Qwen3-VL-4B-GRPO

Qwen3-VL-4B-GRPO: GRPO baseline

This repository contains the Qwen3-VL-4B weights fine-tuned with GRPO, released as a baseline for the paper Learning Spatiotemporal Sensitivity in Video LLMs via Counterfactual Reinforcement Learning.

Overview

This checkpoint is a Group Relative Policy Optimization (GRPO) baseline fine-tuned on top of Qwen3-VL for video question answering, released as a reference RL baseline for the CRPO paper.

Video large language models (Video LLMs) often answer video questions through shortcuts such as single-frame cues and language priors rather than by tracking spatiotemporal dynamics. The CRPO paper studies this problem and compares against multiple RL baselines, of which this checkpoint is one.

Resources

• Paper: Learning Spatiotemporal Sensitivity in Video LLMs via Counterfactual Reinforcement Learning
• Project Page: https://ddz16.github.io/crpo.github.io/
• GitHub Repository: https://github.com/ddz16/CRPO
• CRPO model (same size): ddz16/Qwen3-VL-4B-CRPO

Evaluation

The model is evaluated on DyBench, a paired counterfactual video benchmark with 3,014 videos covering reversible dynamics, moving direction, and event sequence, together with standard video-QA benchmarks (Video-MME, TempCompass, MVBench, TimeBlind). See the paper for full numbers and comparison with CRPO.