Dual Exam: GPT 6 Astra Scores 31.3 Among Multimodal Agents

If an agent learns in a simulation built by a world model, which of the two is to blame when it fails? A group of more than 80 authors proposes giving them separate exams. The Dual Exam benchmark tests multimodal agents and world models independently, then looks at what happens when they are connected. The top agent, GPT 6 Astra, scored 31.3 on the weighted Agent Index. The best fully tested world model, GPT-Image-2, got 21.7 on WM Overall. A pilot pairing of the two, run on 150 cases, did not solve a single task.

Two exams instead of one

The idea is simple. A world model can feed an agent simulated experience, and agent trajectories can in turn help fine-tune the world model. If you only score the final outcome of the whole pairing, you cannot tell where the error came from: the choice of action or the prediction of the next state. So the authors build two independent tracks.

Agents are tested with the environment's response mechanisms held fixed. World models are tested with fixed interventions, meaning actions set in advance. The benchmark covers four domains: physical, digital, social and scientific. This way a result cannot be blamed on the environment or the actions differing from run to run.

What was measured

The agent track compares 10 configurations on 1,088 tasks, grouped into 15 report categories. The world model track includes 400 cases across eight scenarios. 15 configurations have full coverage, and one more completed only six scenarios.

Key numbers:

• GPT 6 Astra: 31.3 on the case-weighted Agent Index, first place among agents;
• GPT-Image-2: 21.7 on WM Overall, first place among world models with full coverage;
• agent track: 10 configurations, 1,088 tasks, 15 categories;
• world model track: 400 cases, 8 scenarios, 15 configurations with full coverage and one covering six scenarios;
• pairing pilot: 150 cases, 0 successful tasks.

The authors themselves warn that 31.3 and 21.7 cannot be compared. The tracks have different tasks and different metrics, so the numbers do not sit on one scale. They show the leaders within their own categories, not that agents are stronger than world models.

Zero successes for the pairing

The most striking line in the paper concerns the pilot. The researchers took one "agent plus world model" pair and ran it on 150 cases. Not one task finished successfully. From this the authors conclude that component-level diagnostics are needed: the pairing's overall score says nothing about what broke.

Across the two tracks, the picture is as follows. Performance is uneven across tasks: a configuration that is strong in one domain falls short in another. World models often fail to preserve state that matters for the task. If an object or parameter in the simulation "drifts" after a step, the agent learns from distorted experience.

The paper does not promise, however, that a high Dual Exam score will predict the success of a pairing or stable mutual learning between an agent and a world model. The authors state outright that the value of these scores for such a prediction has yet to be proven.

How to use the results

For teams building agents that learn in simulation or in an "imagined" environment, the takeaway is practical. Do not trust a single aggregate metric. Test the agent and the world model separately, and first check whether the world model preserves the state the task depends on. If the pairing does not work, start there rather than with fine-tuning the agent.

For those choosing models, the numbers are useful as a reference point within a track. GPT 6 Astra sets the bar for agents, and GPT-Image-2 for world models. But 31.3 and 21.7 remain low, and the industry is far from a reliable pairing of the two components. The paper appeared on Hugging Face Daily Papers on 22 October 2026 under arXiv number 2610.31066 and collected 8 votes.