Megadose The AI curve, compressed.

Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence

· HF Daily Papers ·
ReaLVR adds direct visual-evidence supervision to latent visual reasoning, and the paper reports gains up to 235B-scale models.

The authors identify a “latent evidence-credit gap”: latent tokens change only weakly when image perturbations change the right answer. ReaLVR trains on the model’s own latent trajectories, contrasting correct and wrong answers and matching them against relevant or mismatched visual evidence. Across three model families, it beats the evaluated latent visual reasoning baselines, with a 63.7% five-task average on Qwen2.5-VL-7B. HF Daily Papers' note

score 4

Categories: Research