Multimodal Model Diffing for Feature Discovery and Control

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Hunar Batra, Lachin Naghashyar, Ashkan Khakzar, Philip Torr, Christian Schroeder de Witt, Constantin Venhoff, Ronald Clark
• Venue: arXiv: cs.CV
• Published: 2026-08-10
• Collected: 2026-08-11
• URL: http://arxiv.org/abs/2608.09928v1
• PDF: https://arxiv.org/pdf/2608.09928v1
• DOI: UNKNOWN
• Registry group: academic_apis_and_preprints
• Evidence role: canonical_academic_index
• Language: en
• Country / region: UNKNOWN

Plain-Language Brief

Researchers developed a framework called MMDiff to help understand, control, and audit the behavior of Multimodal Large Language Models (MLLMs). The framework identifies features that are altered by multimodal training, isolates causal features for specific tasks, and allows for controlling these discovered feature directions. The results show that removing or steering these features can improve performance on spatial tasks, OCR, and reduce the success rate of multimodal safety attacks.

What the Document Says

• The research presents a framework called MMDiff to work with Multimodal Large Language Models (MLLMs).
• MMDiff is used for feature isolation, task-specific feature detection, and feature-level control.
• Feature isolation involves comparing a base-LM SAE with its multimodal-adapted counterpart to identify features altered by multimodal training.
• Task-specific feature detection uses per-token contrastive firing analysis to isolate causal features.
• Feature-level control allows for removing or steering discovered feature directions.
• The researchers trained MMDiff on three MLLM families: LLaVA-MORE, PaliGemma 2, and InternVL3.5.
• Evaluations were conducted on visual-spatial understanding, multimodal safety, and OCR tasks.
• Removing or steering the discovered features degrades target behaviors by an average of 12% on spatial tasks and 17% on OCR.
• It reduces attack success rate by 24% on multimodal safety attacks without impacting 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.
• The results suggest that MLLMs can be audited, controlled, and made safer with the help of multimodal SAEs.

Key Concepts Mentioned

• Multimodal Large Language Models (MLLMs)
• feature isolation
• task-specific feature detection
• feature-level control
• sparse autoencoders (SAEs)
• evaluation
• visual-spatial understanding
• multimodal safety
• OCR

Claims Made by the Document

• MMDiff can identify features altered by multimodal training.
• MMDiff can isolate causal features for specific tasks.
• Removing or steering the discovered feature directions degrades target behaviors on spatial tasks and OCR.
• Steering these features improves spatial and OCR accuracy over a standard single-layer steering baseline.

Evidence Used

• The paper presents a framework called MMDiff for Multimodal Large Language Models (MLLMs).
• MMDiff's three uses are explained: feature isolation, task-specific feature detection, and feature-level control.
• Evaluations results show the impact of removing or steering the discovered features on spatial tasks, OCR, and multimodal safety attacks.

Methods / Approach

• Multimodal SAEs training
• per-token contrastive firing analysis

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

The study only evaluates the performance on three MLLM families. It's unclear if the results generalize to other MLLMs.

Open Questions

• How can the framework be applied to other Multimodal Large Language Models besides the ones mentioned in the paper?
• What are the potential long-term implications of using MMDiff for auditing, controlling, and making safer MLLMs?

Notes for Future Q&A

May include interpretation. Labeled where it does.

• The research focuses on Multimodal Large Language Models, which are models that exhibit strong visual understanding capabilities. These models can process both text and visual data simultaneously.
• Sparse autoencoders (SAEs) are used to decompose hidden states into interpretable feature directions.

Raw Abstract / Excerpt

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.