Do LLMs really reason?

LLMs can perform reasoning operations and solve some problems through multiple steps, but their reliability depends on the task and the evaluation conditions. A correct answer or a lengthy explanation is not enough to establish a general ability to reason about unfamiliar situations. Assessing an LLM’s reasoning requires examining its results, how well they hold up under changes, and the methods used to produce them.

What reasoning means for an LLM

In this article, reasoning means the ability to draw conclusions from information and rules. It can involve combining constraints, deducing a consequence, developing a plan, or looking for a counterexample. This definition describes an observable capability; it assumes neither consciousness nor subjective experience.

A simple problem helps make the issue concrete. Imagine a hypothetical exercise in which documents must pass through several departments, subject to restrictions and a required sequence. A correct solution must satisfy every constraint, including those that rule out the most familiar route.

The model might succeed in several ways: by recognizing a familiar exercise, adapting a learned method, or constructing a solution from the supplied rules. These possibilities can overlap. Acquired knowledge can provide the building blocks for a new solution.

Several questions therefore need to be considered separately:

• Does the result satisfy the constraints?
• Does the method still work when the problem changes?
• Does the explanation make it possible to check the solution?
• What resources were needed to succeed?

The debate becomes confused when a single example is used to answer all these questions. Success demonstrates a capability under particular conditions. It does not yet establish that capability’s scope, consistency, or underlying mechanism.

A chain of thought, meaning a sequence of intermediate steps expressed in language, is one approach researchers have studied to improve problem solving. It also provides material to examine, with a limitation: an explanation does not give direct access to all the computations happening inside the network.

Predicting text can help solve a problem

The objection that an LLM merely predicts the next piece of text does not settle the question. It describes a principle of generation without establishing, by itself, what the system can compute. Determining its capabilities requires observing its behavior on controlled tasks.

The paper Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, submitted in January 2022, shows that providing worked examples with intermediate steps improves results on several arithmetic, symbolic, and commonsense reasoning tasks. The finding applies to the models and evaluation procedures studied. It does not mean that every request for an explanation improves every answer.

The generated steps can support computation: state an assumption, retain an intermediate result, and then use it to continue. An error can also propagate through those steps. Their presence therefore offers a way to solve and check a problem, without certifying that the solution is valid.

Other research examines the training process itself. A DeepSeek paper published in Nature in September 2025 describes how reinforcement learning can encourage verification and changes in strategy on tasks with verifiable outcomes. These observations depend on the setup studied and the reward signals it used.

There is experimental evidence, then, that certain methods improve capabilities. The scope of that evidence remains specific: solving more verifiable problems does not automatically demonstrate equivalent progress on ambiguous requests, incomplete information, or goals that are difficult to evaluate.

A convincing explanation can hide faulty reasoning

An answer and its justification need separate assessments. A model can reach the correct result through an invalid argument, or describe an appropriate method and then make a calculation error. In either case, reading only the conclusion discards useful information.

Another difficulty concerns the explanation’s faithfulness: does it describe the factors that actually influenced the answer? In an Anthropic study published in April 2025, researchers introduce hints into questions and then examine whether models acknowledge their influence. Under the tested conditions, explanations often omit hints that nevertheless changed the answer.

This limits the use of chains of thought for monitoring. The study does not establish that every explanation is false or useless. It shows that an explanation cannot be assumed faithful simply because it is detailed and readable.

What is being assessed	What provides evidence	What remains insufficient
Accuracy of the result	Independent verification of the answer	A confident tone
Validity of the method	Checking the steps and assumptions	A lengthy justification
Robustness	Success on controlled variations	Success on an isolated example
Faithfulness of the explanation	Experiments on factors influencing the answer	A plausible account of the process

For practical use, ask primarily for things you can check: assumptions, reproducible calculations, satisfied constraints, sources, and tests. A concise explanation accompanied by verification can be more useful than a long discussion that offers no reliable way to check its claims.

This distinction also helps when assessing public demonstrations. Showing a successful solution is informative. Showing what happens when the data, constraints, and wording change makes it easier to understand the limits of that success.

Generalization separates a method from a fragile recipe

A reasoning method should withstand relevant changes to a problem. Replacing a name should not alter a logical conclusion. Adding a decisive constraint, by contrast, should be able to change the solution.

The GSM-Symbolic study, submitted in October 2024 and revised in August 2025, examines controlled variations of mathematical problems. Among other findings, the authors observe performance variability when the data changes, along with difficulties when information irrelevant to the calculation is added. These results concern the systems evaluated, rather than every model available in October 2026.

The methodological lesson remains useful: test the rule, not just the familiar wording. A slightly modified problem can reveal dependence on habitual phrasing or associations that no longer apply.

For example, you could construct a hypothetical appointment scheduling exercise. Start with compatible availability, then add a conflict, remove a necessary piece of information, or deliberately make the request impossible to fulfill. The expected behavior changes: sometimes the system should propose a schedule, sometimes ask for clarification, and sometimes recognize that no solution exists.

