MMDiff: A Multimodal Model-Diffing Framework for Feature Discovery and Control in MLLMs

Forum topic · 小凯 · 2026-08-11

Summary

This paper introduces MMDiff, a multimodal model-diffing framework that trains multimodal sparse autoencoders (SAEs) and turns them into feature-level interfaces for discovering and controlling multimodal behavior in Multimodal Large Language Models (MLLMs). It addresses the difficulty of identifying, auditing, and controlling internal features that drive visual understanding in MLLMs, where standard SAEs cannot easily isolate features changed by multimodal training or enable targeted control. MMDiff supports three uses: feature isolation by diffing a base-LM SAE against its multimodal-adapted counterpart, task-specific feature detection via per-token contrastive triggering analysis, and feature-level control through causal ablation or steering along discovered directions. The authors train multimodal SAEs for three MLLM families (LLaVA-MORE, PaliGemma 2, InternVL3.5) and evaluate on visual spatial understanding, multimodal safety, and OCR. Causal ablation of discovered sparse features reduces target behavior by 12% on spatial tasks, 17% on OCR, and 24% on multimodal safety attacks without harming VQA performance, while steering outperforms single-layer baselines by +3.6% on spatial and +1.8% on OCR accuracy.

Paper Overview

Field: NLP
Authors: Hunar Batra, Lachin Naghashyar, Ashkan Khakzar
Posted: 2026-08-11
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 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 triggering analysis to isolate causal features; and (iii) feature-level control, through causal ablation or steering along 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 ablation selectively degrades target behavior by 12% on spatial tasks, 17% on OCR, and 24% on multimodal safety attack success rate, without affecting VQA performance. Steering these features yields average gains of +3.6% on spatial and +1.8% on OCR accuracy over standard single-layer steering baselines. These results indicate that multimodal SAEs can serve not only as interpretability tools but also as mechanisms to audit, steer, and control MLLM behavior for safer and more capable generation.

Key Contributions

MMDiff framework: A multimodal model-diffing approach that trains multimodal SAEs and converts them into a feature-level interface for both interpretability and control of MLLMs.
Feature isolation: Compares a base-LM SAE with its multimodal-adapted counterpart to identify features specifically altered by multimodal training.
Task-specific feature detection: Uses per-token contrastive triggering analysis to isolate causally relevant features for a given task.
Feature-level control: Supports both causal ablation and steering along discovered feature directions for targeted behavioral modification.
Broad evaluation: Validated on three MLLM families (LLaVA-MORE, PaliGemma 2, InternVL3.5) across visual spatial understanding, multimodal safety, and OCR.

Key Results

Causal ablation of discovered sparse features reduces target behavior by 12% on spatial tasks and 17% on OCR.
Multimodal safety attack success rate drops by 24% after ablating safety-related features, without harming VQA performance.
Steering along discovered feature directions outperforms standard single-layer steering baselines by +3.6% on spatial accuracy and +1.8% on OCR accuracy on average.
Results suggest multimodal SAEs can function as practical tools for auditing, steering, and controlling MLLM behavior to enable safer and more capable generation.