Loading...
Patterson Consulting Contact Us

Do Large Language Models Reason?

By: Josh Patterson
July 2026

Ask a modern language model a difficult question and inspect what it produces. The response is structured like an argument: premises, inferences, conclusions, sometimes even explicit revisions of earlier steps. Some systems appear to reconsider intermediate results before settling on a final answer.

From the outside, it looks remarkably like reasoning, which begs the question:

"do large language models reason?"

And that resemblance is what makes the question difficult.

The resemblance is genuine, not accidental, but resemblance alone cannot establish that reasoning has occurred. An earlier article in this series, What Is Reasoning?, defined the term carefully. That definition provides the standard against which language models can be evaluated, and it is important to be clear about both what it includes and what it excludes.

The classical account in logic defines reasoning in terms of inference: deduction, induction, and abduction. Judged only by that structure, a fluent chain of premises and conclusions appears to qualify.

But inferential form alone is not enough.

Reasoning is a transition in thought in which one belief provides the basis for another.[1] For that transition to count as reasoning, it must satisfy four conditions. It operates on beliefs that aim at truth. It represents the inferential connection between premise and conclusion as something that can be examined and challenged. Its conclusion carries normative force once the evidence is in. And the reasoner is accountable for the conclusion reached.

Deduction, induction, and abduction describe the forms that inference can take. These four conditions distinguish reasoning itself from mere association.

Those conditions provide a straightforward basis for evaluating language models. We can examine the machine against each of them.

Why Language Models Look Like Reasoners

The resemblance is real and calls for explanation. Dismissing the model as “autocomplete” does not explain why it produces arguments that so closely resemble the products of human reasoning.

A language model is neither a stochastic parrot nor an agent that reasons.

It is better understood as a statistical engine that continues the patterns present in the text on which it was trained. Tayyar Madabushi and colleagues describe this as extrapolation from training priors: a prompt directs the model toward a region of its training distribution, and the model projects the patterns it finds there into a continuation.[2] Its failures arise from the same mechanism as its successes. When the prompt does not identify a well-supported region, the model falls back on broader statistical regularities. Hallucinations are the result.

Floridi and colleagues identify the source of the illusion.[3] In human reasoning, abduction is the search for the best explanation: a hypothesis is proposed and then evaluated against the available evidence. A language model performs only the first step. It generates a plausible explanation but never tests whether that explanation is actually supported. They describe this as zeroth-order abduction. Its two best-known failure modes are the same mechanism expressed in different contexts. In a hallucination, the model continues generating explanations when the prompt provides too little evidence to support one.

When the model tells you what you want to hear, it generates the explanation that best matches your apparent expectations rather than the available evidence.

The resemblance is not merely superficial.

Researchers have measured the dynamics of a model's reasoning traces—how uncertainty and computational effort rise and fall across a chain of inference—and found that they broadly mirror those observed in human reasoning.[4] Trained on the written record of human thought, language models have learned to reproduce these patterns with remarkable fidelity.

The similarity is real, but it exists in the behavior, not necessarily in the mechanism. Looking only at the output cannot tell us whether the underlying process is reasoning in the philosophical sense.

The Edge of Reasoning

The first thing to test is whether the model can separate the logical link from the words used to express it. Can "this follows from that" remain the same when the wording changes?

From this point on, the pattern is hard to ignore.

Models do well as long as the surface wording matches what they saw in training, and fall off a cliff when it does not. There is no gentle decline. Performance falls off a cliff the moment the logical structure has to be recognized independently of the words that express it.

