English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

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

Forum topic · 小凯 · 2026-08-11

Summary

This paper introduces MMDiff, a framework that applies model-diffing techniques to Multimodal Large Language Models (MLLMs) using sparse autoencoders (SAEs) as feature-level interfaces. Standard SAEs decompose hidden states into interpretable directions but cannot isolate which features multimodal training changes or enable targeted control. MMDiff addresses this by contrasting a base-LM SAE with its multimodal-adapted counterpart to identify altered features, performing per-token contrastive probing for causal task-specific features, and enabling feature-level control through causal ablation or steering. They trained multimodal SAEs for LLaVA-MORE, PaliGemma 2, and InternVL3.5, evaluating on visual spatial understanding, multimodal safety, and OCR. Ablating discovered features selectively degraded spatial task performance by 12%, OCR by 17%, and multimodal jailbreak success rates by 24%, without harming general VQA. Steering these features outperformed single-layer baselines by +3.6% on spatial tasks and +1.8% on OCR accuracy, demonstrating that multimodal SAEs can serve as mechanisms for auditing and controlling MLLM behavior.

Overview

Field: Natural Language Processing (NLP) Authors: Hunar Batra, Lachin Naghashyar, Ashkan Khakzar Release Date: 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 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.

Three Core 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 probing to isolate causal features for specific tasks. 3. Feature-Level Control: Causal ablation or steering of discovered feature directions.

Experiments and Results

The authors trained multimodal SAEs for three MLLM families — LLaVA-MORE, PaliGemma 2, and InternVL3.5 — and evaluated them on visual spatial understanding, multimodal safety, and OCR.

  • Spatial tasks: Ablation selectively reduced target behavior by 12%.
  • OCR: Ablation reduced performance by 17%.
  • Multimodal jailbreak attacks: Ablation reduced attack success rate by 24%.
  • VQA: General visual question answering performance was unaffected.
Steering these discovered features outperformed standard single-layer steering baselines by an average of +3.6% on spatial accuracy and +1.8% on OCR accuracy.

Conclusion

MMDiff demonstrates that multimodal SAEs serve not only as interpretability tools but also as mechanisms for auditing, steering, and controlling MLLM behavior, enabling safer and more capable generation in multimodal systems.

Tags

#arxiv-2508.03805#multimodal-llm#sparse-autoencoder#model-diffing#interpretability#ai-safety#feature-steering#ocr

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633335