Abstract
Egocentric video reasoning focuses on the unseen, egocentric agent who shapes the scene, demanding inference of hidden intentions and fine-grained interactions—areas where current MLLMs struggle. We present EgoThinker, a framework that equips MLLMs with strong egocentric reasoning via spatio-temporal chain-of-thought supervision and a two-stage curriculum. We build EgoRe-5M, a large-scale QA dataset derived from 13M egocentric clips, featuring multi-minute segments with detailed rationales and dense hand–object grounding. Trained with SFT on EgoRe-5M and refined with RFT for better spatio-temporal localization, EgoThinker outperforms prior methods on multiple egocentric benchmarks and yields substantial gains in fine-grained localization tasks.