可验证奖励教会视频语言模型时间吗？多模型对照研究
原标题：What Do Verifiable Rewards Teach Video-Language Models About Time? A Controlled Multi-Model Study
AI 摘要
该研究对四个开源视频语言模型（Qwen3-VL-8B/4B、Qwen2.5-VL-7B、Gemma-3-12B）使用GRPO进行可验证奖励强化学习（RLVR），对比了验证合成数据、未验证真实视频及混合三种数据配方。结果显示，验证奖励训练能大幅提升域内准确率，但其中相当一部分增益并非来自视觉感知，而是模型学到了答案先验和格式；仅用验证数据训练会严重损害域外真实视频性能（Qwen3-VL-8B下降26.7和25.2分），而混合真实视频数据可避免退化。研究还发现，没有任何配方能让模型真正学会利用帧序，即时间顺序接地能力未得到提升。
正文节选
What Do Verifiable Rewards Teach Video–Language Models About Time? A Controlled Multi-Model Study Abstract Reinforcement learning from verifiable rewards (RLVR) has produced large reasoning gains in language models, and verifiable video benchmarks make the recipe directly applicable to causal–temporal video question answering. We present a controlled study of what RLVR actually teaches video–language models about temporal structure. We fine-tune four open models (Qwen3-VL-8B/4B, Qwen2.5-VL-7B, G