What are the main approaches to scaling test-time compute?
LLM & AILevel: IntermediateTest-Time Compute Scaling
What are the main approaches to scaling test-time compute?
Hints
- Identify the stage: pretraining, adaptation, generation or evaluation.
- Name the inputs, constraints and a case where the approach fails.
Answer
Three primary TTC strategies:
1. Sequential thinking (chain-of-thought) The model generates a long thinking trace before answering. Simple and effective. All current reasoning models use this as the foundation.
2. Self-consistency / majority voting Generate N independent answers (N parallel samples), pick the most common. Works well when:
- You can sample cheaply
- Correct answers cluster together in output space
- Trade-off: N× cost for sublinear quality gains
3. Search / tree of thoughts The model explores multiple reasoning paths, evaluating partial solutions and backtracking when stuck. More complex but more powerful for hard combinatorial problems.
Root (problem)
/ | \
Path A Path B Path C
/ \ \
A1 A2 C1
[dead] [continue] [best]
MCTS (Monte Carlo Tree Search) is used in some advanced reasoning systems — the model acts as both the policy (which moves to explore) and the value function (which paths look promising).
Source and licence: JustInCache & contributors — Awesome AI Interviews · Apache-2.0