FaithEyes-8B-RL
This is the final reinforcement-learning (RL) checkpoint of FaithEyes on the Qwen3-VL-8B initialization, obtained by running GRPO on top of the FaithEyes-8B-SFT cold-start model, with the base model Qwen3-VL-8B-Instruct.
FaithEyes: Towards Faithful Tool Use via Multi-Agent Process-Image Self-Verification
Haoqing Wang, Xingrun Xing, Ziheng Li, Jianyuan Guo, Yehui Tang
📖 Model Description
FaithEyes is a multi-agent self-judging framework for agentic vision-language models (VLMs). A single VLM plays two roles: a main agent that solves the visual question by interleaving reasoning with executable code-based tool calls, and a subagent that judges whether each process image (e.g. a crop) produced by the main agent is helpful for answering the question. The verdict {"is_helpful": true/false, "reasons": ...} is both injected into the tool observation to steer subsequent reasoning and used to scale the tool reward by the helpful-tool ratio, suppressing reward hacking.
This is the final RL checkpoint on the stronger Qwen3-VL-8B initialization. The RL stage (GRPO) is what drives the model from merely imitating demonstrations to autonomously producing faithful tool calls — i.e. calls that genuinely capture the queried target — rather than decorative ones. It combines four reward terms: accuracy, format, consistency, and a faithfulness-targeting tool reward $r_{\mathrm{tool}} = 1 - (n_{\mathrm{fail}} + n_{\mathrm{unhelpful}}) / n_{\mathrm{tool}}$ that scales the bonus by the helpful-tool ratio. Crucially, the tool reward is accuracy-independent: it always credits an executable and helpful call and always penalizes a failed one, which keeps both the code-failure ratio and tool usage stable throughout training (an accuracy-gated variant degenerates into tool avoidance).
For the 8B model, the RL stage also largely repairs the MathVision drop introduced during SFT and pushes MathVista beyond both the SFT model and the base model. The self-judging subagent is instantiated by the model itself (under a separate prompt), so the judgement is available at inference with no external model dependency, keeping train–test consistency.
📚 Citation
If you find this work useful, please consider citing:
@article{wang2026faitheyes,
title={FaithEyes: Towards Faithful Tool Use via Multi-Agent Process-Image Verification},
author={Wang, Haoqing and Xing, Xingrun and Xia, Wei and Li, Ziheng and Tang, Yehui},
journal={arXiv preprint arXiv:2607.28225},
year={2026}
}
🙏 Acknowledgements
This model is trained on top of Qwen3-VL-8B-Instruct. The training framework builds on verl, DeepEyes, Thyme, ms_swift, and VLMEvalKit.