Towards Understanding Multimodal Fine-Tuning: A Case Study into Spatial Features

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

Keywords: Sparse Autoencoders, Vision transformers, Causal interventions

TL;DR: We apply stage-wise model diffing with sparse autoencoders to multimodal fine-tuning and show how vision-language models develop spatially grounded features and the attention heads that drive them.

Abstract: Vision–Language Models (VLMs) demonstrate strong performance on a wide range of tasks by fine-tuning pretrained language backbones to process projected visual tokens alongside text. Yet despite these empirical gains, it remains unclear how backbone representations adapt during multimodal training and when vision-specific capabilities emerge. In this work, we present the first mechanistic analysis of VLMs adaptation with stage-wise model diffing, a technique that isolates representational changes introduced during multimodal fine-tuning to reveal how a language model learns to "see". Concretely, we fine-tune sparse autoencoders trained on LLaMA-3.1-8B over multimodal activations from LLaVA-More (based on LLaMA-3.1-8B) using 50k VQAv2 pairs. We first isolate vision-preferring features that appear or reorient during multimodal fine-tuning. We then test for spatial selectivity using a controlled shift to spatial prompts and use attribution patching to identify the attention heads that causally activate these units. Our findings show that stage-wise model diffing reveals when and how spatially grounded multimodal features arise. It also provides a clearer view of modality fusion by showing how visual grounding reshapes features that were previously text-only. This methodology enhances the interpretability of multimodal training and provides a foundation for refining training regimes as well as auditing and steering models in safety-critical or domain-specific settings.

Submission Number: 325