This approach examines generalization: what a system can usefully carry over when it encounters a different situation. It also helps avoid confusing how difficult a problem looks with how difficult it proves to be in testing.

Failure on a variation does not establish that the model memorizes everything. It might also indicate a misinterpretation, an unsuitable strategy, or insufficient resources. Distinguishing between these explanations requires repeated trials, changing only one element at a time.

What François Chollet proposes measuring

In On the Measure of Intelligence, published on arXiv in November 2019, François Chollet distinguishes acquired proficiency on a task from the efficiency with which a system acquires new skills. His proposal accounts for the prior knowledge and experience available to solve unfamiliar problems.

Applied to LLMs, this approach encourages looking beyond the final result. A system trained on many similar exercises and a system that discovers a rule from a few examples might succeed at the same task without demonstrating the same efficiency of adaptation.

This distinction does not make specialized skills useless. It clarifies what those skills allow us to claim about general intelligence. Our explanation of the essay on measuring intelligence explores this framework in more detail.

The ARC-AGI benchmark, presented by ARC Prize, examines, among other things, the ability to infer transformations from examples. Its grid tasks require finding a rule consistent with the demonstrations and then applying it to a new input.

Even within this framework, a score needs to come with its evaluation protocol: accessible data, permitted attempts, tools, and computing resources. The ARC-AGI reference page explains the test and its different versions.

To place this work alongside other approaches to artificial general intelligence, the AGI and research section brings together the main questions about measurement. No isolated test can replace examining a range of capabilities under documented conditions.

Evaluate the complete system and its problem-solving costs

A result attributed to a model may depend on a broader setup. That setup might generate several candidate solutions, execute code, consult documents, or use a verifier to select an answer. An evaluation needs to specify what was available.

These are legitimate ways to build a useful tool. They do, however, change what the measurement covers: a solution validated by a program tells us about the complete system, while an answer produced without tools tests a different configuration.

The budget matters too. Allowing more attempts or steps can change the results, so comparisons need explicit conditions. The number of successful solutions, the time taken, and the resources consumed together describe the practical value of a method.

For your own evaluation, assemble a set of exercises with an expected answer or a precise success criterion. Include routine problems, edge cases, and impossible requests. Keep the instructions, settings, and outputs so that you can examine discrepancies later.

Then distinguish the types of failure: misreading the request, overlooking a constraint, making a calculation error, misusing a tool, or answering despite missing information. This analysis points more clearly to what needs fixing than an overall score does.

The article on LLM benchmarks and interpreting rankings explains these precautions. When moving from a model to a system that acts through tools, the guide to evaluating an AI agent offers an approach suited to real tasks.

What you can expect from a reasoning model

To use these capabilities effectively, start by expressing the problem in a way that allows the answer to be checked. Provide the constraints, the available information, and the expected result. Also specify when the system should ask for clarification or report that it cannot reach a conclusion.

Separate generating the solution from checking it. A model can propose a calculation or a program; an independent check should examine the result. For a consequential decision, the quality of the prose cannot replace the evidence that decision requires.

Finally, choose the amount of computation to suit the need. Summarizing a passage and solving a problem with multiple constraints do not necessarily call for the same setup. Assess reasoning models by the improvement they deliver on your task, taking their response times and limitations into account.

The answer to the title’s question remains conditional, but useful: some behaviors meet criteria for reasoning without establishing a general, stable, and transparent capability. Before placing trust in an answer, look for a verified solution and a method that holds up when the problem changes.

Key takeaways

• LLMs can solve problems through multiple steps; the scope of that capability needs to be measured task by task.
• A readable explanation guarantees neither valid reasoning nor a faithful account of the factors that influenced the answer.
• Novel variations, impossible cases, and independent checks provide stronger evidence than an isolated demonstration.
• A result needs to be accompanied by details of the computing conditions, available tools, and problem-solving budget.

Frequently asked questions

Does an LLM understand what it says?

Understanding can mean a practical ability to use concepts or a subjective experience. Tests can examine the former, for example by asking a model to apply a rule in a different context. Those tests alone cannot settle the latter.

Why does a reasoning model get a simple question wrong?

What humans find difficult does not always predict what a model will find difficult. Ambiguity, unusual wording, or a mistaken initial assumption can lead to an error. Examine the conditions behind the failure before drawing a broad conclusion from a single response.

Should I always ask an LLM to think step by step?

That instruction offers no guarantee, and its usefulness depends on the model and the task. Compare results with and without it on examples whose solutions you already know. Give priority to explicit assumptions and reproducible checks.

Can you prove that an LLM reasons without knowing its training data?

Novel problems, controlled variations, and independent checks can provide strong evidence. Without access to the training data, however, ruling out all prior exposure remains difficult. Any conclusion should specify the tasks and conditions examined.