Apple proposes TGPO to improve temporal reasoning in egocentric video models

Apple’s Machine Learning Research team proposes Temporal Global Policy Optimization (TGPO), a reinforcement learning with verifiable rewards (RLVR) algorithm designed to explicitly incentivize temporal awareness in multimodal large language models (MLLMs) for egocentric video understanding.

The authors argue that current MLLMs often lack temporal awareness because training objectives fail to reward temporal reasoning, leading models to rely on frame-level spatial shortcuts instead. TGPO addresses this by contrasting model outputs generated from temporally ordered versus temporally shuffled video frames, producing calibrated, globally normalized reward signals that favor temporally coherent reasoning.

TGPO is integrated with GRPO and GSPO to support cold-start RL training and to suppress spatial shortcut behaviors learned by existing MLLMs. The authors report that experiments across five egocentric video benchmarks show TGPO consistently improves temporal grounding and causal coherence, outperforming prior RL-based video reasoning approaches.

The paper positions TGPO as a simple and scalable pathway toward temporally robust MLLMs for egocentric video understanding, highlighting its potential to improve real-world applications where event ordering and causal relationships are essential.