Ask not "Why" but "Why Not": Harnessing the Power of Reasoning LLMs 💪⛽

Ask not "Why" but "Why Not": Harnessing the Power of Reasoning LLMs

By Leo Kee Chye

With a wave of advanced reasoning Large Language Models (LLMs) such as OpenAI's o3 mini, DeepSeek R1, Google’s experimental Gemini 2.0 Flash, xAI’s Grok 3, Perplexity’s reasoning and deep research model and Alibaba’s QwQ entering the market, many users may find ourselves overwhelmed—especially as we are still exploring the full potential of earlier models.

This article argues that the rise of these sophisticated reasoning LLMs will dramatically enhance researchers’ capabilities. Rather than simply seeking answers from these models, researchers should focus on refining the questions we ask. Furthermore, instead of limiting our inquiries to “What,” “How,” and “Why,” we should also ask, “Why not?”

Understanding Reasoning LLMs

Core Features and Evolution

Reasoning LLMs are designed to think step-by-step, deconstruct complex problems into components, solve them either sequentially or concurrently, and reflect on their methods to optimise their approach or explore alternatives.

While earlier models required explicit instructions for step-by-step thinking, modern reasoning models have built-in Chain-of-Thought (CoT) processing that runs automatically for every prompt. However, these capabilities come with longer response times and higher token consumption, leading to increased costs (OpenAI, 2025; OpenAI, 2024).

Training Process and Emergence

According to Karpathy (2025), reasoning capabilities emerge primarily in the third stage of training. The process involves:

1. Pre-training

2. Supervised Fine-tuning

3. Reinforcement Learning

During reinforcement learning, models develop independent problem-solving methods with minimal human assistance. The models may deviate from conventional human approaches, potentially creating specialised reasoning processes that could be incomprehensible to humans.

Notable Examples

DeepSeek R1 Breakthrough

A significant "Aha" moment occurred during the development of the DeepSeek R1 reasoning model. During reinforcement training, the model independently discovered it needed more Test-Time Compute (TTM) for accurate results (Gao et al., 2025).

AlphaGo's Innovation

AlphaGo's Move 37 against Lee Se-dol demonstrated unprecedented strategic creativity, showing that AI systems can explore novel solutions beyond human strategies. This breakthrough expanded the understanding of AI capabilities and its potential for enhancing human strategic thinking (Zarkadakis, 2016).

Tesla's Neural Network Advancement

Tesla's Full Self-Driving (FSD) version 12 implements an end-to-end neural network that learns directly from video clips of human driving. This system replaces over 300,000 lines of code with AI-driven decisions, enabling adaptation to various scenarios without explicit programming (Ramey, 2024).

How Do Reasoning LLMs Empower Researchers?

The advantages of these models include:

• Interpreting complex datasets

• Proposing experimental setups

• Deriving math solutions or proofs

• Solving advanced domain-specific problems

• Writing and debugging specialised code

• Synthesising insights from multiple research papers across disciplines

• Encouraging cross-pollination of ideas between fields

What and How to Ask? A Simplified Model of Scientific Inquiry

Level One: Recognising Common Patterns

Identify shared features among different phenomena.

• Example: “What is a dustbin?” or “What do dustbins have in common?”

Level Two: Juxtaposing Dissimilar Concepts

Combine unrelated ideas to uncover new synergies or perspectives.

• Example: “How can ‘dustbin’ be used figuratively?” or “How can the placement of fewer dustbins optimise waste collection efficiency?”

Level Three: Flipping an Idea on Its Head

Challenge fundamental assumptions to generate innovative solutions.

• Example: “Why must we rely on physical dustbins to keep streets clean? Can we eliminate them while maintaining hygiene and convenience?”

More Real-World Cases of AI in Action

Level One Inquiry: Uncovering Hidden Patterns

Deep neural networks (DNNs) have achieved high accuracy in predicting gender from retinal fundus images—a task considered nearly impossible by ophthalmologists (Indu Ilanchezian et al., 2021). This discovery suggests AI can reveal subtle biomarker differences between genders and possibly other undetected health indicators.

