At India’s biggest GenAI Summit for developers, MLDS 2025, Rohit Thakur, GenAI Lead at Synechron, explored the next phase of AI development, moving beyond mere text prediction to models capable of logical reasoning.
“LLMs today largely function on next-word prediction, which works well for a variety of NLP tasks,” said Thakur. “But not all tokens have the same information density. Some require much deeper reasoning, and that’s where new models are changing the game.”
Beyond Next-Word Prediction
For years, LLMs have been trained to predict the next word in a sequence, enabling tasks such as translation, summarisation, and chat-based interactions. However, this method has its limitations, particularly when handling complex reasoning tasks like mathematical problem-solving or multi-step logical deductions.
Thakur explained how Ashish Vaswani’s seminal paper, Attention is All You Need, introduced the transformer architecture in 2017, which laid the foundation for modern LLMs. The evolution from BERT to GPT to reinforcement learning with human feedback (RLHF) has improved models, enabling preference tuning. However, they still rely on probabilistic text generation rather than structured reasoning.
Chain of Thought reasoning, which breaks down multi-step problems, marked a breakthrough, but LLMs trained on next-word prediction continue to imitate reasoning rather than truly reasoning.
“There’s a difference between emulating reasoning and actually reasoning,” Thakur noted. “LLMs trained only on next-word prediction might arrive at the correct answer, but they lack the ability to logically deduce it in the same way a human would.”
From AI Assistants to Industrial Use Cases
The rise of ‘reasoning AI’ is already having tangible effects across industries. Thakur shared a case study from a manufacturing company, where an AI assistant was deployed for engineering-related chatbot interactions. Engineers frequently queried the chatbot for specifications requiring mathematical calculations.
“When we used a normal prompt meant for chat models, we got an incorrect output because some amount of calculation was needed—it said both models have these specifications.”
Traditional LLM-based assistants often failed because they relied on probabilistic text completion rather than actual computation. By implementing reasoning AI, the chatbot improved its responses, correctly filtering product models based on kilowatt power ranges.
This shift has implications beyond industrial applications. In fields such as finance, legal analysis, and healthcare, where AI-generated responses require verifiable logical steps, reasoning AI can significantly enhance accuracy and trustworthiness.
“Imagine an AI that doesn’t just provide an answer but actually explains how it arrived at it,” said Thakur. “That’s the leap we’re witnessing today.”
Interestingly, DeepSeek’s entry has reintroduced the focus on a structured reinforcement approach. Thakur explains how this method assists DeepSeek in building logical pathways rather than merely predicting probable answers.
Proprietary models like OpenAI’s latest versions and Google’s Gemini AI are exploring similar methods, although details remain limited. However, deploying such models isn’t straightforward—Ben Hilack’s analysis revealed that reasoning models require different prompt structures compared to traditional LLMs, emphasising the need for customised input designs.
The Road Ahead for AI That Thinks
While reasoning AI is still in its early stages, its trajectory is clear. The transition from statistical prediction to structured reasoning marks a significant milestone in AI development. Open-source projects such as DeepSeek and refinements in reinforcement learning-based training will continue to push the boundaries of what AI can achieve.
However, challenges remain. Future research should focus on ensuring that models generalise across diverse problem sets, avoid biases in reasoning pathways, and remain computationally efficient. Moreover, as Thakur highlighted, the shift towards task-specific prompting strategies will be crucial in maximising the potential of these models.
“The future of AI isn’t just about predicting words—it’s about understanding and reasoning. And we’re only at the beginning of that journey,” he concluded. You can read more about Synechron’s transformative AI solutions here.
