FaithEyes-8B-RL-GGUF

FaithEyes-8B-RL is the final reinforcement-learning checkpoint of FaithEyes on Qwen3-VL-8B-Instruct, obtained by running GRPO on top of the FaithEyes-8B-SFT cold-start model, as introduced in the paper "FaithEyes: Towards Faithful Tool Use via Multi-Agent Process-Image Self-Verification." FaithEyes is a multi-agent self-judging framework in which a single VLM plays two roles: a main agent that solves visual questions by interleaving reasoning with executable code-based tool calls (like image crops), and a subagent — instantiated by the same model under a separate prompt, requiring no external model dependency — that judges whether each process image the main agent produces is actually helpful, with that verdict both injected into the tool observation to steer subsequent reasoning and used to scale the tool reward. The RL stage combines four reward terms (accuracy, format, consistency, and an accuracy-independent tool-faithfulness reward that credits helpful, executable calls and penalizes failed ones) to push the model from merely imitating demonstrations toward autonomously producing genuinely faithful rather than decorative tool calls, while keeping code-failure ratio and tool usage stable throughout training (an accuracy-gated variant was found to degenerate into tool avoidance). For this 8B model, the RL stage largely repairs the MathVision regression introduced during SFT and pushes MathVista performance beyond both the SFT checkpoint and the base model. Training built on the verl framework alongside DeepEyes, Thyme, ms_swift, and VLMEvalKit, and the model is released under the Apache 2.0 license. FaithEyes-8B-RL on Hugging Face

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp