Beyond Visual Memory: Mechanistic Diagnostics of Latent Visual Reasoning 事件

Beyond Visual Memory: Mechanistic Diagnostics of Latent Visual Reasoning arXiv:2606.01287v1 Announce Type: new Abstract: Recent latent visual reasoning methods achieve substantial gains by inserting continuous latent tokens into multimodal language models. These gains are commonly attributed to the tokens encoding visual evidence; recent analyses, however, reveal a paradox: the tokens are loosely tied to the image and contribute little to the answer. Critically, these analyses treat latent toke