Multimodal large language models (MLLMs) increasingly rely on latent visual reasoning (LVR), a paradigm in which intermediate computation happens in continuous latent tokens instead of being spelled out in words. But unlike textual chain-of-thought, these latent tokens are not directly observable, which makes it hard to supervise what they actually learn.
A new paper posted on Hugging Face analyzes latent-token behavior and reports a latent evidence-credit gap: latent tokens respond only weakly to image perturbations that change the correct answer. The authors hypothesize this stems from the absence of explicit supervision during GRPO training, arguing that a final-answer reward offers too little guidance on which visual evidence to preserve and how credit should be distributed across latent tokens.
To address this, the authors propose ReaLVR, which brings visual-evidence supervision into the model's own free-running latent trajectories. The method contrasts correct answers with model-generated wrong answers to decide where stronger supervision is needed, and contrasts relevant versus mismatched visual evidence to specify what should be preserved.
Across three model families, ReaLVR consistently outperforms the evaluated LVR baselines, reaching the highest five-task average of 63.7% on Qwen2.5-VL-7B. The authors state they are the first to scale visual reasoning in latent space, reporting robust improvements at frontier model scales up to 235B parameters.
Further analyses indicate more question-sensitive latent-token positions, stronger alignment with relevant visual regions, and greater fixed-context dependence on the most attended latent tokens.