Multimodal Model Diffing for Feature Discovery and Control

2026-08-10

AI Summary

This work addresses the limited interpretability and controllability of internal mechanisms driving visual understanding in multimodal large language models. To this end, we propose MMDiff, a novel framework that, for the first time, integrates model differencing with multimodal sparse autoencoders to enable feature-level behavior discovery and intervention. Our approach facilitates task-relevant feature identification, causal manipulation, and feature isolation, supported by token-wise contrastive activation analysis and directional feature intervention techniques. Experimental results demonstrate that MMDiff incurs only a modest performance drop of 12%–17% on spatial reasoning and OCR tasks while reducing the success rate of multimodal safety attacks by 24%. Furthermore, feature-guided interventions improve accuracy by 1.8%–3.6% without compromising performance on visual question answering (VQA) benchmarks.

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.

Problem

• Multimodal Large Language Models
• feature discovery
• model interpretability
• behavior control
• multimodal safety

Innovation

• Multimodal Model Diffing
• Sparse Autoencoders
• Feature-level Control
• Causal Feature Detection
• Multimodal Large Language Models