TTC (Test-Time Compute)

Test-Time Compute (TTC) is the strategy of spending more computational resources at inference time (when a model is answering a question) rather than at training time (when a model learns from data) to improve output quality, reasoning depth, and solution accuracy. This inverts the scaling intuition from the era of large language models, where bigger models trained on more data generally performed better. The key insight is that reasoning can be "thought out" as well as "learned in"—sometimes it's better to allocate 1000 inference steps to a single problem than to train a larger model.

Technically, test-time compute manifests through several mechanisms. Chain-of-thought prompting (asking the model to "think step-by-step") is a lightweight form, expanding reasoning tokens without extra training. Beam search, where the model generates multiple candidate solutions and ranks them, is another approach. More recent models like OpenAI's o1 and o3 employ reinforcement learning during inference, allowing models to search through reasoning pathways, backtrack when stuck, and verify intermediate steps. Some frameworks use ensembles, running multiple inference passes and aggregating results. Others use verification loops, where a language model proposes a solution and a second system checks its correctness. All of these techniques increase the compute spend at test time while keeping the base model fixed or even smaller.

TTC matters because it suggests a new scaling frontier: instead of doubling model parameters every year, systems can allocate compute dynamically based on problem difficulty. A simple query might need minimal reasoning; a complex math problem might benefit from 100x more compute. This enables more efficient systems—smaller, cheaper base models augmented with adaptive reasoning. It also changes how we think about AI capability: a system that reasons longer isn't necessarily smarter, but it's more thorough. However, TTC increases latency and cost per query, creating new tradeoffs between speed, accuracy, and expense.