Frontier · Advanced
Reasoning models
Chain-of-thought, test-time compute and process supervision: why newer models think before answering.
Why it matters
Spending more computation at inference time has become a main lever for capability. It changes cost, latency and how you prompt.
Key ideas
- Chain-of-thought prompting
- Test-time compute
- Self-consistency
- Process vs outcome supervision
- When reasoning helps and when it does not
You will be able to
- Choose between fast and reasoning models
- Budget latency and tokens
- Check reasoning faithfulness
Primary reading
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (opens in a new tab)Wei et al. · 2022
- Self-Consistency Improves Chain of Thought Reasoning (opens in a new tab)Wang et al. · 2022
- Let's Verify Step by Step (opens in a new tab)Lightman et al. · 2023
Learn it here
Planned course