Consider one number: 0.380876. That is the new record for the Erdős minimal overlap problem — a decades-old open problem posed by Paul Erdős, where the best human result of 0.380927 had stood for a long time. The new record was set in January by a system called TTT-Discover, pushing the upper bound at the fifth decimal place. Advances at the boundary of mathematics often look like this — nearly invisible to the eye — but a record is a record, and this one passed review by the organizers and experts.
More interesting than the number itself is how the system works.
The mainstream paradigm for AI scientific discovery had been "frozen model + search": the model's weights stay fixed, and massive sampling plus evolutionary rewriting hunts for better solutions — DeepMind's FunSearch and AlphaEvolve both take this route. TTT-Discover (a collaboration between Stanford, NVIDIA, UCSD, the Astera Institute, and Together AI; arXiv 2601.16175) takes a different approach: continue training during the exam. Given a problem, the model's parameters undergo continual reinforcement learning on that problem, evolving while it tests. When moving to the next problem, the base model is untouched — each problem evolves from scratch.
The objective function hides a subtle misalignment
Ordinary reinforcement learning optimizes "average performance": averaging 80 points over 100 attempts is good. But scientific discovery is the opposite — only the best single attempt counts. A strategy with a dismal average score is a winner if it ever touches a new record; a strategy that averages 90 but never breaks out is useless for discovery. The paper calls this the fundamental divergence between naive RL and discovery. TTT-Discover's answer is an entropy-regularized objective: it prevents the policy from collapsing onto safe high-reward actions and forces it to keep taking risks — low-scoring solutions retain nontrivial probability of further exploration, because high-risk, high-reward paths are where records hide.
The second component, borrowed from AlphaZero: PUCT-style search reuse
The system stores attempted solutions in a buffer; high-scoring starting points are preferentially pulled out and improved further, while low-scoring ones retain some chance. Each reuse extends that solution's timeline — many records were not conceived from zero but continued from a half-finished attempt.
The record sheet: four rows
Mathematics: Besides the Erdős problem, the bound of an autocorrelation inequality was also improved. One detail from that campaign: the model first used gradient optimization with Adam plus softmax parameterization, pushing the bound down to 1.510, then switched itself to linear programming, reaching 1.504, and finally used heuristic focusing to inch toward the record — in a single test session, it changed its mathematical tool three times.
GPU kernels: On the GPUMode community's TriMul competition (a core operation from AlphaFold 3), it was 2.06x faster than the human leader on A100 and 18% faster on H100, winning outright on B200 and MI300X as well — and no AI had ever previously submitted a valid competition kernel.
Algorithms: Across historical AtCoder competition problems, it scored 567,062 total, beating the human leader by 65 points.
Biology: On denoising for single-cell analysis, it scored 0.71 versus the human best of 0.64.
The base model matters
All of this ran on gpt-oss-120b — an open-weight model, not a closed-source flagship. Training on each problem cost only a few hundred dollars, executed via Thinking Machines' Tinker API, with code public and results reproducible. As a contrast, prior best results in these fields came almost entirely from closed-source frontier models. A university-lab budget is enough to hire such a tireless collaborator that bets on records across five subjects.
What this means for AI research automation
TTT-Discover is a prototype of a new species: the boundary between training and inference loses meaning, as the same model is both the object being searched and the object being trained during problem-solving. It does not yet accumulate across tasks — what it learns on Erdős does not carry to the next problem. But the direction is clear: evolving from "searching the space of model outputs" to "searching the space of model parameters," the closest step yet to weight-level self-improvement. The four battlefields share one trait: cheap verification. Math has comparable bounds, kernels have stopwatches, contests have judges, denoising has ground truth. Wherever verification is cheap enough, machines are already climbing on their own.
0.380876. The fifth decimal place, a decades-old open problem, a few hundred dollars, one open-weight model. The next line on the record board is still blank.
---
Source note: Paper arXiv 2601.16175, "Learning to Discover at Test Time" (v2, 2026-02-05; first released 2026-01-22), abstract and full text verified; project page test-time-training.github.io; authorship Stanford/NVIDIA/Astera Institute/UC San Diego/Together AI. All figures (0.380876/0.380927, 2.06x/18%, 567,062 vs 566,997, 0.71 vs 0.64, a few hundred dollars per problem, gpt-oss-120b, Tinker) come directly from the paper and project page; the entropy objective and PUCT reuse descriptions follow Section 3.2 of the paper; the "three tool switches" comes from the paper's Erdős campaign section.