Jackwang111/FaithEyes-8B-SFT
FaithEyes-8B-SFT is an 8 billion parameter supervised-fine-tuned (SFT) checkpoint of the FaithEyes framework, built upon Qwen3-VL-8B-Instruct. This vision-language model (VLM) is designed for multi-agent self-judging, where a single VLM acts as both a main agent for visual question-answering with code-based tool use and a subagent for judging the helpfulness of process images. It specializes in integrating feedback-driven reasoning and faithfulness judging into tool-use observations, representing the initial SFT stage before reinforcement learning for autonomous discrimination.
Loading preview...
FaithEyes-8B-SFT: Multi-Agent Self-Judging VLM
FaithEyes-8B-SFT is the supervised-fine-tuning (SFT) checkpoint of the FaithEyes framework, developed by Haoqing Wang et al. It is based on the Qwen3-VL-8B-Instruct model and serves as the first stage in a two-stage SFT + RL pipeline. This model introduces a novel multi-agent self-judging mechanism for agentic vision-language models (VLMs).
Key Capabilities
- Code-based Tool Use: Enables the VLM to generate and execute Python code for image manipulation tasks such as cropping, zooming, rotation, contrast adjustment, and arithmetic operations.
- Faithfulness Judging: Allows the VLM to act as a subagent, evaluating whether a generated process image is helpful for answering a given question and outputting a JSON verdict (
{"is_helpful": true/false, "reasons": ...}). - Feedback-Driven Reasoning: Integrates the subagent's judgment directly into the main agent's reasoning process, allowing subsequent steps to be conditioned on this feedback rather than ignoring it.
- Joint Training: The main agent (solving visual questions) and subagent (judging image helpfulness) share the same VLM and are trained jointly in this SFT stage.
Good For
- Developing agentic VLMs that can self-verify their intermediate steps.
- Research into faithful tool use and multi-agent VLM architectures.
- Applications requiring visual question answering with integrated image manipulation and self-correction capabilities.
This SFT model primarily imitates demonstrations; for autonomous discrimination of helpful tool calls, the subsequent RL checkpoint is required.