Test-time compute is an AI concept where the model spends more time thinking at inference (when answering a question) instead of just at training. Tech jargon. Not concerning. If your kid uses it constantly, ask them where they first heard it, not to police their speech but to understand which online community they are in. Most slang of this kind moves through TikTok and group chats, so seeing it in your child's vocabulary just means they are plugged into normal teen internet culture.
test-time compute
Using more compute at inference time — via longer chain-of-thought, sampling, or verification — to boost model quality without retraining.
Etymology timeline
Born ~2024 · no dated citations yetSeen in the wild
This reflection is a form of "test-time compute", where additional computational resources are used during inference.
Some models also exhibit performance gains by scaling inference through increased test-time compute (TTC), extending neural scaling laws beyond training to the deployment phase.
Definitions
Using more compute at inference time — via longer chain-of-thought, sampling, or verification — to boost model quality without retraining. The OpenAI o1 + Claude extended-thinking paradigm. You pay seconds to gain accuracy.
"We're throwing more test-time compute at the eval. 3x latency, much better answers."
Test-time compute is an AI concept where the model spends more time thinking at inference (when answering a question) instead of just at training. Tech jargon. Not concerning. If your kid uses it constantly, ask them where they first heard it, not to police their speech but to understand which online community they are in. Most slang of this kind moves through TikTok and group chats, so seeing it in your child's vocabulary just means they are plugged into normal teen internet culture.
Test-time compute is a 2024 AI-industry term meaning the computational resources used during inference, particularly when models engage in extended reasoning. Use in AI research and technical writing. Formal equivalent: "inference-time computation." Common non-native mistake: confusing with training-time compute. Test-time compute happens when answering queries, not during training. Read or listen to the word in context several times before trying it yourself; the right register is harder to learn than the dictionary meaning.
“This reflection is a form of "test-time compute", where additional computational resources are used during inference.”
“Some models also exhibit performance gains by scaling inference through increased test-time compute (TTC), extending neural scaling laws beyond training to the deployment phase.”
Comments
The amount of computational power or processing time required for an artificial intelligence system to generate a result or perform a task after it has already been trained.
Comments
For specific readers
Why is this section here?
Slang is read by different audiences for very different reasons. The For parents view re-explains the term in plain, non-judgemental language so a caregiver can understand what their child means without needing the surrounding subculture. The For English learners view focuses on register (when it's safe to use), formal alternatives, and common mistakes — what a textbook won't teach you.
Used in
Have a better definition?
Add your own interpretation of "test-time compute".
⚡About Tech, Dev & AI slang
The vocabulary of software engineers, AI researchers, and anyone living in a terminal or on GitHub — from LLM to MCP, CORS to vibe coding, agentic to enshittification.
See all Tech, Dev & AI slang on Slangora.
Tags
Other words near "test-time compute" (T)
Browse all slang words starting with T.
Related terms
People also read
Semantically related
- temperature75%
An LLM setting that controls randomness. Temperature 0 = deterministic, always pick the most likely next token. Higher values = more creative / less predictable.
- prompt engineering75%
The craft of writing inputs to LLMs to reliably produce the output you want. Includes role-setting, few-shot examples, chain-of-thought, and careful phrasing.
- reasoning model72%
An LLM trained or configured to produce long chains of internal "thinking" tokens before answering — trading latency for higher accuracy on complex problems. OpenAI's o1 and Claude's extended thinking are examples.
- tool use72%
The ability of an LLM to call external functions (search, code execution, databases) mid-conversation. The building block of agentic systems.