Multimodal Model Diffing for Feature Discovery and Control

📄 arXiv: 2608.09928v1 📥 PDF

作者: Hunar Batra, Lachin Naghashyar, Ashkan Khakzar, Philip Torr, Christian Schroeder de Witt, Constantin Venhoff, Ronald Clark

分类: cs.CV, cs.AI, cs.CL, cs.LG

发布日期: 2026-08-10

备注: Preprint. Accepted at ICML 2026 Trustworthy AI for Good Workshop


💡 一句话要点

提出MMDiff框架以发现和控制多模态特征

🎯 匹配领域: 支柱九:具身大模型 (Embodied Foundation Models)

关键词: 多模态学习 特征控制 稀疏自编码器 模型审计 视觉理解 任务特定检测 因果特征

📋 核心要点

  1. 现有方法难以识别和控制多模态大型语言模型的内部特征,导致审计和控制能力不足。
  2. 本文提出MMDiff框架,通过训练多模态稀疏自编码器,提供特征级接口以发现和控制多模态行为。
  3. 实验结果显示,MMDiff在空间任务和OCR上分别提升了3.6%和1.8%的准确率,同时降低了多模态安全攻击的成功率。

📝 摘要(中文)

多模态大型语言模型(MLLMs)展现出强大的视觉理解能力,但其内部特征的识别、审计和控制仍然困难。本文提出MMDiff,一个多模态模型差异化框架,通过训练稀疏自编码器(SAEs)来创建特征级接口,以发现和控制多模态行为。MMDiff支持特征隔离、任务特定特征检测和特征级控制。实验表明,MMDiff能够发现稀疏且因果特征,其去除会导致空间任务和OCR性能平均下降12%和17%,而在多模态安全攻击中成功率降低24%。

🔬 方法详解

问题定义:本文旨在解决多模态大型语言模型内部特征难以识别和控制的问题。现有方法无法有效隔离多模态训练所改变的特征,限制了对模型行为的审计和控制。

核心思路:MMDiff框架通过训练多模态稀疏自编码器(SAEs),将模型的隐藏状态转化为可解释的特征方向,从而实现特征的隔离和控制。该设计使得研究者能够更好地理解和操控模型的多模态行为。

技术框架:MMDiff的整体架构包括三个主要模块:特征隔离、任务特定特征检测和特征级控制。特征隔离通过对比基线SAE和多模态适应SAE来识别特征变化;任务特定特征检测利用逐标记对比分析来隔离因果特征;特征级控制则通过去除或引导发现的特征方向来实现。

关键创新:MMDiff的主要创新在于其能够通过特征级接口实现对多模态行为的审计和控制,这与现有方法的特征提取和控制机制有本质区别。

关键设计:在训练过程中,MMDiff采用了特定的损失函数来优化SAEs的稀疏性和可解释性,同时设计了针对不同多模态模型(如LLaVA-MORE、PaliGemma 2和InternVL3.5)的适应性策略,以确保特征的有效性和准确性。

🖼️ 关键图片

fig_0
fig_1
fig_2

📊 实验亮点

实验结果表明,MMDiff能够发现稀疏且因果特征,其去除导致空间任务性能平均下降12%,OCR性能下降17%,同时在多模态安全攻击中成功率降低24%。此外,特征引导的准确率在空间任务和OCR上分别提升了3.6%和1.8%。

🎯 应用场景

MMDiff框架具有广泛的应用潜力,尤其在多模态人工智能系统的开发中。它可以用于提高模型的安全性和可靠性,帮助研究者理解模型决策过程,并在实际应用中实现更精确的控制。这一研究为未来的多模态模型设计提供了新的思路和工具。

📄 摘要(原文)

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.