Research map / Vision-language and multimodal / Multimodal LLMs

MMDiff: Multimodal Model Diffing for Feature Discovery and Control

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MMDiff, a model-diffing pipeline that warm-starts multimodal SAEs from base-LM SAEs, isolates vision-adapted features by decoder rotation and visual energy, and uses per-token contrastive firing to find task-specific features for causal ablation and steering (MMDiff-CAA) on LLaVA-MORE, PaliGemma 2, and InternVL3.5.

Problem

MLLM internal features that drive visual behaviors are hard to identify, audit, or control, and standard SAEs trained on MLLM activations mix inherited language-backbone features with those altered by multimodal training.

Approach

Diff a base-LM SAE against its multimodally adapted counterpart (warm-started, text-only masked training) to find features that both rotate in decoder space and respond to visual input; then use per-token contrastive firing with lexical-invariance filtering to select task-specific features for causal removal and steering.

Key results

• Feature-level ablation of top spatial features lowers VSR accuracy by about 10-15% on average across the three MLLMs with |ΔVQA| ≤ 1.5%.
• Ablating MMDiff-discovered unsafe features reduces VLSBench ASR by 9.67% on average (per-feature 17-28%) with no VQA cost.
• MMDiff-CAA improves spatial steering by +3.6% and OCR steering by +1.8% on average over vanilla CAA.

Surprising findings

• Standard SAE trained directly on MLLM activations yields top spatial features whose ablation leaves VSR unchanged (mean +0.22), indicating the causally effective set requires diffing.
• Instruction tuning amplified causal effects of spatial features roughly 3x, and two features reversed sign after instruction tuning.

Limitations

• Safety and OCR evaluated only on PaliGemma 2; larger or other MLLM families not covered.
• MMDiff-CAA requires an instruction-tuned reference model to extract steering directions.
• A minority of safety candidates cause generation collapse rather than refusal when ablated.

Experimental setup

Datasets: VQAv2 VSR VLSBench MSSBench OCRBench

Benchmarks: VSR VQAv2 VLSBench MSSBench OCRBench

Metrics: accuracy delta attack success rate odds ratio FVU

Baselines: vanilla CAA random feature ablation standard SAE trained on MLLM activations

Builds on: stage-wise model diffing TopK SAE JumpReLU SAE contrastive activation addition

Closest papers by meaning

• Multi-Layer Visual Feature Fusion in Multimodal LLMs: Methods, Analysis, and Best Practices (CVPR 2025 · similarity 0.88)
• Where MLLMs Attend and What They Rely On: Explaining Autoregressive Token Generation (CVPR 2026 · similarity 0.88)
• OddGridBench: Exposing the Lack of Fine-Grained Visual Discrepancy Sensitivity in Multimodal Large Language Models (CVPR 2026 · similarity 0.88)
• Predictive Regularization Against Visual Representation Degradation in Multimodal Large Language Models (CVPR 2026 · similarity 0.88)
• Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings (NeurIPS 2025 · similarity 0.88)
• Implicit Multimodal Alignment: On the Generalization of Frozen LLMs to Multimodal Inputs (NeurIPS 2024 · similarity 0.88)