First-hand information for everyone
Researchers introduce MMDiff, a new way to find and change the internal visual features that drive multimodal large language models (MLLMs). Many of these models understand images well, but it is hard to see which internal signals—or features—cause those behaviors. The authors build a tool that makes those hidden features easier to identify, test, and manipulate.
The core idea is to train sparse autoencoders (SAEs) on model hidden states and then use comparisons, or “diffing,” between a text-only base model and its multimodal-adapted version. A sparse autoencoder is a small model that compresses internal activations into a few interpretable directions. By comparing the SAE for the base language model to the SAE for the multimodal model, MMDiff highlights which features changed during multimodal training. The method also includes a per-token contrastive firing analysis, which spots features that reliably “fire” for particular words or tokens, and a set of interventions that can remove or steer those feature directions to test causality.
The team trained multimodal SAEs for three families of MLLMs: LLaVA-MORE, PaliGemma 2, and InternVL3.5. They used these learned feature interfaces to study three kinds of behavior: visual-spatial understanding (for example reasoning about positions and relationships in images), multimodal safety (resistance to unsafe or adversarial image+text prompts), and optical character recognition (OCR, reading text in images).
MMDiff found sparse features that appear to be causally linked to specific behaviors. Removing those features selectively degraded the target behaviors by measurable amounts: on average about 12% worse on spatial tasks and 17% worse on OCR. The same removals cut attack success rates on multimodal safety tests by about 24%. Importantly, these targeted removals did not harm performance on visual question answering (VQA). Steering the discovered features—nudging them rather than removing them—improved spatial and OCR accuracy by about +3.6% and +1.8% on average compared with a standard single-layer steering baseline.
Why this matters: MMDiff turns an interpretability technique into a practical interface for auditing and controlling multimodal models. It offers a way to probe what changed when models were made multimodal, find features tied to specific behaviors, and directly test how changing those features affects outputs. Important caveats are that the results are reported for the three tested model families and for the particular tasks and attacks the authors used. The paper frames the approach as a post-hoc inspection and control tool; generalization to other models, behaviors, or deployment settings would need further study.