[ArXiv 2026] OmniGAIA & OmniAtlas: Breaking the Bi-Modal Ceiling for Native AI Agents

OmniGAIA: Towards Native Omni-Modal AI Agents

Summary
Problem
Method
Results
Takeaways
Abstract

This paper introduces OmniGAIA, a pioneering benchmark for evaluating "Omni-modal" AI agents capable of interleaved reasoning across video, audio, and images. To address the gap in agentic capabilities, the authors also propose OmniAtlas, a native foundation agent optimized for tool-integrated reasoning and active perception, setting a new SOTA for open-source models (improving Qwen3-Omni from 13.3 to 20.8 Pass@1).

In the evolution of AI, we have moved from text-only LLMs to vision-language models. However, human intelligence doesn't just "see" or "read"—it perceives the world as a continuous, omni-modal stream of audio, video, and imagery, while using tools to fill gaps in its knowledge.

A team of researchers from Renmin University of China, Xiaohongshu, and other top institutions has released OmniGAIA, a benchmark designed to test this exact frontier. Alongside it, they introduced OmniAtlas, a recipe to turn standard omni-modal models into high-functioning agents.

TL;DR

OmniGAIA is the first benchmark to demand multi-hop reasoning and multi-turn tool use across video, audio, and images. The results are a wake-up call: most current "Omni" models fail miserably, often confusing movie-inspired priors with real-world evidence. OmniAtlas, their proposed agent, provides a training framework—including OmniDPO—that significantly closes the gap between open-source models and proprietary giants like Gemini-3-Pro.


The Problem: Why "Omni" Doesn't Mean "Agent"

Most current multi-modal models are tested on perception: "What color is the car?" or "What is the speaker saying?". In reality, an AI assistant needs to:

  1. Ground Evidence: Connect a specific bridge visible in a video to a historical site mentioned in the audio.
  2. Execute Tools: Use web search to find that bridge's construction date.
  3. Reason & Compute: Subtract the construction date from a filming date to answer a complex query.

Current benchmarks (like OmniBench or WorldSense) are too simple, focusing mostly on short clips and multiple-choice questions. They don't capture the logic collapse that happens when a model makes a perception error that cascades into a reasoning failure.


Methodology: Building the Event Graph

To create hard but solvable tasks, the authors developed an Omni-modal Event Graph pipeline. Instead of just asking questions about a video, they:

  • Extract time-aware signals (ASR, OCR, object detection).
  • Build a graph of entities and cross-modal relations.
  • Expand the graph using AI agents that browse the web to find "next-hop" evidence.
  • Fuzzify nodes: Masking key entities to force the model to use the full reasoning path rather than relying on internal memory.

OmniGAIA Construction Pipeline


OmniAtlas: The TIR Paradigm

The researchers didn't just stop at a benchmark; they built OmniAtlas. Two key innovations define this model:

  1. Active Omni-Modal Perception: Instead of downsampling a 10-minute video into a few frames (and losing detail), OmniAtlas can explicitly call read_video(start, end) or read_audio(segment). It "looks" and "listens" where its reasoning suggests evidence might be.
  2. OmniDPO (Fine-Grained Correction): Traditional DPO looks at whole trajectories. OmniDPO identifies the first point of failure (was it perception? tool call? logic?) and generates a corrected prefix to teach the model exactly where it went wrong.

OmniAtlas Training Strategy


Experimental Results: A Massive Gap

The benchmark revealed a "stark proprietary-open gap."

  • Gemini-3-Pro: 62.5% Pass@1.
  • Qwen3-Omni (Baseline): 13.3% Pass@1.
  • OmniAtlas-Qwen3: 20.8% Pass@1 (A significant boost).

Key Insight: Scaling doesn't solve this. the 560B LongCat-Flash-Omni actually performed worse than the 30B Qwen3 model. This suggests that agentic policy, not parameter count, is the primary bottleneck.

Fine-Grained Error Analysis

The error analysis (above) shows that while OmniAtlas helps with tool use, perception remains the ultimate ceiling. If the model mishears a location name, no amount of web searching will find the correct answer.


Case Study: "The Blues Brothers" Trap

The paper highlights a fascinating case where the model sees a bridge and hears the speaker mention The Blues Brothers.

  • The Failure: Most models jump to a "Chicago Bridge" because that's what's in their training data (the movie prior).
  • The Success: A strong agent (Gemini-3-Pro or OmniAtlas) ignores the movie distraction, grounds the location at "Joliet Iron Works" (from the audio), identifies the local "Ruby Street Bridge," and then uses a search tool to find its 1935 construction date.

Conclusion & Future Outlook

OmniGAIA proves that the next generation of AI isn't just about having "more" modalities—it's about the interplay between them. The authors suggest that the future lies in Omni-modal Agentic RL and Embodied Agents that can use these same reasoning patterns in the physical world.

Takeaway: For researchers and developers, the focus should shift from blanket perception to "Tool-Integrated Reasoning." Native perception is the floor; effective tool use is the ceiling.

Find Similar Papers

Try Our Examples

  • Search for recent papers on multi-modal AI agents that utilize external tools or browsing capabilities for multi-step reasoning.
  • Which study first introduced the concept of 'active perception' in the context of Large Multimodal Models (LMMs), and how does the approach in OmniAtlas differ?
  • Explore research that applies omni-modal foundation models to embodied AI or physical robotics tasks requiring simultaneous audio-visual processing.
Contents
[ArXiv 2026] OmniGAIA & OmniAtlas: Breaking the Bi-Modal Ceiling for Native AI Agents
1. TL;DR
2. The Problem: Why "Omni" Doesn't Mean "Agent"
3. Methodology: Building the Event Graph
4. OmniAtlas: The TIR Paradigm
5. Experimental Results: A Massive Gap
6. Case Study: "The Blues Brothers" Trap
7. Conclusion & Future Outlook