Soothsayer recently conducted a 2 nd round of independent validation of the capabilities of the AXIOM model released by VERSES AI. The validation team consisted of Akshay Deshpande (Project Research Lead, Soothsayer), Haritima Chauhan (Project Research CoLead, Soothsayer), and Varun Agrawal (Research Assistant, Verses). AXIOM is designed to learn in ten diverse 2D arcade-style environments from the Gameworld-10k challenge. In these settings, agents learn to control characters or objects across a range of tasks. These environments are specifically designed to evaluate an agent’s general control capabilities, testing its ability to learn efficiently from limited data, sustain reasoning and planning over time, and handle delayed or sparse rewards
AXIOM (Active eXpanding Inference with Object-centric Models) achieves significant gains over deep reinforcement learning (DRL) models, even those employing model-based methodologies such as DreamerV3. In the "Gameworld-10K" benchmark, AXIOM demonstrated higher scores, better sample efficiency, and faster processing using fewer parameters.
AXIOM's efficiencies stem from its use of Variational Inference, which approximates Bayesian posteriors by leveraging priors akin to human learning, and Active Inference, where agents act to minimize free energy or surprise, thereby reducing uncertainty in decision-making. Moreover, utilizing core priors within a fast structure learning mixturemodel framework with Bayesian reduction to simplify complex models further supports efficient learning.
“The AXIOM architecture is a powerful demonstration of biologically inspired AI. Its ability
to adapt and generalize with fewer parameters and faster learning cycles makes it a
compelling candidate for real-world decision support systems.”
-Akshay Deshpande: Sr Director of AI and Research Lead for the project.
Our validation focused on evaluating the theoretical model and the performance of the proposed approach. This involved reviewing various mixture models and the Active Inference framework within the target environments, as well as examining hyperparameters and other key design components. From a performance standpoint, we evaluated three key aspects: reward structure (how the model performed in terms of step scores and cumulative rewards), sample efficiency (convergence based on the number of steps needed to reach benchmark performance), and computational efficiency (quantified by per-step total training time). All experiments were conducted using a single NVIDIA A100 GPU per environment, with results aggregated over 10 random seed trials.
Our analysis focuses on individual environments given the unique learning in each setting. We report similar average payoffs, all within one standard deviation, and the differences in cumulative reward are statistically indistinguishable at the 1% level. Additionally, with the exception of an outlier report in one environment, the maximum gameplay scores remained within the 99% margin of error. In our runs, AXIOM converged to its final performance within 2,950 steps in seven environments and 5,800 steps in the remaining three, with the fastest convergence observed at just 947 steps. Furthermore, total training time across environments ranged from 304 ms per step to 592 ms per step, with the exception of one. Overall, this is a very promising approach. We look forward to exploring its mainstream applications and potential model advancements.
“It was both exciting and intellectually rigorous to examine AXIOM’s architecture. AXIOM’s
performance held up well across all ten games, with measurable gains in efficiency and
performance. It was valuable to see the model’s core principles reflected in the outcomes,
not just the theory.”
- Haritima Chauhan, PhD: Associate Professor, Economics and Decision Science and Research Co-Lead for the project.
While AXIOM was tested on arcade-style game simulations, the same variational and active-inference principles it employs are also relevant in real-world decision-making contexts. For instance, its approach to managing uncertainty and adapting to new information in a human-like learning manner lends itself to applications such as procurement and supplier management in manufacturing sectors, or risk modeling in business information and financial services. Similarly, in applications such as predictive maintenance and dynamic inventory planning within industrial sectors, its use of core priors and mixture models can support domain-specific and micro-level learning, facilitating more responsive decision-making.
“We see our collaboration with VERSES not just as technical validation, but as part of a broader partnership to bring biologically inspired AI closer to real-world applications. -Gaurav Agrawal: CEO, Soothsayer Analytics
We extend our gratitude to Gabriel René, James Hendrickson, Steven Swanson and Hari Thiruvengada, and the entire VERSES management team, whose vision and commitment continue to push the boundaries of intelligent systems. A special thanks as well to the VERSES technical team for their responsiveness and cooperation during the validation— especially in supporting our lead researcher, Haritima Chauhan, in the evaluation process.
About Soothsayer Analytics
Soothsayer Analytics is a global leader in AI strategy, solutions, and training—helping forward-thinking organizations decode complexity, anticipate the future, and build lasting digital capabilities. With over a decade of experience across the U.S., Middle East, Europe, and India, we deliver bespoke AI solutions, scalable ML pipelines, and AI Centers of Excellence. Our interdisciplinary team of data scientists, engineers, and domain experts brings deep expertise in generative AI, computer vision, statistical modeling, and real-time forecasting. We have delivered 100+ AI-powered solutions across industries such as manufacturing, energy, automotive, healthcare, insurance, and retail—driving measurable ROI and efficiency for Fortune 1000 and Global 2000 clients
About VERSES
VERSES® is a cognitive computing company building next-generation agentic software systems modeled after the wisdom and genius of Nature. Designed around first principles found in science, physics and biology, our flagship product, Genius, is an agentic enterprise intelligence platform designed to generate reliable domain-specific predictions and decisions under uncertainty. Imagine a Smarter World that elevates human potential through technology inspired by Nature. Learn more at verses.ai, LinkedIn and X.
To learn more, please visit www.SoothsayerAnalytics.com & www.Verses.ai
Media Contact
Christopher Dole, Soothsayer Analytics
Chris@SoothsayerAnalytics.com
+1 (614) 902-0294
