Reasoning AI: Cracking the IIT test is a giant leap for machine-kind

These headlines are striking not just because machines can now ace exams designed to pick the top one in a thousand students, but because AI has acquired the ability to reason.

From language to logic: The first generation of large language models were, at heart, articulate parrots. They could produce convincing sentences because they had been trained on vast libraries of text. But they lacked a systematic way of thinking. Confronted with problems that demanded multiple steps in mathematics, science or logic, they crumbled.

Also Read: SubwayTakes, Jimmy Fallon and the Jonas Brothers may not be enough for Google to attract AI sceptics

Two innovations changed that. The first was reinforcement learning. Instead of training AI only to mimic  patterns in data, researchers began to reward it for following coherent chains of reasoning. Models learned not just to provide the right answer, but to show their work step-by-step. This nudged them away from guessing and towards disciplined thought.

The second was tool-use. Newer systems were given the ability to call on external programs, run a Python script to crunch numbers or use a browser to fetch information. The hybrid approach mirrored how humans solve problems: intuition first, then calculation and then fact-checking. A model that once ‘hallucinated’ answers could now defer to a calculator; one that once relied only on memory could consult the web. This turned them from eloquent talkers into capable problem-solvers.

Also Read: Is an AI winter upon us? There seems a chill in the air

Practice makes genius: Olympiad hopefuls do problem-set drills for years; so do modern AI models, except at superhuman speed. They are trained on vast collections of worked examples, from Olympiad papers and university-level problem sets to synthetic challenges generated by other AIs. As they can attempt millions of problems in weeks, they develop an almost athletic familiarity with different types of reasoning.

Exams provided the perfect training ground. Their strict rules and objective answers forced models to refine their reasoning until it was reliable. What began as preparation for geometry proofs or calculus questions quickly spilled over into other domains. Once you have learnt how to reason carefully, the skill is transferable.

Reasoning everywhere: Reasoning is not confined to mathematics. It is the connective tissue of human decision-making. A scientist designing an experiment, an auditor tracing a discrepancy or a regulator weighing the impact of a tax policy all rely on step-by-step logic. An AI that can reason is not just a better exam-taker, it can be a collaborator in science, business and government.

In the sciences, reasoning machines could help sift through vast data-sets to spot hypotheses worth testing or simulate molecular interactions in search of new drugs. In finance and accounting, they could trace numbers through complex systems, flag anomalies or model risks with a precision that would make regulators envious. In operations, they could untangle the complex web of supply chains, hospital workflows or logistics networks. The Olympiad gold signals more than triumph in a contest. It shows the arrival of machines able to shoulder a part of the cognitive load across fields.

Also Read: Free-AI: How well India adopts AI will shape its future

A new benchmark: It is telling that technology companies now use exams as public demonstrations. In the 1990s, chess was the test of AI and in the 2010s, it was the game Go. Both were rule-bound games with limited practical relevance. Mathematics, by contrast, is a universal language of reasoning. To solve Olympiad problems is to show mastery in general problem-solving skills. These scores reflect a shift from machines that sound intelligent to machines that can actually ‘think.’

The road ahead: The next phase will be about applying reasoning to overcome real-world challenges. In healthcare, reasoning AIs could help doctors trace symptoms through complex case histories or design treatment plans that account for multiple conditions. In science, they could accelerate discovery by testing thousands of hypotheses in silicon before a single experiment is run in a lab. Businesses may use them to map supply chains, model financial risks or design more efficient processes. Governments too could benefit. India often faces intricate choices in trade, energy and foreign policy. These are not simple calculations but judgements about competing costs and benefits. Reasoning machines can help policymakers navigate these dilemmas.

A few years ago, these systems failed at basic arithmetic. Today, they match Olympiad champions. Tomorrow, they may be indispensable in research and policy. The challenge is not to compete with language models in the exam hall, but to put their growing capabilities to work on the problems that matter most.

The authors are, respectively, former chief technology officer of Aadhaar and co-founder of eGovernments Foundation; and director of engineering, Klarity.ai