Test-Time Training: the model that changes while you use it

By Turing Post.

Before getting into the substance, the newsletter announces a format change: the 'AI 101' section is being renamed '∇ Guide' (nabla, the mathematical symbol representing a gradient) and will move to the end of the week. The idea is to go beyond basic explanations of AI concepts, models and techniques, connecting them to Turing Post's 'Almanac'.

The article opens with an analysis of OpenAI's experiment on the Navier-Stokes problem (one of the Millennium Problems), which serves as a starting point for discussing the inference or test-time stage. A chart compares GPT-6 Astra with OpenAI's new internal next-generation model on a set of open mathematical problems. According to the email, two things happen at once: the new model starts from a much higher level, and additional test-time compute continues to improve the performance of both models. Part of the jump comes from training a stronger model, and another part comes from giving it more time, more attempts and more compute after it receives the problem. In the Navier-Stokes run, some 10,000 agents spent 88 hours exploring different approaches, sharing useful findings with each other and improving the result through context, coordination, tools and test-time compute.

From there, the newsletter argues for a broader shift in the nature of inference: it has stopped being a fixed model producing a single answer and has become an active process that can stretch out over hours and improve its path toward the solution. But, they note, most of that improvement happens in the process surrounding the model. The question they raise is what happens to the evolution of the model itself during inference.

Here they connect to an earlier Turing Post article (from more than a year ago) on test-time compute and test-time scaling, in which they already identified test-time training (TTT) as one of the possible directions for improving a model's reasoning during inference. According to the email, there has not yet been a TTT 'boom', but we are reaching an interesting point at which several researchers are beginning to explore it.

The central distinction the article draws is this: test-time compute consists of using more computation during inference while keeping the model's weights fixed, whereas test-time training means the model keeps adapting or updating itself during the test phase. TTT challenges the traditional idea that learning ends when a model is deployed, blurring the line between training and inference. This is especially useful when data changes rapidly, which is why it would bring major benefits to world models, which are focused precisely on adapting quickly to new states, actions and environments.

The episode is structured into several sections announced in the table of contents: essential definitions; what test-time training is and how it works; studies that mark TTT as an emerging trend (including 'Test-Time Policy Optimization', TTPO, and 'Agentic Test-Time Training', aTTT); why the pairing between world models and TTT matters so much; examples of world models that use TTT (such as 'WAM-TTT: Test-Time Training While Watching Humans'); what the 'Test-time Mixture of World Models' (TMoW) is; limitations of the approach; and a conclusion. The body of the email received develops the first two parts in detail, while the remaining sections are announced in the index without being developed in the available text.

In the definitions section, the email runs through the necessary concepts: 'training time' (the phase in which the model learns from a dataset by updating its parameters) versus 'inference' or 'test time' (the phase in which the already-trained model receives new data and generates predictions, traditionally with fixed parameters). It defines model parameters as the internal weights learned during training, which TTT also updates during inference. It explains the 'training distribution' (the patterns in the training data) and the 'test distribution' (the patterns in the new data the model encounters after deployment), and 'distribution shift', that is, the difference between the two — for example, new lighting, image noise, a different accent or a different sensor. It also defines self-supervised learning (learning without human labels by creating a task out of the data itself, such as reconstructing masked content or predicting an image's rotation), the 'main task' (the prediction the model actually has to make, such as recognizing an object, transcribing speech or predicting a sequence), the 'auxiliary task' (the self-supervised task used to adapt the model at test time, which is not the final task but whose updates should help complete it), the 'auxiliary loss' (the measure of error on that self-supervised task, used as a signal to update the model), and 'test-time adaptation' (the process of adjusting the model to the new test data, of which TTT is one specific form, via parameter updates and self-supervised learning).

On what test-time training itself is, the email cites the 2020 paper by Yu Sun and other researchers at the University of California, Berkeley, titled 'Test-Time Training with Self-Supervision for Generalization under Distribution Shifts', which describes TTT as a broad approach in which a model adapts while it is being used. The central idea is that inference need not be passive: training can continue during inference to improve the model's performance on completely new data it encounters while operating in the real world.

On how TTT works, the email explains that the process begins once the model has already been trained and deployed, and its goal is to let the model adjust to the data in front of it at that moment (for example, an image uploaded by a user, a frame from a robot's camera, or a snippet of recorded speech). The model needs to make a prediction about that new input — its main task — but it does not know the correct answer, so it cannot learn from it as it would from a normal labeled training task. To solve this, the TTT system creates a separate, self-supervised task from that same new data, called the auxiliary task, whose answer can be generated automatically: because the answer is generated from the input itself, the model already knows it and does not need a human-provided label. The text of the email breaks off further down, before developing the remaining sections announced in the index.

As additional content, outside the main article, the email mentions an episode of Turing Post's YouTube channel, 'Attention Span', devoted to Meta's personal agent Muse. According to the text, Meta gave Muse a notable degree of freedom, but deliberately designed one aspect of the system to be hard for the agent itself to control: its own authority. The episode analyzes the 'Muse Secure VM' and the security architecture surrounding the agent.