Researchers introduced MMDiff, a multimodal model-diffing framework that trains multimodal sparse autoencoders (SAEs) to discover and control features in multimodal large language models (MLLMs). MMDiff supports feature isolation by diffing a base-LM SAE against its multimodal-adapted counterpart, task-specific feature detection via per-token contrastive firing analysis, and feature-level control through causal removal or steering of discovered feature directions. 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.
MMDiff is a framework for discovering and controlling internal features in multimodal large language models. It uses multimodal sparse autoencoders to isolate features altered by multimodal training, detect task-specific features, and enable causal control. The approach was validated on three MLLM families across tasks including visual-spatial understanding, multimodal safety, and OCR.
MMDiff leverages sparse autoencoders trained on multimodal model activations to decompose hidden states into interpretable feature directions. By comparing SAEs from a base language model and its multimodal-adapted version, it isolates features specifically introduced or modified by multimodal training. Per-token contrastive firing analysis identifies features causally linked to specific tasks, and causal interventions (removal or steering) demonstrate direct control over model behavior.
This work addresses a critical gap in AI safety and interpretability for multimodal systems, which are increasingly deployed in commercial applications. The ability to audit and control internal features could accelerate responsible deployment of MLLMs in sensitive domains like healthcare, autonomous driving, and content moderation. The framework's applicability across multiple model families suggests potential for standardization in model auditing tools.
MMDiff provides a pathway to safer and more controllable multimodal AI products, reducing risks of unintended behaviors and enabling fine-grained customization. This could lower barriers to enterprise adoption in regulated industries and create new markets for AI interpretability tools and services.
Next signals include open-sourcing of MMDiff code and SAE weights, integration with popular MLLM evaluation suites, and extension to other modalities (e.g., audio, video). Adoption by model developers for pre-deployment safety checks and by regulators for compliance auditing is plausible. Further research may explore real-time feature steering for dynamic behavior adjustment.