Online course · advanced
Reasoning Models & Test-Time Compute
Understand chain-of-thought, test-time compute, self-consistency, and verifiers. Measure when a reasoning model is worth its cost and latency, using offline labs that need no paid API.
With Admin
THE TAKEAWAYS
What you’ll learn
- Explain how intermediate reasoning tokens give a model more computation per answer, and where that does not help.
- Write a chain-of-thought prompt whose steps can be checked by code.
- Distinguish a reasoning trace from an explanation of the model's actual computation.
- Calculate the cost and latency of serial and parallel test-time compute from dated, user-supplied prices.
- Compute majority-vote accuracy under independence and explain how correlated errors break it.
- Compare outcome and process checks, and design a problem-specific verifier.
- Describe reinforcement learning with verifiable rewards and its risk of reward hacking.
- Choose between configurations using a labelled evaluation set, recording accuracy, cost, and latency together.
- Deliver a measured reasoning decision report with explicit limitations.
Is this for you?
Developers and analysts who can prompt an LLM and write basic Python, and who need to decide when extra inference compute improves a multi-step task.
Before you start
- Complete LLMs: From Prompts to Practice, or demonstrate prompt design and basic evaluation.
- Python 3.10 or newer; comfort with functions, lists, dictionaries, and running a script.
- No paid API or GPU required. Measuring a real model is optional and must be labelled with its name, settings, and date.
PUT IT TO WORK
What you’ll build
Reasoning-model decision report. Client scenario: a team must decide how to answer a recurring multi-step question reliably and affordably. Deliverable: an evaluation set, strategy comparison, budget, custom verifier, and decision memo.
Learn what reasoning models do differently, why a visible reasoning trace is evidence to check rather than an audit log, and how sampling, majority voting, and verifiers spend inference compute. Work through an offline lab that checks arithmetic steps, counts votes, and computes budgets over hand-written fictional traces, including a correlated mistake that defeats voting. Finish by producing a measured decision report for one multi-step task without claiming results you did not run.
Course curriculum
SECTION 1What reasoning models do and how far to trust the trace3 lessons
- What a reasoning model isReading
- Chain-of-thought promptingReading
- Is the reasoning faithful?Reading
SECTION 2Spending inference compute: sampling, voting, and verification3 lessons
- Test-time compute and its trade-offsReading
- Self-consistency: sampling and votingReading
- Verifiers: checking outcomes and processesReading
SECTION 3Training, model choice, and a measured decision3 lessons
- How reasoning behaviour is trainedReading
- Choosing and budgeting a reasoning modelReading
- Capstone: a measured reasoning decision reportReading