CoLVR: Enhancing Exploratory Latent Visual Reasoning via Contrastive Optimization

Name: CoLVR: Enhancing Exploratory Latent Visual Reasoning via Contrastive Optimization
Uploaded: 2026-10-20T18:43:14.250657+00:00
Duration: 618 s
Description: CoLVR introduces a latent contrastive training framework to improve how Large Language Models explore their hidden states during visual reasoning. It uses contrastive supervision and trajectory-level rewards instead of hard alignment, resulting in more diverse, flexible, and robust latent representations that significantly boost performance on visual tasks.

CoLVR introduces a latent contrastive training framework to improve how Large Language Models explore their hidden states during visual reasoning. It uses contrastive supervision and trajectory-level rewards instead of hard alignment, resulting in more diverse, flexible, and robust latent representations that significantly boost performance on visual tasks.

Open the full episode