FaithEyes: Towards Faithful Tool Use via Multi-Agent Process-Image Verification

Researchers proposed FaithEyes, a multi-agent framework that checks whether tool-generated process images are actually useful for vision-language models.

The system feeds the model’s own judgment signal back into reasoning and down-weights rewards for unhelpful tool calls.

It is trained with a two-stage pipeline of supervised fine-tuning and reinforcement learning.

The approach aims to improve tool faithfulness, reliability, and interpretability without needing an external judge at inference time.