What they found
The authors tested several multimodal SI models that mix text with hidden visual "latents" (internal picture-like states that are never drawn as images). They changed these latents by dropping, replacing or shuffling them. The model's final-answer probability changed only a little. The same changes to the written reasoning hurt much more. The authors call this Mental Simulation Collapse: the model looks like it reasons visually, but most of the answer work runs through text. They argue current training goals allow this, because they make latents look like helper-image embeddings without forcing the model to rely on them. They then propose AnchorSim, a contrastive distillation method (training by contrasting examples) that makes the latents consequential. The abstract says AnchorSim improves results on multiple visual reasoning benchmarks, and that visual changes now affect predictions. It gives no numbers.
Why it matters
A model can look like it is thinking in pictures while doing the work in words. This paper gives a direct test for that gap. It also shows that gains from a new reasoning format can be hollow unless the model actually uses the new part. Readers should check whether the parts of a model that look important truly drive its answers.
How it changes the race
If the result holds, builders of multimodal SI get a simple diagnostic and a training fix for latent visual reasoning. It is a targeted improvement to one family of methods, not a change in overall capability.
Safety and control
The perturbation test shows that a model's visible reasoning steps may not reflect what drives its answer. That is a useful reminder for anyone who relies on reasoning traces to understand or oversee SI systems. The abstract does not discuss safety directly.
The abstract, by the authors
Mental simulation, the ability to reason by constructing and updating internal visual states, is central to flexible human cognition, but remains elusive in current multimodal large language models. Latent multimodal reasoning offers a promising route toward this capability by interleaving text with continuous visual latent blocks, allowing models to maintain internal visual workspaces without decoding them into pixels. In this paper, we ask whether these latent visual states actually matter. Through a direct perturbation analysis across representative latent-reasoning MLLMs, we find that they often do not: dropping, replacing, or shuffling generated visual latents has only minor effects on the model's own final-answer probability, while equivalent perturbations to the linguistic reasoning trace are substantially more damaging. We call this failure mode Mental Simulation Collapse (MSC): the model appears to perform interleaved visual reasoning, but routes most answer-relevant computation through text. We further diagnose that MSC follows naturally from existing training objectives: visual alignment can make generated latents resemble helper-image embeddings, but it does not require the decoder to rely on them; subsequent text-only supervision further leaves the semantic role of the latent pathway unidentifiable. To address this, we introduce AnchorSim (Anchored Mental Simulation), a contrastive distillation framework that trains models to make their generated visual latents consequential. Across multiple visual reasoning benchmarks, AnchorSim improves performance, while diagnostic re-evaluation shows that visual perturbations now meaningfully affect predictions. Our results expose a fundamental failure mode of latent multimodal reasoning and demonstrate that mental simulation must be anchored to model behavior in order to matter.
Quoted from arXiv:2611.10423. The research and its words belong to its authors.