You already know: chain-of-thought and self-consistency voting (D7 Advanced prompting); reward models and how they get hacked (D6 Preference optimization); policy gradients and GRPO with verifiable rewards (D6 RL for LLMs); reflection loops that need a grounded critic (Agent architectures); binomial probabilities (D2 Probability).
Why this matters for Atlas. For years the main way to get a smarter model was to train a bigger one. Test-time compute is the other lever: spend more computation per question, by generating longer reasoning, sampling several answers and voting, checking candidates with a verifier, or searching through possible reasoning steps. Reasoning models (OpenAI's o-series, DeepSeek-R1, Claude with extended thinking) are trained with RL to use this lever themselves. It's how a model gets a hard maths proof right, and it's why some answers take a minute and cost 20× more. This lesson builds each method, measures exactly what it buys, and shows when it's worth paying for.
1. Two kinds of thinking, and why more samples help
The question. If a model is right 60% of the time on a question, can it ever be right more often than that without retraining?
Start from nothing. A standard model answers in one pass, token by token: fast, intuitive, "System 1". Some problems need "System 2": trying an approach, noticing it fails, backtracking, checking. You can buy System 2 behaviour at inference time in two ways: