Gravity & Flight · 4B

Evidence-grounded reasoning for fine-grained remote sensing.

Project website · GitHub · Dataset preview

Gravity and Flight (G&F) develops Qwen3.5-4B through Evidence-of-Thought supervised fine-tuning and Relational Transfer Policy Optimization. Gravity anchors visual claims to the image; Flight checks how reasoning uses those observations and adapts to related contexts. Inference uses one image and one question.

Release status

This is the official model repository reserved for the forthcoming release. Model weights are not uploaded yet. We plan to release the model and full dataset in stages after paper acceptance. There is no model download or inference package in this preview.

Three Stage-1 demonstrations and three Stage-2 image-pair examples are provided in the dataset preview. Reported experiments and updates are documented on the project website and in the research notes.

本页为 G&F 官方模型发布入口,目前仅公开项目介绍,尚未上传模型权重。模型与完整数据集计划在论文中稿后逐步开放;数据集页面现提供少量 Stage 1、Stage 2 样例预览。

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for zongling/Gravity_and_Flight-4B

Finetuned
Qwen/Qwen3.5-4B
Finetuned
(1,264)
this model