Multimodal Models Fake Their Visual Thinking

🔥 What happened Researchers at UT Austin and Meta AI found that latent visual states in multimodal LLMs are mostly decorative. Perturb the generated visual latents and the model's answer barely changes; perturb the text trace and performance tanks. 💡 Why it matters They call it “Mental Simulation Collapse”: models pretend to reason visually but route almost all computation through text. Their fix, AnchorSim, uses contrastive distillation to force visual latents to actually influence predictions — and boosts benchmark scores. ⚡ Our take Selling visual reasoning while secretly doing text math is expensive theater. AnchorSim is the first honest fix — make it the default.