Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Models

In three linesResearchers use Transcoders to interpret how vision-language models transform images into text. Applied to Gemma 3-4B-IT, the framework decomposes the model into computational pathways linking image patches to token generation. Transcoder attributions outperform SAEs in identifying hallucinations (AUC 0.68).

Summary generated by Claude — human-verified