Researchers have released OmniGAIA, a benchmark designed to stress-test AI agents across video, audio, and image simultaneously — not just in pairs.
Most multi-modal models today handle two inputs at once, typically vision and language. OmniGAIA is built to probe what happens when an agent must reason across all three perceptual channels at once, through multi-turn tool calls and multi-hop queries pulled from real-world data. The benchmark is constructed using an "omni-modal event graph" method that chains cross-modal dependencies together, making it harder for a model to get the right answer by leaning on a single strong modality. Alongside the benchmark, the team released OmniAtlas, a foundation agent trained on synthesized reasoning trajectories and fine-tuned with a process the authors call OmniDPO for targeted error correction.
The gap this research targets is real: most deployed AI assistants are still glorified vision-language pipelines bolted to a text interface. If agents are ever going to handle the messy, multi-sensory nature of real-world tasks — think a video call summary that also flags an alarm in the background audio — they need to reason across modalities rather than process them separately. OmniGAIA gives the field a shared measuring stick for that capability.
Whether OmniAtlas actually closes that gap at scale is the open question; benchmarks built by the same team that trains the model being evaluated tend to flatter the home side.