The stateless request-response mode of using a language model: a prompt goes in, one completion comes out, and the interaction ends — with no intermediate tool execution, environmental feedback, self-correction, or persistent state; it is the cheapest and most predictable inference pattern, suited to classification, extraction, translation, and summarisation, and serves as the baseline against which iterative agentic workflows are defined and evaluated.
Semantic Classification
Content
Definition
Single-turn inference is the simplest way to use a large language model: send one prompt, receive one completion, done. The model has no opportunity to observe the consequences of its output, call tools, gather missing information, or revise a mistake — everything it needs must be present in the prompt, and everything it produces must be right first time. This is the mode in which most production LLM workloads still run: classification, entity extraction, translation, summarisation, rewriting, and retrieval-augmented question answering where the retrieval happens before the model is invoked.
The pattern’s virtues are operational. Cost and latency are bounded and predictable (one forward pass over a known context), behaviour is easier to evaluate (one input maps to one output, so standard test sets and metrics apply), failure modes are contained (an error affects one response, not a cascading action sequence), and the interaction is stateless, which makes horizontal scaling and caching straightforward. Prompt Engineering — instructions, few-shot examples, and Chain of Thought elicitation — is the main lever for quality, since there is no second chance within the interaction.
Its defining contrast in this graph is with the Agentic Workflow: iterative loops in which the model acts, observes real feedback, and corrects course. Tasks whose difficulty exceeds what one forward pass can reliably deliver — multi-file code changes, research requiring search, anything needing environment interaction — motivate Autonomous Task Execution. The boundary has blurred somewhat with reasoning models that perform long internal deliberation within a single API call: such calls remain single-turn in interface terms (no external feedback), even though substantial computation happens inside the turn.
Current Landscape
-
Where it dominates: high-volume pipelines (moderation, tagging, extraction), latency-sensitive interfaces, and any task with a well-defined input-output contract; the majority of enterprise LLM spend remains single-turn.
-
Scaling within the turn: chain-of-thought prompting, self-consistency sampling, and inference-time reasoning all buy accuracy without adding interaction rounds — test-time compute inside a single turn.
-
When to escalate: empirically, agentic loops justify their 5-100x token cost when tasks require environmental feedback or verification; when they do not, single-turn baselines are frequently as accurate and far cheaper — a comparison every agent evaluation should include.
-
Evaluation role: benchmarks such as MMLU-style QA measure single-turn capability, while agentic benchmarks (e.g. SWE-bench) measure what looping adds; the delta between them quantifies the value of iteration for a task family.
-
Test-time compute is now a scaling axis within the turn: OpenAI’s o1 (2024) and o3 report performance that “consistently improves” with more thinking time; o1 reached 74.4% pass@1 on AIME 2024 versus GPT-4o’s 9.3%, entirely inside a single API call with no external feedback.
-
Small models plus better inference can beat larger ones: Stanford’s s1 (2025) used SFT on 1,000 examples plus “budget forcing” to exceed o1-preview on competition maths by up to 27%, a clean demonstration that inference-time strategy, not just parameters, drives single-turn accuracy.
-
Diminishing returns are real: 2025 analyses of o1-like models (QwQ, R1, LIMO) found longer chains of thought do not monotonically improve accuracy — correct solutions are often shorter than incorrect ones — so test-time compute has task-dependent optima rather than unlimited upside.
-
The interface distinction holds: reasoning models remain single-turn in interface terms (hidden deliberation, one completion), which is why the single-turn/agentic boundary is still defined by the presence of external tool execution and environmental feedback, not by internal compute.
Sources:
-
https://ashitaorbis.com/understanding-ai/wiki/test-time-compute/