# Steering Vision-Language Models with Joint Sparse Autoencoders

arXiv: https://arxiv.org/abs/2606.25657 · HTML: https://mlsanity.com/paper/2606.25657/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Placed in VLM benchmarks and visual reasoning, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 18 of its 28 references that are in the map.
- First posted on arXiv 24 Jun 2026.

## Placed in

- Multimodal LLMs / vision-language models / [VLM benchmarks and visual reasoning](https://mlsanity.com/mllm/vlm-benchmarks-and-visual-reasoning/)
- Mechanistic interpretability / [Interpreting vision and multimodal models](https://mlsanity.com/mechinterp/interpreting-vision-and-multimodal-models/)

## Builds on (most cited first, 18 of 18)

- [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://mlsanity.com/paper/1810.04805/) (2018)
- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [Improved Baselines with Visual Instruction Tuning](https://mlsanity.com/paper/2310.03744/) (2023)
- [Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution](https://mlsanity.com/paper/2409.12191/) (2024)
- [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://mlsanity.com/paper/2102.05918/) (2021)
- [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://mlsanity.com/paper/2305.06500/) (2023)
- [MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models](https://mlsanity.com/paper/2304.10592/) (2023)
- [Locating and Editing Factual Associations in GPT](https://mlsanity.com/paper/2202.05262/) (2022)
- [PaLM-E: An Embodied Multimodal Language Model](https://mlsanity.com/paper/2303.03378/) (2023)
- [Microsoft COCO Captions: Data Collection and Evaluation Server](https://mlsanity.com/paper/1504.00325/) (2015)
- [Sparse Autoencoders Find Highly Interpretable Features in Language Models](https://mlsanity.com/paper/2309.08600/) (2023)
- [Steering Language Models With Activation Engineering](https://mlsanity.com/paper/2308.10248/) (2023)
- [LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model](https://mlsanity.com/paper/2304.15010/) (2023)
- [Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models](https://mlsanity.com/paper/2403.19647/) (2024)
- [Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders](https://mlsanity.com/paper/2407.14435/) (2024)
- [Improving Dictionary Learning with Gated Sparse Autoencoders](https://mlsanity.com/paper/2404.16014/) (2024)
- [Weight-sparse transformers have interpretable circuits](https://mlsanity.com/paper/2511.13653/) (2025)
- [Decomposing multimodal embedding spaces with group-sparse autoencoders](https://mlsanity.com/paper/2601.20028/) (2026)

## Built on by (most cited first, 4 of 4)

- [Prior Directions: Why GUI Grounding Gets Locked in the Past](https://mlsanity.com/paper/2607.26913/) (2026)
- [Taming CoT Obfuscation in VLMs: From Mechanistic Evidence to Activation Enforcement](https://mlsanity.com/paper/2609.24243/) (2026)
- [Geometric Similarity in VLM Low-Level Vision Representations](https://mlsanity.com/paper/2610.00848/) (2026)
- [Cross-Modal Latent Influence: Patch-Level Training Sample Attribution for Vision-Language Models](https://mlsanity.com/paper/2610.14190/) (2026)

## Abstract (arXiv)

Sparse Autoencoders (SAEs) have shown promise for analyzing language models, but applying them to vision-language models (VLMs) often yields representations that are difficult to use as controllable cross-modal steering directions. We introduce the Joint Sparse Autoencoder (JSAE), which uses an explicit alignment constraint to jointly factorize sequence-pooled vision and language activations into shared, interpretable image/caption-level features. Applied to LLaVA, JSAE recovers cross-modal features for recognizable concepts (e.g., food and animals). Through bidirectional interventions (additive steering and suppression), we observe a layer-dependent asymmetry under our protocol: additive steering peaks at mid-to-late (pre-output) layers and weakens at both ends, whereas suppression scores remain within a comparable range across all probed layers within statistical noise. Experiments on three VLMs, namely LLaVA-v1.6-Mistral-7B, Llama3-LLaVA-8B, and the MoE-based Qwen3-VL-30B, show related layer-localized effects across architectures. Together, these results suggest that explicitly aligned sparse representations support more controllable intervention-based analysis of multimodal features, within an identifiable layer range, than the unconstrained alternatives tested here.