Since the release of OpenAI's o1 in September 2024, and especially following DeepSeek-R1 in January 2025, a technical question has dominated research labs: are reasoning models actually reasoning — deriving novel conclusions from premises — or are they executing sophisticated retrieval of patterns memorized during training?

The answer matters because it determines the real limits of these architectures. If it is genuine reasoning, the models can solve problems they have never seen. If it is advanced memorization, their utility collapses exactly on the problems where it matters most: the novel ones.

What Reasoning Models Are

Models like o1, o3, and DeepSeek-R1 differ from conventional LLMs in one fundamental dimension: they receive a compute budget to "think" before responding. This intermediate chain of thought is visible in DeepSeek-R1 (open weights) and Gemini 2.5, but opaque in OpenAI's models.

The o3 model introduced adaptive compute: simple questions receive short chains; complex problems receive long chains. The practical result was substantial — DeepSeek-R1 improved its accuracy on the AIME benchmark from 15.6% to 71% using extended reasoning. This leap did not come from additional parameters, but from more compute during inference time.

What DeepSeek-R1 Revealed About How Models Learn to Reason

The most scientifically interesting aspect of DeepSeek-R1 was not its performance — it was its training method. Instead of supervising the reasoning process step-by-step (as o1 likely does), DeepSeek used GRPO — a form of reinforcement learning where the model discovers how to reason by trial and error, rewarded only for getting the final answer right.

The model was not instructed on "how to reason." It developed reasoning structures on its own. This is a significant result: sophisticated reasoning behavior can emerge from RL without massive supervised reasoning datasets. The "Thoughtology" research published in the Transactions on Machine Learning Research in 2026 systematically analyzed DeepSeek-R1's reasoning behavior and found that its thoughts follow a consistent structure — similar to human language processing patterns.

The Problem: Reasoning Is Not Universal

The evidence complicating the optimistic narrative: reasoning is not uniformly better. Research from 2025–2026 shows that chain-of-thought can degrade performance when verbalizing the reasoning introduces unnecessary complexity — or when the task simply does not require sequential reasoning.

A study comparing DeepSeek-R1 and o3-mini in evaluating machine translation and summarization showed counterintuitive results: o3-mini improved with extended reasoning in translation, but DeepSeek-R1 generally underperformed its non-reasoning variant on the same tasks — except in summarization consistency evaluation. Reasoning helps on problems that require multi-step inference. It hinders on tasks that do not.

Test-Time Compute: The Paradigm Inversion

The most profound architectural implication of reasoning models is the inversion of the compute paradigm. Since 2017, AI has been dominated by training — the bulk of computational investment goes into creating the model, and inference is cheap. Reasoning models partially inverted this: inference compute grows with the complexity of the problem.

Analysts project that by 2030, inference compute will represent 75% of total AI compute — a complete inversion of the dominant paradigm of the 2020s. This has direct implications for hardware (who benefits from efficient inference instead of massive training) and for architecture (sparse models and MoE gain additional relevance because they reduce the cost per token in inference).

The Open Question

Mechanistic interpretability research — especially Anthropic's 2025–2026 work on circuit tracing — is beginning to illuminate what happens internally during a chain-of-thought. Early results suggest that models form genuine intermediate representations during reasoning, not just retrieving patterns. But "genuine" and "useful for novel problems" are not synonymous.

The reasoning vs. memorization question remains open — and it is one of the most important in frontier research in 2026. Because the answer determines whether the next models will be more capable because they are smarter, or just because they have memorized more.