Complete AI Training
Sign inGet my AI kit

Your job's AI kit

Get your AI kit

Tell us who you are and what you do. We show you your kit right away and email you the link: skills, prompts, AI agents, MCP servers and courses for your job.

500+ jobs ready, and we make a kit for any other job. No payment needed to look.

Share

Video course · 11 chapters · 63 min · certificate

Test-Time Compute Scaling: Smarter LLMs Without Retraining

Make LLMs smarter at inference without retraining: test-time compute scaling laws, repeated sampling and coverage, the shift in compute allocation, parallel versus sequential methods, reward and process reward models, and architecture search for inference.

Become a memberAlready a member? Sign in

What you'll learn

  • Explain the three stages of LLM development
  • Describe repeated sampling and coverage
  • Explain the shift from pre-training to test-time compute
  • Compare parallel sampling and sequential revision
  • Use outcome and process reward models
  • Optimize inference with architecture search

Chapters

11 chapters · 63:20
  1. 2:20 01Intro Members Inference scaling The three stages of LLM development and test-time compute scaling laws.
  2. 2:40 02Sampling Members Repeated sampling and coverage More samples let a model match much stronger models on some tasks.
  3. 6:00 03Economics Members A shift in compute allocation Predictable pre-training scaling, and the shift toward spending compute at inference.
  4. 9:00 04Applications Members Formal proofs and code Why verifiable domains like formal proofs and code benefit most.
  5. 7:38 05Research Members Knowledge graphs and efficiency Research directions: building on test-time scaling and analysing efficiency.
  6. 1:05 06Methods Members Parallel and sequential methods Parallel sampling versus sequential revisions of an answer.
  7. 3:23 07Rewards Members Process reward models Selecting answers by outcome versus scoring each reasoning step.
  8. 7:54 08Analysis Members Analysing test-time scaling Step-level scoring and how to analyse where test-time compute pays off.
  9. 5:47 09Trade-offs Members Pre-training versus test-time compute Pre-training happens once; test-time cost recurs on every run.
  10. 4:13 10Optimization Members Architecture search for inference An architecture search framework for combining inference-time techniques.
  11. 13:20 11Techniques Members Inference techniques and optimization Designing mechanisms to reach high-quality answers without wasting tokens.