OmniOPSD: Rationale-Privileged On-Policy Self-Distillation for Affective Computing

Researchers have proposed new frameworks to address limitations in reinforcement learning for multimodal large language models, achieving state-of-the-art performance in human-centered scenarios.

Two new frameworks, OmniOPSD and d-OPSD, have been introduced to tackle the challenges of reinforcement learning in multimodal large language models (MLLMs). OmniOPSD uses frontier-generated rationales as teacher-side privileged evidence, rather than student imitation targets, allowing the student to learn on its own trajectory distribution without directly imitating frontier-model completions^[1]. d-OPSD, on the other hand, is tailored for diffusion large language models (dLLMs) and reframes self-teacher construction using self-generated answers as suffix conditioning, shifting supervision from token-level to step-level^[2]. According to the authors, OmniOPSD achieved an average score of $84.19 on MER-UniBench^[1]. d-OPSD consistently outperformed RLVR and SFT baselines with superior sample efficiency, requiring only around 10% of the optimization steps by RLVR^[2]. The proposals were submitted to arXiv on 14 Jun 2026 and 16 Jun 2026, respectively^[1][2].