ResearchPod Summary
Multimodal Large Language Models (MLLMs) possess impressive visual capabilities, but their internal mechanisms remain largely opaque. Researchers struggle to identify which internal features drive specific behaviors, making it difficult to audit or steer these models without expensive retraining. This paper introduces MMDiff, a framework that uses sparse autoencoders (SAEs) to create a feature-level interface for MLLMs. By comparing SAEs trained on a base language model against those trained on the multimodal-adapted version, the authors isolate features that were specifically reshaped or repurposed by multimodal training. The pipeline involves three key steps: isolating visually responsive features, identifying task-specific subsets through contrastive firing analysis, and applying causal interventions to remove or steer these features.
MMDiff successfully identifies sparse, causally specific features across three MLLM families: LLaVA-MORE, PaliGemma 2, and InternVL3.5. The authors demonstrate that these features are not just interpretability artifacts but functional control handles. Causal removal of these features selectively degrades target behaviors—reducing spatial reasoning accuracy by 12%, OCR accuracy by 17%, and multimodal safety attack success rates by 24%—without negatively impacting general visual question answering performance. Furthermore, the authors introduce MMDiff-CAA, a steering method that combines backbone activation steering with the injection of specific SAE decoder directions. This approach improves spatial and OCR accuracy by an average of 3.6% and 1.8%, respectively, outperforming standard single-layer steering baselines.
This work provides a bridge between interpretability and model control. By demonstrating that multimodal SAEs can serve as mechanisms for auditing and steering, the authors offer a path toward safer and more reliable MLLMs. The ability to selectively suppress unsafe behaviors or enhance specific capabilities like spatial reasoning without retraining the entire model is a significant step forward for model governance and alignment.
Alex: Welcome to another episode of ResearchPod. Today, we're looking at a paper that asks a deceptively simple question: when an AI model that handles both images and text makes a mistake, how do you find and fix the exact internal part responsible — without rebuilding the whole system from scratch?
Sam: That sounds like a genuinely hard problem. Why is the usual fix so costly?
Alex: Because these models are enormous. Retraining one from scratch to correct a single flaw takes weeks of computing time and enormous expense — and it risks breaking skills the model already had. It's like tearing down a wall to fix one bad pipe, and accidentally knocking out the electricity in the process.
Sam: So the researchers are looking for a more targeted approach. What's their starting point?
Alex: They start by trying to understand what actually changed inside the model when it learned to handle images. They call this "model diffing." Think of it like comparing two blueprints of the same building — one from before a renovation, one from after — to see exactly what was added or moved. Here, the "before" is a text-only AI, and the "after" is the same model once it's been trained to understand images as well.
Sam: So rather than just looking at what the model outputs, they're comparing its internal activity at each stage of processing.
Alex: Right. To do that, they attach small helper networks to the model's internal states. These helpers are trained to compress the model's activity into a set of clean, distinct directions — like taking a massive, chaotic library and sorting every book onto labelled shelves, so you can actually see what's where.
Sam: And once things are sorted, you can spot which shelves got reorganised when the visual wing was added.
Alex: Precisely. But there's a subtlety worth flagging. The model processes both text and images, so how do you make sure you're only identifying features that genuinely changed because of visual training — and not just background noise?
Sam: How do they handle that?
Alex: Two ways. First, they train the helper networks using text tokens only. That keeps the internal measuring stick anchored to the original language model, so any shift they detect is specifically caused by the visual training — not by general drift. Like calibrating your ruler before you measure anything.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.
Sam: That makes sense. What's the second step?
Alex: They use two signals to score each internal feature. One measures how strongly a feature activates when an image is present. The other measures how much that feature's internal wiring actually shifted during training. A feature that scores high on both has genuinely been repurposed to handle visual input — rather than just reusing old text habits.
Sam: So you're not just looking for features that react to images — you're looking for ones that were rebuilt to handle them.
Alex: Exactly. And there's one more filtering step. They check whether a feature fires because of the image itself, or just because of how the question is phrased. If swapping the wording changes the response but the image stays the same, that feature isn't really doing visual reasoning — it's doing language pattern-matching.
Sam: So they end up with a set of features that are genuinely tied to what the model sees, not how the prompt is worded.
Alex: Right. And once those features are identified, researchers can intervene on them directly — nudging the model's behavior by adjusting specific internal directions during generation, without touching anything else.
Sam: What does that look like in practice?
Alex: Imagine you've found the exact shelf in the library that handles spatial reasoning — like understanding whether an object is to the left or right of another. You can quietly reorganise just that shelf, and the rest of the library stays intact. The paper reports that focusing on these specific adapted units produces predictable shifts in behavior without disrupting the wider network.
Sam: Though I imagine it's not always that clean.
Alex: It isn't, and the authors are upfront about that. A small number of safety-targeted interventions — where you try to suppress an unsafe behavior — can occasionally cause the model to stop producing coherent sentences altogether, rather than simply declining the request. Push too hard on the wrong feature, and you can destabilise the surrounding language generation.
Sam: So there's a real risk of overcorrection.
Alex: There is. The authors also note that the whole framework depends on having a responsibly trained reference model to begin with. You need a good map before you can do good surgery.
Sam: When you step back, what's the broader significance here?
Alex: The study suggests that when a model learns to handle images, the changes don't spread evenly across the entire system. Instead, visual reasoning concentrates into a compact, identifiable set of internal circuits. That's a meaningful finding — it means these systems are more structured than they might appear, and that structure can be used.
Sam: And if you can identify those circuits reliably, you could potentially monitor them in real time — catching errors before they reach the output.
Alex: That's exactly the direction the authors point toward. Models that can audit and correct their own visual reasoning during inference, rather than waiting for a human to catch the error after the fact. It's a step toward AI systems that are genuinely transparent from the inside.
Sam: A quiet but meaningful shift in how we might think about AI reliability.
Alex: Well put. Thanks for listening to ResearchPod.