OmniVChat defines and studies native audio-visual dialogue, where an omni model receives simultaneous audio and video (with the query embedded in them, no separate text) and responds in text. To overcome the scarcity of real recordings and the unreliability of keyword-based scoring, the authors build OmniVChat-Studio, a multi-agent engine that synthesizes single- and multi-turn dialogues, and use it to construct the OmniVChat-Bench evaluation suite plus OmniVChat-RL, a reinforcement-learning reward design targeting correctness, efficiency, and style. Training Qwen3-Omni-Instruct with OmniVChat-RL on synthesized data improves performance on both the synthetic benchmark and a human-recorded validation set, showing transfer to real dialogues.
