Decompose, Look, and Reason: Reinforced Latent Reasoning for VLMs

Build a lightweight plug-in for popular VLMs that performs recursive latent visual decomposition before answering queries. This would drastically improve accuracy on complex spatial and multi-step visual reasoning tasks.

Suggested repo: latent-look

"Stop guessing visual context—decompose, look, and reason."

Estimated effort: 40h