Vanguard: Knowing When to Trust the Draft in Speculative Vision-Language Agents
Summary
- Researchers introduce Vanguard, a training-free judge designed for step-level speculative vision-language agents.
- Vanguard utilizes screenshots alongside candidate steps to evaluate small model drafts in a single forward pass, calling larger target models only when drafts are rejected.
- The training-free decision model improves accuracy and efficiency across web, mobile, and embodied agent tasks compared to standalone target models.
Why it matters
Vanguard reduces the latency and cost of deploying vision-language agents by enabling smaller draft models to perform accurately, improving both speed and task completion rates.