Making Every Tool Call Count: Necessary Tool-Evidence Path Rewards for Agentic Vision-Language Models

Explain Like I'm Normal

Researchers developed a new training method for vision-language models that rewards them for finding 'necessary evidence' rather than just guessing the right final answer. By penalizing redundant or off-target tool calls, the model learns to strategically use image cropping and web search to solve complex visual tasks. This moves agents away from 'lucky guesses' and toward actual reasoning-based information gathering.