Paper Overview
Field: NLP Authors: Hunar Batra, Lachin Naghashyar, Ashkan Khakzar Published: 2026-08-11 arXiv: 2508.03805
Summary
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 hidden states decomposed into interpretable feature directions using sparse autoencoders (SAEs) are applicable to post-hoc inspection, they 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. 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 trigger analysis to isolate causal features. 3. Feature-level control — causal ablation or steering of the discovered feature directions.
Experiments and Results
Multimodal SAEs were trained for three MLLM families — LLaVA-MORE, PaliGemma 2, and InternVL3.5 — and evaluated on visual spatial understanding, multimodal safety, and OCR.
Key findings:
- MMDiff discovers sparse, task-specific features whose ablation selectively reduces targeted behavior:
- Spatial tasks: −12%
- OCR: −17%
- Multimodal safety attacks: attack success rate reduced by 24%, with no impact on VQA performance
- Steering the discovered features improves accuracy over standard single-layer steering baselines by +3.6% on spatial tasks and +1.8% on OCR, on average.
Conclusion
These results indicate that multimodal SAEs can serve not only as interpretability tools but also as mechanisms for auditing, steering, and controlling MLLM behavior toward safer, more capable generation.
Original paper: arXiv:2508.03805