VisualThink-VLA: Visual Intermediate Reasoning for Effective and Low-Latency Vision-Language-Action Policies

VisualThink-VLA introduces compact visual intermediate reasoning for vision-language-action policies, avoiding text chain-of-thought overhead while preserving spatial detail.

Selective routing and visual evidence checks improve both faithfulness and efficiency, supported by the new VisualEvidence-Kit dataset.

The system achieved top success rates on most benchmarks and greatly reduced inference latency for robot control tasks.

On BridgeData V2, latency dropped from 8.377 seconds to 0.367 seconds, making reasoning-enabled VLA policies far more usable in real time.