ToolGate: Token-Efficient Pre-Call Control for Tool-Augmented Vision-Language Agents

Explain Like I'm Normal

Researchers found that vision-language agents are terrible at knowing when they actually need to call a tool, resulting in high costs and context bloat. ToolGate acts as a 'pre-call' filter that decides if a tool execution is necessary before the agent wastes tokens. It significantly reduces hallucination and redundant processing by pruning unnecessary perceptual requests.