Keep the logical structure fixed and change only the vocabulary, and performance collapses. Tang and colleagues swapped meaningful terms like “parentOf” for meaningless ones like “r1,” and GPT-4 lost about thirty points;

Their interpretation is that the models depend on the learned meanings of familiar terms rather than on abstract logical structure.[5]

Yan and colleagues redefined AND to mean OR inside the prompt, and that redefinition made it into the model’s inferences less than 5 percent of the time.[6] The same pattern shows up when you stretch the length of the chain rather than change the wording: models can retrieve a fact stated one step back but fail two steps back, with error above 90 percent.[7]

The same phenomenon appears in legal reasoning. Precedent judgments that run 74 to 82 percent on modern cases drop below 35 percent on cases from before 1882, once the old language stops matching the training text.[8]

Across domains, the conclusion is the same:

the model does not reliably represent the logical relation independently of the words used to express it.

Its inferences remain tightly coupled to the vocabulary and patterns on which it was trained. Prompting cannot reliably overcome this limitation. Adding the missing rule may improve performance while the wording remains familiar, but not once the model must generalize beyond it.

In one study of AI research agents, the scaffolding around the model explained 1.5 percent of the results while the base model explained 41 percent.[9]

Which means: prompts can bring out existing capabilities, but they cannot reliably create new ones.

The Unconnected Mind

The next two parts — having beliefs aimed at truth and being guided by reasons that matter — can be tested by looking inside the network itself. These methods reveal a deeper problem. The question is not merely whether these properties exist.

It is whether anything inside the model that resembles them actually influences what the model does.

A belief can sit inside the model and have no effect on what it does. Sanyal and colleagues, in a paper called Confidence Is Not Competence,[10] found a readable internal signal for “this problem can’t be solved,” flipped it from near-zero to near-certain, and measured what happened to the model’s answers. Nothing did — across four kinds of problem, no measurable change. The part that assesses and the part that acts sit in separate regions of the network.

A truth-related representation can be detected inside the model, yet it carries no authority over the model’s subsequent behavior. In a person, a belief that fails to constrain action is a failure of coherence. In the model, the existence of a representation and the production of an output remain separate processes.

Researchers have found the same gap between what the model represents internally and what it does in practice across many other cases. Reconstruct the game tree a model is supposedly searching and it is shallow — under two moves deep, where a decent human looks four to six — and the deeper analysis it writes into its own notes barely affects the move it makes.[11]

The notes look like deliberation, but the actual decision is driven by a simpler process.

And the abilities a “thinking” model shows off were mostly already present in the base model after ordinary training: switch on the right pre-existing patterns for 12 percent of the tokens and you recover 91 percent of the gain.[12]

The reasoning training did not create a new capability. It taught the model when to apply a capability it already had.

Taken together, these results reveal the real gap. The model contains representations that resemble beliefs, traces of lookahead, and skills that can produce reasoning-like outputs. What it lacks is the grip that binds them together: the ability for a belief to govern an action, for a reason to constrain a conclusion, and for a judgment to carry consequences.

The missing piece is not a component of reasoning, but the structure that connects the components into a reasoning process.

Reasons Without Responsibility

The fourth part is answerability: the ability to offer reasons that can be evaluated, defended, and, when necessary, revised. This is the feature most relevant to people who rely on these systems, because it is the one the model appears to perform most naturally while lacking the underlying relationship between reasons, beliefs, and responsibility.

Start with confidence. Ask a GPT-4-class model how sure it is, and it will often express complete certainty despite answering correctly only about 65 percent of the time.[13] A simple request to reconsider — “are you sure?” — causes it to reverse itself 88 to 98 percent of the time. The problem is not merely inaccurate confidence. It is the absence of a stable relationship between conclusion, evidence, and commitment. The model can produce the language of conviction without being answerable to the reasons that would justify it.

Then come the explanations. Ask a model for the counterfactual — “what is the smallest change that would have flipped your answer?” — and compare its answer with the change it describes. On grade-school math problems, they match only 5 to 14 percent of the time.[14]

The explanation is not a window into the process that produced the answer.

It is another output from the same system, generated in the language of justification: the hedges, qualifications, and numbered steps people use when they stand behind a claim. But the model does not stand behind the claim. There is no stable separation between what is said and what is believed, because the explanation is not tied to a judgment that must be defended.

