Paper Overview
Field: NLP Authors: Hunar Batra, Lachin Naghashyar, Ashkan Khakzar arXiv: 2508.03805
Abstract
Multimodal Large Language Models (MLLMs) exhibit strong visual understanding, yet the internal features that cause these behaviors remain difficult to identify, audit, or control. While applicable to post-hoc inspection, hidden states decomposed into interpretable feature directions using sparse autoencoders (SAEs) neither readily isolate which features are changed by multimodal training, nor are they directly useful for targeted control. The authors introduce MMDiff, a multimodal model-diffing framework that trains multimodal SAEs and turns them into feature-level interfaces for discovering and controlling multimodal behavior.
Key Capabilities
MMDiff supports three uses:
1. Feature isolation — diffing a base-LM SAE against its multimodal-adapted counterpart to identify features altered by multimodal training. 2. Task-specific feature detection — per-token contrastive activation analysis to isolate causal features. 3. Feature-level control — causal ablation or steering of the discovered feature directions.
Experiments and Results
The authors trained multimodal SAEs for three MLLM families — LLaVA-MORE, PaliGemma 2, and InternVL3.5 — and evaluated on visual spatial understanding, multimodal safety, and OCR.
- Removing discovered features selectively degraded target behavior by 12% on spatial tasks and 17% on OCR, and reduced attack success rate by 24% on multimodal safety attacks, without affecting VQA performance.
- Steering these features improved spatial and OCR accuracy by +3.6% and +1.8% on average over standard single-layer steering baselines.
Conclusion
These results show that multimodal SAEs can serve not only as interpretability tools but also as mechanisms for auditing, steering, and controlling MLLM behavior toward safer and more capable generation.