Intel / article

Test-Time RL Beats Frozen Frontier Models on Discovery Tasks

Source: test-time-training.github.io
Author: Stanford AI Lab
Date: Published Jan 22, 2026

Context

Learning to Discover at Test Time, from researchers at Stanford, NVIDIA, the Astera Institute, UC San Diego and Together AI, tests whether a gets better at one hard problem if it keeps training while it works on it. Earlier systems such as Google DeepMind's AlphaEvolve prompt a frozen model many times and feed promising past attempts back into new prompts, but the model itself never improves. The method in the paper, TTT-Discover, runs reinforcement learning on the problem it is trying to solve. Each attempt is scored automatically, for example by a kernel's runtime or a mathematical bound, and the model's weights are updated on those scores. Because only one great answer matters, the training objective favors the highest-scoring attempts over the average, and the next starting points are picked from the best results while still exploring others.

The authors say they report every problem they attempted, and the results are mixed in useful ways. TTT-Discover set new records on Erdős' minimum overlap problem (an upper bound of 0.380876, against AlphaEvolve's 0.380924) and on one autocorrelation inequality. Its GPU kernel for a triangular matrix multiplication contest ran in 1,161 microseconds on an H100, against 1,371 for the best human entry, and its programs would have placed first in two past AtCoder optimization contests. It also beat the best published score on one single-cell RNA denoising , though the authors warn that benchmark wins do not guarantee biological validity. It did not beat AlphaEvolve on a second autocorrelation inequality, only matched known circle-packing results, and its AMD kernel was not significantly faster than the top human submission. Experts and contest organizers reviewed the main results.

The cleanest evidence that learning matters is the paper's same-model baseline. Sampling gpt-oss-120b 25,600 times with no training, the same budget, produced a best kernel of about 5,352 microseconds, roughly four and a half times slower than TTT-Discover's, and ablations removing either the training objective or the selection rule did worse. The Stanford AI Lab post frames the work as an open model beating closed models like Gemini and GPT-5. That is a comparison with earlier systems, such as AlphaEvolve and AtCoder agents using Gemini, GPT-5 and Claude models, run with different code and budgets. It does not show those models losing under the same training. The paper puts one run at about $500 on Thinking Machines' Tinker training service. The authors say the method needs problems with continuous scores and name sparse or pass-fail problems as the main open challenge.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.

Read the source test-time-training.github.io