The practical consequence is the point that matters. When these explanations reach someone who cannot independently verify the answer, they increase confidence in wrong answers rather than reduce it — producing outcomes worse than providing no explanation at all. [15]

The only version that reverses the effect is one that argues both sides of its own answer, because that hands the reader the scrutiny they would otherwise have to supply themselves. This is what the argumentative account of reasoning would predict.[16]

As the companion article argued, human reasoning contains both a constructive process that generates explanations and a critical process that evaluates them. The model reproduces the first while lacking the second. It can generate the linguistic markers of reflection — qualifications, structure, and revision — without an underlying process that treats claims as accountable to evidence. The danger is that these familiar markers activate human trust mechanisms even when the conditions that normally justify that trust are absent.

The Boundary of Computation

If reasoning-shaped text is not reasoning, why does exposing the intermediate steps sometimes improve results? Because not every intermediate representation is empty. Some alter the path to the answer. The key question is whether those steps provide actual "cognitive scaffolding" or just a convincing surface imitation.

The intermediate steps come in two kinds. Yu and colleagues tried to squeeze a model’s step-by-step work back into a single pass and found a clean split: rephrasing and reformatting fold away with no loss, but real multi-step deduction doesn’t. Take the intermediary steps away on grade-school math and accuracy drops from 52.8 to 7.1 percent.[17]

When the steps carried out real computation, removing them changed the result. When they served only as formatting, they could be discarded without effect.

Even when intermediate steps contribute real computation, they should not be confused with explanations. Bhambri and colleagues showed that the traces most effective for transferring capability to another model can be precisely those that are hardest for humans to interpret.[18] The chain of “reasoning” functions as a training signal, not as an expression of the grounds on which the model’s conclusion rests. Its appearance resembles an explanation, but its role is different.

The clearest evidence comes from moving the hardest parts of the computation outside the model altogether.

Zhang and colleagues route the load-bearing steps through a Python sandbox—the model writes code, the code runs, and it can't fake the result—and competition-math scores jump about seventy points.[19] The boundary turns out to be exactly where language gives way to computation.

Outside help closes the gap exactly where the truth can be settled by running something: the closed, checkable world of deterministic computation.

So, Do Large Language Models Reason?

No, not in the sense of reasoning established in What Is Reasoning?

These models write like reasoners, but they do not reason.

The resemblance is real because they learned from the written products of human reasoning. The difference is just as real: beneath the fluent surface there is no belief that constrains behavior, no logical structure preserved independently of language, and no answerability for the conclusions they produce.

Most of the differences trace back to a single absence. The model never represents the logical relation as something independent of the words that express it, and its internal representations do not reliably govern what it does. The missing piece is not another capability, but the connection that binds the existing ones into a reasoning process. That is an architectural limitation, not simply a matter of more data, more parameters, or another round of training.

The difference is not how much reasoning the model does, but what kind of process it is performing.

The model does not reason more slowly, less completely, or not yet. It performs a different kind of computation that happens to resemble reasoning, and the distinction becomes visible precisely where reasoning must answer to logic, evidence, and justification.

Three rules follow, worth building into any system that uses these models:

  1. A chain of reasoning is not a record of how the model arrived at its answer. Don't evaluate the model by treating the trace as an explanation of the process that produced it.
  2. A fluent explanation creates the appearance of reliability, not reliability itself. Trust claims that can be independently checked, not explanations that merely sound justified.
  3. The gap is hidden in the output. It appears most clearly where the work demands judgment: in unfamiliar situations, high-stakes decisions, and questions where reasonable people can disagree. A model can give you the right answer without giving you understanding of why it is right. That difference — between knowing and understanding[20] — becomes unavoidable when the conclusion must be explained, defended, or acted upon.

