Abstract
Egocentric social-affective understanding is central to embodied affective intelligence, yet remains challenging for multimodal large language models because correct interpretations often depend on subtle egocentric evidence and fine-grained distinctions among plausible candidates. We propose EgoAffect-R1, a process-aware and elimination-guided reinforcement learning framework for egocentric embodied emotional reasoning. We construct EgoAffect-CoT-18k for cold-start CoT supervision and EgoAffect-RL-5k for verifiable reinforcement learning. Our EgoAffect reward scores format, question type, emotional tone, and elimination-based candidate discrimination, penalizing trajectories that discard the correct semantic option. To reduce option-order shortcuts, we introduce Option-Perturbed DAPO, which trains on reordered candidates and aligns rewards with their semantic identities. Experiments on EgoLink show that EgoAffect-R1 achieves 92.75% accuracy on the EgoLink validation split and 92.46% accuracy in the EgoLink 2026 Challenge. These results show that structured cold-start data, elimination-guided rewards, and option-perturbed RL improve egocentric affective reasoning. Code and data are available at https://github.com/ZhengQinLai/EgoAffect-R1.
Funding Information
  • National Key R&D Program of China (2025YFC3811300)
  • National Natural Science Foundation of China (62376070)
  • National Natural Science Foundation of China (62076195)