AI summaryⓘ
The authors developed MMDiff, a new method to understand and control how multimodal large language models (MLLMs) handle visual information. They use special sparse autoencoders to find specific internal features that change when the models learn from images and text together. MMDiff can identify which features are important for tasks like reading text in images or understanding spatial info, and it can also tweak these features to improve or weaken certain model behaviors. Their tests show that this approach works on several MLLM types, helping to make models safer and better at visual tasks without hurting question-answering abilities. This means their method can help researchers both study and steer complex multimodal models more effectively.
Multimodal Large Language ModelsSparse AutoencodersFeature IsolationModel DiffingContrastive AnalysisVisual-Spatial UnderstandingOptical Character Recognition (OCR)Model SteeringMultimodal SafetyInterpretability
Authors
Hunar Batra, Lachin Naghashyar, Ashkan Khakzar, Philip Torr, Christian Schroeder de Witt, Constantin Venhoff, Ronald Clark
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 that are 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. We 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: (i) feature isolation, by diffing a base-LM SAE against its multimodal-adapted counterpart to identify features altered by multimodal training; (ii) task-specific feature detection, via per-token contrastive firing analysis that isolates causal features; and (iii) feature-level control, by causally removing or steering the discovered feature directions. We train multimodal SAEs for three MLLM families, LLaVA-MORE, PaliGemma 2, and InternVL3.5, and evaluate on visual-spatial understanding, multimodal safety, and OCR. MMDiff discovers sparse, causally specific features whose removal selectively degrades target behaviors by an average of 12% on spatial tasks and 17% on OCR, and reduces attack success rate by 24% on multimodal safety attacks, with no impact on VQA performance. Steering these features improves spatial and OCR accuracy by +3.6% and +1.8% on average over a standard single-layer steering baseline. These results show that multimodal SAEs can serve not only as interpretability tools, but as mechanisms for auditing, steering, and controlling MLLMs behavior toward safer and more capable generations.