Incentivizing Temporal-Awareness in Egocentric Video Understanding Models

2026-03-28

AI Summary

Existing multimodal large language models often lack temporal awareness in egocentric video understanding, tending to rely on frame-level spatial shortcuts while neglecting the sequential evolution of events. This work proposes Temporal Global Policy Optimization (TGPO), a novel algorithm that, for the first time, integrates reinforcement learning with a verifiable reward mechanism. By comparing model outputs on temporally ordered versus shuffled video frames, TGPO constructs a globally normalized calibration reward signal that explicitly encourages temporally coherent reasoning. The method supports cold-start training and effectively suppresses reliance on spatial shortcuts. Evaluated across five egocentric video benchmarks, TGPO significantly improves temporal localization accuracy and causal coherence, outperforming current reinforcement learning–based approaches for video reasoning.

Abstract

Multimodal large language models (MLLMs) have recently shown strong performance in visual understanding, yet they often lack temporal awareness, particularly in egocentric settings where reasoning depends on the correct ordering and evolution of events. This deficiency stems in part from training objectives that fail to explicitly reward temporal reasoning and instead rely on frame-level spatial shortcuts. To address this limitation, we propose Temporal Global Policy Optimization (TGPO), a reinforcement learning with verifiable rewards (RLVR) algorithm designed to incentivize temporal awareness in MLLMs. TGPO contrasts model outputs generated from temporally ordered versus shuffled video frames to derive calibrated, globally normalized reward signals that explicitly favor temporally coherent reasoning. Integrated with GRPO and GSPO, TGPO supports cold-start RL training and effectively suppresses spatial shortcut behaviors learned by existing MLLMs. Experiments across five egocentric video benchmarks demonstrate that TGPO consistently improves temporal grounding and causal coherence, outperforming prior RL-based video reasoning approaches. Our results suggest that TGPO offers a simple and scalable pathway toward temporally robust MLLMs for egocentric video understanding.

Problem

• temporal awareness
• egocentric video understanding
• multimodal large language models
• temporal reasoning
• spatial shortcuts

Innovation

• Temporal Global Policy Optimization
• Reinforcement Learning with Verifiable Rewards
• Temporal Awareness
• Egocentric Video Understanding
• Multimodal Large Language Models