Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Models
The paper uses transcoders to trace how image inputs shape token generation in a VLM, and finds the method helps explain visual grounding and hallucinations.
Intelligence analysis by GPT-5.4 Mini
The paper proposes a function-centric interpretability method for vision-language models, using transcoders as a causal proxy for layer-wise computation. On Gemma 3-4B-IT, it maps image patches to token-generation pathways and reports that these traces are more stable than SAE-based attributions for visually grounded tokens, while also supporting hallucination prediction.
A vision-language model is a computer program that looks at a picture and writes words about it. This paper tries to peek inside that program to see how the picture changes what it says.
The authors use a tool like a map of little roads inside the model. Each road shows how a part of the image can help choose the next word, like following crumbs from the picture to the sentence.
They also find that some of these road patterns can help guess when the model is making things up about the image. It is a bit like noticing that a storyteller is starting to wander off from the facts.
Analysis
What the paper does
The paper argues that existing interpretability methods for vision-language models miss an important part of the story: the updates inside the model that actually drive cross-modal interaction. Instead of only decomposing static residual representations with Sparse Autoencoders, it uses Transcoders, described here as sparse approximations of MLP sublayers that serve as a causal proxy for layer-wise computation.
What they found
Applied to Gemma 3-4B-IT, the method breaks the model into interpretable computational pathways that connect image patches to directions in token generation. The authors report that transcoder attributions show stronger and more stable effects on visually grounded tokens under patch ablation than SAE attributions, and that they line up better with semantically relevant image regions. They also run a False Visual Grounding counterfactual analysis and say the recovered pathways are specific to vision-language behavior.
Hallucination signal
The paper also studies hallucinated generations by extracting graph-based indicators from the circuit traces produced by the transcoders. A logistic classifier built on these mechanistic graph features reaches an AUC of 0.68 for predicting hallucinations. The result is modest rather than decisive, but it suggests the traces carry useful signal about when the model is likely to stray from the image.
Bottom line
The paper’s main claim is that a function-centric circuit decomposition can provide both interpretable and predictive views of multimodal computation. In plain terms, it tries to show not just what the model says, but which internal pathways turn a picture into words and when those pathways go wrong.
Key points
- The paper studies interpretability in vision-language models using transcoders rather than only static residual decompositions.
- It applies the method to Gemma 3-4B-IT and traces pathways from image patches to token generation.
- The authors say transcoder attributions are stronger and more stable than SAE attributions for visually grounded tokens.
- A graph-based classifier over circuit-trace features predicts hallucinations with an AUC of 0.68.
- The work frames transcoders as a function-centric way to explain multimodal computation.
