Aryabhata 2: Scaling Reinforcement Learning for Advanced STEM Reasoning

Aryabhata 2 is a specialized language model designed for competitive STEM examinations that uses reinforcement learning to improve reasoning capabilities while reducing computational output by up to 64%. Trained on PhysicsWallah's question banks, it outperforms its base model on JEE and NEET exams, addressing the practical challenge of deploying AI at scale for educational applications.

Aryabhata 2 represents a shift in AI development toward domain-specific optimization rather than general-purpose scaling. The model addresses a genuine market need: educational institutions and edtech platforms struggle to deploy large language models cost-effectively for millions of student queries. By fine-tuning GPT-OSS-20B through reinforcement learning with verifiable rewards—a technique gaining traction in AI development—the researchers achieved superior performance while dramatically reducing computational overhead. This efficiency gain matters because it directly impacts deployment viability in resource-constrained educational settings across developing economies where PhysicsWallah operates. The curriculum-based training approach using high-quality internal question banks demonstrates how proprietary datasets create competitive advantages in specialized domains. Aryabhata 2's strong performance across both in-distribution benchmarks (JEE, NEET) and challenging out-of-distribution datasets (AIME, GPQA) suggests genuine reasoning improvement rather than test-specific overfitting. For the broader AI industry, this work validates reinforcement learning post-training as a practical method for improving reasoning without proportional compute increases. The edtech sector specifically gains a blueprint for deploying AI tutoring systems at scale. Looking ahead, similar domain-specific optimization approaches will likely proliferate across professional licensing exams, medical training, and technical certifications, fragmenting the market away from one-size-fits-all foundation models toward specialized alternatives.

• →Aryabhata 2 reduces output tokens by up to 64% while improving STEM reasoning performance through reinforcement learning
• →Domain-specific fine-tuning on verified question banks creates competitive advantages for edtech applications
• →The model's strong performance on out-of-distribution benchmarks indicates genuine reasoning improvements beyond test overfitting
• →Efficient deployment addresses the practical challenge of scaling AI for educational use in resource-constrained markets
• →Reinforcement learning with verifiable rewards emerges as a viable alternative to scaling model size for performance gains