Slow Thinking for Models

The Extra Budget

A trained language model can answer a prompt directly, or it can spend extra work before committing to the answer. That extra work is test-time reasoning: computation used during inference rather than during training.

Why It Helps

Hard tasks often need intermediate state. Arithmetic, coding, planning, and multi-hop questions can require partial results, comparisons, or checks. Reasoning tokens, samples, verifiers, and tools give the system ways to hold and test those intermediate steps.

Length Is Not the Mechanism

The useful part is not simply producing more text. Extra computation helps when it changes the distribution of possible answers toward better ones: by decomposing the task, exploring alternatives, rejecting weak paths, or delegating exact substeps.

The Design Question

A reasoning system has to decide when to spend the budget. Easy prompts should stay cheap. Hard prompts may deserve longer traces, more samples, revision loops, tool calls, or stronger verification before the answer is trusted.

What is the practical purpose of test-time thinking in an LLM system?

Answer: It lets the system spend additional computation during inference on decomposition, search, checking, revision, or tool use before accepting a final answer.

The course treats thinking time as an inference resource. It is useful when the extra work changes answer quality, not when it merely makes the response longer.

Which scenario best illustrates useful test-time thinking?

• A model is pretrained on a larger corpus before deployment.
• A model generates several solution paths, checks them with a verifier, and returns the highest-scoring answer.
• A model always writes five paragraphs even for one-word factual questions.
• A model changes its UI label from answer to reasoning.

Answer: A model generates several solution paths, checks them with a verifier, and returns the highest-scoring answer.

This spends inference-time computation on search and selection, which can shift the final answer toward a better candidate.

A reasoning system should usually spend the same long chain-of-thought budget on every prompt, regardless of difficulty.

• True
• False

Answer: False

Weng emphasizes adaptive thinking time. Easy prompts should stay cheap; hard prompts may justify more samples, tools, or revision.