The authors decoded each video once at a maximum frame count, then constructed multiple aligned temporal views by uniformly subsampling that decode at different ratios.
Varying sampled video frames strengthens reinforcement learning for temporal reasoning
FrameGRPO trains video models on multiple temporally aligned views, preserving useful reward differences between rollouts without adding networks or changing the model architecture.
Chinese Tech
Jiaxuan Sun · Biao Wang · Shuai Li
Shanghai Jiao Tong University · Shanghai Innovation Institute · Tencent Hy
Research Digest··3 min read
Sun, Wang and Li address a weakness in Group Relative Policy Optimization for video: rollouts normally see identical frames, so they may all succeed or fail and provide little comparative learning signal.
§