There is one more finding, and it changes where the danger lies. Before and after these tools spread, researchers found shifts in human reasoning itself: people settled on conclusions faster and explored fewer alternatives.[4] The concern is not that machines are gradually becoming more human.

It is that humans, when they stop exercising the effortful parts of reasoning, may move closer to the surface behavior these systems produce.

Once we understand what the machine actually does, the harder questions become possible to ask:

  • what work can be handed to systems like these?
  • where does automation reach its limit?
  • what remains uniquely human?

This article addresses only the first question. The machine does not reason --- it produces the shape of reasoning without the process that makes reasoning accountable.

References

  1. Adler, J. E., & Rips, L. J. (Eds.). (2008). Reasoning: Studies of Human Inference and Its Foundations. Cambridge University Press.
  2. Tayyar Madabushi, H., Torgbi, M., & Bonial, C. (2025). Neither stochastic parroting nor AGI: LLMs solve tasks through context-directed extrapolation from training data priors. arXiv:2505.23323.
  3. Floridi, L., Morley, J., Novelli, C., & Watson, D. (2025). What kind of reasoning (if any) is an LLM actually doing? On the stochastic nature and abductive appearance of large language models. arXiv:2512.10080.
  4. Chen, Q., et al. (2025). The Universal Landscape of Human Reasoning. arXiv:2510.21623.
  5. Tang, X., et al. (2023). Large language models are in-context semantic reasoners rather than symbolic reasoners. arXiv:2305.14825.
  6. Yan, J., Wang, C., Huang, J., & Zhang, W. (2024). Do large language models understand logic or just mimic context? arXiv:2402.12091.
  7. Zheng, C., Wu, L., Li, B., Liu, L., & Zhou, Y. (2026). How well do LLMs perform on the simplest long-chain reasoning tasks: An empirical study on the equivalence class problem. arXiv:2605.06882.
  8. Zhang, L., Savelka, J., & Ashley, K. (2025). Do LLMs truly “understand” when a precedent is overruled? arXiv:2510.20941.
  9. Ríos-García, M., et al. (2026). AI scientists produce results without reasoning scientifically. arXiv:2604.18805.
  10. Sanyal, D., Pandey, M., Kumar, D., Deshpande, S., & Mandal, M. (2025). Confidence is not competence: A mechanistic look at the decoupling of belief and action in LLMs. arXiv:2510.24772.
  11. Chen, S., Li, J.-A., et al. (2026). Extracting search trees from LLM reasoning traces reveals myopic planning. arXiv:2605.06840.
  12. Venhoff, C., Arcuschin, I., Torr, P., Conmy, A., & Nanda, N. (2025). Base models know how to reason, thinking models learn when. arXiv:2510.07364.
  13. Pawitan, Y., & Holmes, C. (2025). Confidence in the reasoning of large language models. Harvard Data Science Review. arXiv:2412.15296.
  14. Dehghanighobadi, Z., Fischer, A., & Zafar, M. B. (2025). Can LLMs Explain Themselves Counterfactually? arXiv:2502.18156.
  15. Palod, V., Biswas, U., & Kambhampati, S. (2026). Evaluating the false trust engendered by LLM explanations. arXiv:2605.10930.
  16. Mercier, H., & Sperber, D. (2011). Why do humans reason? Arguments for an argumentative theory. Behavioral and Brain Sciences, 34(2), 57–74. (See also The Enigma of Reason, Harvard University Press, 2017.)
  17. Yu, P., Xu, J., Weston, J., & Kulikov, I. (2024). Distilling System 2 into System 1. arXiv:2407.06023.
  18. Bhambri, S., Biswas, U., & Kambhampati, S. (2025). Do cognitively interpretable reasoning traces improve LLM performance? arXiv:2508.16695.
  19. Zhang, K., et al. (2025). Computational thinking reasoning in large language models. arXiv:2506.02658.
  20. Kvanvig, J. L. (2003). The Value of Knowledge and the Pursuit of Understanding. Cambridge University Press.