FollowAgents AI AGENTS DIRECTORY
Home / AI Agent Glossary / Large Reasoning Model (LRM)
Ecosystem & Emerging Terms

Large Reasoning Model (LRM)

Also called: LRM · reasoning model · thinking model

A large reasoning model (LRM) is a language model trained to work through a problem step by step in a long internal reasoning trace before giving its final answer.

A large reasoning model is still a language model, but one trained so that, before answering, it generates an extended chain of intermediate reasoning: breaking the problem down, trying an approach, checking it, backing up when it fails. Many products expose this as a "thinking" or "reasoning effort" mode.

The idea took off in 2024-2025 as vendors and open-weight labs shipped reasoning-tuned models, and it is now a standard tier next to faster non-reasoning models. Reasoning models tend to do noticeably better on math, coding, and multi-step logic, and they are commonly the model behind an agent's planning step.

The cost is latency and tokens: the reasoning trace is generated, and usually billed, before the visible answer. That is why most providers let you choose how much thinking to allow.

How it works

Two ingredients are usually cited. First, training with reinforcement learning on problems whose answers can be checked (rlvr), which rewards reasoning traces that reach correct answers. Second, spending more computation at answer time by letting the trace run longer (test-time-compute). The result is the model learning behaviors like verifying its own steps and trying alternatives. Providers differ in whether the raw trace is shown, summarized, or hidden.

Example

Asked to find why a test fails intermittently, a non-reasoning model may guess the most common cause. A reasoning model works through the code path, considers a race condition, checks whether the ordering it hypothesizes is possible, rejects one theory, and then reports the one that survives.

How it differs

LRM vs. prompting for chain-of-thought: chain-of-thought is a prompting technique you apply to any model; a reasoning model has the behavior trained in, so it reasons at length without being asked.

Common misconceptions

Often assumed: Reasoning models are always better, so use them for everything.
Actually: They cost more and respond slower, and for simple lookups, formatting, or classification the extra thinking adds little. Many agents use a fast model for routine steps and a reasoning model for hard ones.
Often assumed: The visible reasoning trace is a faithful record of how the model decided.
Actually: Research on reasoning traces suggests they can omit or misstate factors that influenced the answer, so treat them as a useful signal, not a guaranteed explanation.

FAQ

What is a large reasoning model?
A language model trained to generate a long step-by-step reasoning trace, including checking and revising, before it gives a final answer.
What is the difference between an LRM and a normal LLM?
An LRM has extended reasoning trained in and spends more tokens and time thinking; a standard LLM answers more directly, faster and cheaper, and reasons only if prompted to.
Are reasoning models better for AI agents?
Often for planning and hard decisions, but not for every step; mixing fast and reasoning models by task difficulty is common.

Last checked: 2026-09-20

Related terms

Test-Time Compute

Test-time compute is computation spent while a model answers a query — rather than during training — used to improve the result through longer chains of reasoning, multiple sampled attempts, search over options, or self-checking.

Chain-of-Thought (CoT)

Chain-of-thought is a prompting technique where a model is asked to write out its intermediate reasoning steps before giving a final answer, rather than producing the answer directly.

RLVR (Reinforcement Learning with Verifiable Rewards)

RLVR is reinforcement learning where the reward comes from an automatic check of the outcome, such as whether code passes its tests or a math answer matches the known solution.

System One Model

A System One model is a class of AI model, named after fast intuitive thinking, that answers structured questions in a single fast pass and returns typed values with calibrated probabilities instead of generated text.