijacking the language generation pathway. Text-only steering or prompt steering fails to preserve scene coherence.

Performance Impact

Small Models: In smaller models (e.g., Gemma-2-2B), SAE constraints induce a marginal accuracy regression on some benchmarks, likely due to limited representational capacity.
Large Models: In larger architectures (e.g., LLaMA-3.1-8B), SAE constraints act as a beneficial regularizer, improving performance across the majority of benchmarks.
OOD Robustness: VISTA's alignment generalizes to Out-of-Distribution (OOD) data (DomainNet), maintaining consistent reconstruction, sparsity, and matching rates, indicating the alignment is a structural property rather than a training-set artifact.

5. Significance and Claims

The paper claims that interpretability is a transferable property across modalities. By enforcing SAE constraints during the alignment of vision encoders to LLMs, VISTA maps visual tokens into a pre-existing, labeled linguistic feature space. This eliminates the need for the computational and labeling costs associated with training vision-specific SAEs.

The authors emphasize that the success of localized interpretability and steering is intrinsically coupled with the spatial fidelity of the vision backbone. They demonstrate that DINOv2's local objective provides the necessary precision for this transfer, significantly outperforming contrastive models like CLIP. This "interpretability transfer" paradigm offers a scalable path for understanding and controlling complex multimodal systems, enabling fine-grained causal interventions and feature circuit analysis without the prohibitive costs of retraining foundational models.

The work is validated across multiple LLM architectures (Gemma, LLaMA) and visual encoders (CLIP, DINOv2, I-JEPA), with code made available to support further research into mechanistic interpretability in multimodal systems.