CNBeta — 2026-05-01#
Top Story#
According to a cnbeta report, DeepSeek briefly published and then deleted a highly revealing technical paper detailing its new visual reasoning model. The model bypasses the “Reference Gap” bottleneck found in top-tier Western models by using point and bounding box coordinates as cognitive anchors during its chain-of-thought process, rather than relying solely on linguistic descriptions. This breakthrough allows the AI to simulate human “point-to-reason” synergy, significantly outperforming competitors like GPT-5.4 and Claude 4.6 in complex spatial tasks such as maze navigation, all while utilizing a mere fraction of the computational tokens required by other multimodal architectures.