# Latent Visual Reasoning

Interleaved multimodal reasoning inserts continuous visual latent blocks between text spans, but direct perturbation shows that dropping, replacing or shuffling these blocks barely changes the model's own answer probability, while the same perturbations to the textual reasoning trace hurt far more. The authors call this mental simulation collapse and introduce AnchorSim, which during training contrasts predictions with the visual pathway against predictions with it masked, forcing the latents to carry answer-relevant information. Built on Qwen2.5-VL-7B, AnchorSim-7B averages 62.4 across nine visual reasoning benchmarks, 13.5 points above the base model, and visual perturbations become consequential again.

Interleaved multimodal reasoning inserts continuous visual latent blocks between text spans, but direct perturbation shows that dropping, replacing or shuffling these blocks barely changes the model's own answer probability, while the same perturbations to the textual reasoning trace hurt far more. The authors call this mental simulation collapse and introduce AnchorSim, which during training contrasts predictions with the visual pathway against predictions with it masked, forcing the latents to carry answer-relevant information. Built on Qwen2.5-VL-7B, AnchorSim-7B averages 62.4 across nine visual reasoning benchmarks, 13.5 points above the base model, and visual perturbations become consequential again.

The work first analyzes latent-token behavior in latent visual reasoning (LVR) and identifies a latent evidence-credit gap: latent tokens respond only weakly to image perturbations that alter the correct answer, which the authors attribute to the lack of explicit supervision during GRPO training; it then proposes ReaLVR, which brings visual-evidence supervision to the model's own free-running latent trajectories, using correct-versus-wrong answer contrast to decide where stronger supervision is needed and relevant-versus-mismatched visual evidence to specify what to preserve, consistently outperforming evaluated LVR baselines across three model families, reaching a 63.7% five-task average on Qwen2.5-VL-7B, and being the first to scale visual reasoning in latent space up to 235B parameters.

The work first analyzes latent-token behavior in latent visual reasoning (LVR) and identifies a latent evidence-credit gap: latent tokens respond only weakly to image perturbations that alter the correct answer, which the authors attribute to the lack of explicit supervision during GRPO training; it then proposes ReaLVR, which brings visual-evidence supervision to the model's own free-running latent trajectories, using correct-versus-wrong answer contrast to decide where stronger supervision is needed and relevant-versus-mismatched visual evidence to specify what to preserve, consistently outperforming evaluated LVR baselines across three model families, reaching a 63.7% five-task average on Qwen2.5-VL-7B, and being the first to scale visual reasoning in latent space up to 235B parameters.

The work proposes Scaffolding Minds, a two-stage framework in which Stage 1 replaces the frozen off-the-shelf vision encoder with a scaffolding encoder trained end-to-end through the downstream task loss to produce the latent target, and Stage 2 replaces deterministic latent regularization with a Gaussian sampler whose mean and variance are learned so that residual latent actions can be sampled for RL exploration, improving over the strongest latent-reasoning baseline by 9.5 points on average on FrozenLake (19 points at Level 32) and by 5.6 points on average across nine visual-centric reasoning benchmarks.

The work proposes Scaffolding Minds, a two-stage framework in which Stage 1 replaces the frozen off-the-shelf vision encoder with a scaffolding encoder trained end-to-end through the downstream task loss to produce the latent target, and Stage 2 replaces deterministic latent regularization with a Gaussian sampler whose mean and variance are learned so that residual latent actions can be sampled for RL exploration, improving over the strongest latent-reasoning baseline by 9.5 points on average on FrozenLake (19 points at Level 32) and by 5.6 points on average across nine visual-centric reasoning benchmarks.