Reasoning

Reasoning describes an LLM's ability to work through a problem in multiple steps rather than just predicting the next Token in one shot. With OpenAI's o1 and DeepSeek's R1 in 2024, 'reasoning models' became a recognised category — they generate a hidden Chain-of-Thought before producing the final answer, often spending thousands of tokens of internal deliberation. This delivers striking gains on math, code and logic-heavy benchmarks, but at the cost of significant test-time-compute, Latency and price. Tree-of-Thought, Self-Critique and tighter evaluation loops are natural extensions of the same paradigm.