Level Two Inquiry: Thinking Outside the Box

A research team developed a small mechanical robot capable of crawling on its legs. The AI was programmed to minimise the number of steps taken to reach a destination. However, instead of walking conventionally, the robot flipped onto its back and used its elbows to manoeuvre forward—an unexpected but highly efficient solution (Cully, Antoine, et al., 2015).

https://goo.gl/9cwFtw

Similarly, deep learning has challenged a long-held forensic assumption: that no two fingerprints, even from the same person, are alike. AI has demonstrated that ridge orientation patterns, rather than minutiae points, are key to cross-finger matching (Guo et al., 2024).

Level Three Inquiry: Revolutionising Historical Analysis

A 22-year-old, Luke Farritor, designed an AI programme that successfully deciphered previously unreadable ancient scrolls from Pompeii, providing new insights into classical history (Steinberg, 2024).

Reasoning LLMs shift the focus from seeking answers to asking better questions.

For example, if I hand you a toy car, you’d likely push it along a surface. A child, however, might make it fly through the air. Perhaps we need to think like children again—to ask not just why, but why not?

Conclusion

The power of reasoning LLMs lies not in the answers they provide, but in how they inspire us to rethink our questions. By challenging conventional assumptions and adopting a "Why not?" mindset, researchers can unlock new realms of knowledge and innovation. 

A breakthrough in computer science emerged from an unexpected source when undergraduate computer engineer Andrew Krapivin developed an innovative hash table design that accelerates data searches. His achievement was particularly remarkable as it disproved a long-standing Yao’s conjecture about hash table efficiency that had stood unchallenged for four decades. Ironically, Krapivin's fresh perspective came from his unfamiliarity with the established theory – his ignorance of the conventional wisdom freed him to explore solutions that others might have dismissed (Nadis, 2025). 

“Why Not” childlike inquiry may possibly be the mother of discovery.

References:

 

Cully A, Clune J, Tarapore D, Mouret JB. Robots that can adapt like animals. Nature. 2015;521(7553):503–507.

Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., ... & He, Y. (2025). Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.

Guo, G., Ray, A., Izydorczak, M., Goldfeder, J., Lipson, H. and Xu, W. (2024). Unveiling intra-person fingerprint similarity via deep contrastive learning. Science Advances, [online] 10(2). doi:https://doi.org/10.1126/sciadv.adi0329.

Indu Ilanchezian, Kobak, D., Faber, H., Focke Ziemssen, Berens, P. and Murat Seçkin Ayhan (2021). Interpretable gender classification from retinal fundus images using BagNets. medRxiv (Cold Spring Harbor Laboratory). [online] doi:https://doi.org/10.1101/2021.06.21.21259243.

Karpathy, Andrej (2025). Deep Dive into LLMs like ChatGPT . YouTube. https://youtu.be/7xTGNNLPyMI?si=hcu7xSObJz4F8DVs

Nadis, S. (2025). Undergraduate Upends a 40-Year-Old Data Science Conjecture. [online] Quanta Magazine. Available at: https://www.quantamagazine.org/undergraduate-upends-a-40-year-old-data-science-conjecture-20250210/ [Accessed 12 Feb. 2025].

‌Openai.com. (2024). Learning to reason with LLMs. [online] Available at: https://openai.com/index/learning-to-reason-with-llms/ [Accessed 11 Feb. 2025].

Openai.com. (2025). OpenAI Platform. [online] Available at: https://platform.openai.com/docs/guides/reasoning [Accessed 11 Feb. 2025].

Ramey, J. (2024). Tesla Bets on AI in Latest FSD Update. [online] Autoweek. Available at: https://www.autoweek.com/news/a46535912/tesla-fsd-ai-neural-networks-update/ [Accessed 17 Feb. 2025].

Steinberg, J. (2024). The 22-Year-Old Who Unlocked the Secrets of Ancient Rome. [online] Thefp.com. Available at: https://www.thefp.com/p/luke-farritor-vesuvius-challenge-scrolls-rome [Accessed 11 Feb. 2025].

Zarkadakis, G. (2016). Move 37, or how AI can change the world. [online] HuffPost. Available at: https://www.huffpost.com/entry/move-37-or-how-ai-can-change-the-world_b_58399703e4b0a79f7433b675 [Accessed 17 Feb. 2025].