[Exclusive] Tech Mahindra is Working on World Models, Eyes ARC-AGI Milestone This Year

Tech Mahindra is building AI world models and working on the ARC-AGI challenge as part of a broader push towards AGI, with the company targeting tangible outcomes from its research efforts later this year, AIM has learnt.

The initiative is being led by the company’s Makers Lab, which is exploring world models, next-generation vision-language models (VLMs), and AGI systems from an Indian context, and was selected under the IndiaAI Mission to build sovereign language models. 

People familiar with the matter said the research goes beyond conventional LLMs, focusing on enabling AI systems to understand, simulate, and reason about the world.

The company is also working on the ARC-AGI challenge, a benchmark widely regarded as one of the toughest tests of general intelligence in machines. The effort forms part of a larger programme to build foundational AI capabilities from the ground up rather than relying solely on existing foundation models.

The work marks Tech Mahindra’s most ambitious AI research project yet and signals an expansion of its focus from sovereign language models to more advanced reasoning systems. 

Earlier this year, the company launched an 8-billion-parameter Hindi-first education LLM under the IndiaAI Mission, becoming the only IT services firm selected by the programme to build its own foundational model.

The latest research direction is closely aligned with the vision articulated by Nikhil Malhotra, who has for several years advocated developing world models as a path towards more capable AI systems.

In an earlier interaction with AIM, Malhotra described his concept of "Dream AI", a neurosymbolic architecture that combines symbolic reasoning with deep reinforcement learning. He argued that today's AI systems excel at pattern recognition but lack a genuine understanding of the world.

"Symbolism helps us build world models, and Transformers enable actions based on those models," Malhotra had said.

According to him, world models allow AI systems to simulate environments and reason about consequences before taking action, reducing dependence on massive datasets and enabling more contextual decision-making. The approach combines symbolic AI, reinforcement learning, and neural networks to create agents capable of learning from simulated experiences rather than purely from historical data.

The research areas now being pursued by Makers Lab suggest that several of the ideas behind Dream AI are moving from concept to active development. Alongside world models, the team is also investigating new VLM architectures and broader AGI-related capabilities.