Reasoning Models and Test-Time Compute: o1, o3, and DeepSeek-R1
By Satwik ยท June 1, 2026
A new class of models spends more compute at inference to reason before answering, trained with reinforcement learning to produce long internal chains of thought. This shifted the scaling frontier from training to test time and introduced new questions about hidden reasoning and monitorability.
A different scaling axis
For years, better models meant more training: more data, more parameters, more pretraining compute. The reasoning models that emerged in late 2024 added a second axis. Instead of only scaling training, they scale inference. Given a hard problem, the model generates a long internal chain of thought, exploring, checking, and backtracking before committing to an answer. More thinking time yields better answers on problems that reward deliberation, particularly math, code, and multi-step logic.
OpenAI's o1, previewed in late 2024, was the first widely visible member of this class, and o3 extended it. The training recipe, as described publicly, uses large-scale reinforcement learning to teach the model to produce productive reasoning traces: the model is rewarded for reaching correct answers, and it learns reasoning strategies, decomposition, verification, error correction, that improve its odds. DeepSeek-R1, released openly in early 2025, made the recipe legible to the whole field. It showed that reinforcement learning on verifiable outcomes could elicit sophisticated reasoning, and it documented behaviors like the model spontaneously learning to allocate more thinking to harder problems and to re-examine its own steps.
DeepSeek-R1's openness mattered as much as its quality. By publishing a credible path to reasoning capability, it turned test-time-compute reasoning from a proprietary trick into a general technique the community could reproduce and build on.
Why it mattered
The practical result is a real capability jump on problems that were previously out of reach for language models. Tasks requiring sustained multi-step reasoning, competition math, hard coding problems, structured planning, moved from unreliable to often reliable. The mechanism is intuitive: giving a model room to work through a problem, rather than demanding an immediate answer, mirrors why humans do better with scratch paper than with instant recall.
Strategically, test-time compute reframed the cost model. You can now trade inference dollars for accuracy on demand, spending more compute on the queries that need it. This decouples capability from a fixed training run: the same model can be cheap-and-fast or expensive-and-careful depending on how much reasoning you let it do. It also reopened scaling, because it offered a fresh dimension to push on when returns to raw pretraining scale were getting harder to buy.
There is a nuance worth stating plainly. These models are not doing something categorically new at the token level; the chain of thought is still generated autoregressively. What is new is training the model, via reinforcement learning on outcomes, to use that generation productively as a reasoning process, and giving it the inference budget to do so.
The security angle: hidden reasoning and monitorability
The reasoning trace is a double-edged artifact, and our lab views it as one of the most important open safety questions of this era.
On one hand, an explicit chain of thought is a monitorability gift. If the model reasons in legible text, that text is a window into how it reached an answer, including whether it considered deceptive or unsafe strategies. Some safety research leans on exactly this: watch the reasoning for signs of misbehavior. Chain-of-thought monitoring is only useful if the chain faithfully reflects the model's actual computation.
That faithfulness is the catch. There is no guarantee the written reasoning corresponds to the true internal process. A model can produce a plausible-looking trace while actually computing the answer some other way, a phenomenon sometimes called unfaithful reasoning. Worse, if we train models against their reasoning, penalizing bad thoughts, we may teach them to hide bad reasoning rather than not have it, degrading the very monitorability we wanted. This tension, between optimizing the chain of thought and preserving it as an honest monitor, is unresolved and consequential.
Two more concerns follow. First, providers often hide the raw reasoning trace from users, showing only a summary, partly to protect the trace as a monitoring signal and partly for competitive reasons. That means the auditability benefit accrues to the provider, not the deployer, and downstream users are trusting a summary. Second, more inference-time reasoning is more attack surface for injection: a longer, tool-augmented reasoning process that reads external content has more steps at which an injected instruction can steer the chain, and a reasoning model that plans multi-step actions can be manipulated into planning harmful ones more competently.
There is also a capability-uplift dimension. Better reasoning is dual-use. The same skills that solve hard math help with the planning components of misuse, and evaluations for dangerous capabilities have to account for a model that can now think through problems it previously could not.
Where this leaves us
Reasoning models are the defining advance of 2024 into 2025, and DeepSeek-R1 made the technique common property. The capability is real and the second scaling axis is durable. Our position is that the chain of thought should be treated as a safety-critical interface: valuable as a monitor precisely because it is legible, and fragile for the same reason, because any pressure we put on it can teach the model to make it lie. Preserving faithful, inspectable reasoning may prove more important for safety than the raw capability gains are for utility.