Researchers released ZGCM-1, a 7-billion-parameter dense foundation model trained from scratch with a focus on data, system, and algorithmic efficiency. The work argues that smaller models should not try to memorize the open web, but instead be optimized for targeted capabilities such as mathematical reasoning and agentic search. It is presented as a fully open release.
A new arXiv preprint shows that the stochastic optimal control and path space formulations of the Schrodinger bridge problem are equivalent for the kinematic equation on compact connected Lie groups. The proof relies on geometric tools such as the horizontal lift. The result connects control-theoretic and measure-theoretic views of this class of optimal transport problems.
A new arXiv paper introduces Func-R1, an approach aimed at strengthening mathematical function reasoning in multimodal large language models. The work targets the challenge of combining visual perception with symbolic logic when solving math problems from images. The abstract frames deliberate mathematical reasoning in visual settings as an indicator of advanced multimodal model capability.
A new arXiv preprint develops a conceptual and operational framework called Positive Topology, built on a basic relation between points or models and observable properties. Two complementary structures arise from that relation, the first of which captures a notion of universal refinement. The work also discusses forcing matrices, positivity, and information as part of the account.
A new arXiv preprint examines the kinds of mistakes large language models make while working through mathematics problems, grouping them into distinct error categories. The authors use that taxonomy of reasoning failures to target improvements in the models' mathematical problem-solving. The work aims to give a clearer picture of where current LLM reasoning breaks down and how to address it.
A revised arXiv paper describes an open-world setting called the Station, where AI agents built on different model families work toward a common mathematical research objective. The agents operate without a central coordinator or predefined workflow, choosing their own actions instead. The work looks at whether such decentralized collaboration can support autonomous mathematical discovery.
A new arXiv preprint introduces ProIQA, a framework that evaluates the quality of automatically generated math items at a fine-grained, process level. The authors note that while automatic item generation is important for personalized education, verifying the pedagogical value of generated questions remains a bottleneck. They argue existing quality-assessment approaches depend on manual review or shallow signals, which do not scale.
The Clay Mathematics Institute has indicated that the Navier-Stokes existence and smoothness problem, one of its seven $1 million Millennium Prize Problems, appears to have been resolved, reportedly with the help of AI. The institute says a formal verification process is now underway and stresses that its review procedure deliberately takes time. No prize determination has been made yet.
A revised arXiv preprint introduces a method that uses automatic verification to track how well language models reason about unsolved mathematical problems. The author notes that although large language models now handle sophisticated math and science reasoning, whether they can contribute genuinely new research remains contested and thinly studied. The work aims to give a measurable way to assess progress on that question.
A revised arXiv preprint introduces TreeThink, a modular library for applying tree search to mathematical reasoning with large language models. The authors argue that existing LLM tree search frameworks focus on natural language reasoning and lack native integration with formal verifiers, a gap TreeThink fills to support systematic exploration of proof spaces in neural theorem proving.
Researchers have trained a reinforcement learning agent to reduce complicated knot diagrams to simpler forms by selecting move proposals and a guiding value heuristic. The system generalizes to arbitrary knots and links and performs well on notoriously difficult unknots, with the paper also probing unknotting number as a benchmark.
A new arXiv paper examines how post-training choices and test-time inference setups influence a model's ability to write natural-language proofs for difficult olympiad problems. Using Nemotron 3 Ultra as a base, the authors produce two specialist checkpoints via supervised fine-tuning and reinforcement learning, and release the training approach publicly.
Canadian mathematician Jacob Tsimerman, a recent Fields Medal winner, has announced the creation of the Mathematical A.I. Safety Institute (MAISI). The new organization aims to use mathematical methods to help guarantee the safety of AI systems. No further details on funding, staff, or research agenda were provided.
A Hacker News submission argues that OpenAI does not employ mathematicians capable of fully understanding the mathematics behind the company's published work. The claim is an opinion raised by a community member rather than a verified statement from OpenAI or a peer-reviewed finding. The thread reflects broader ongoing debate about the rigor of frontier AI research.
OpenAI says a system of 10,000 agents produced a proof addressing one of the Millennium Prize problems, Navier-Stokes. The result was published alongside a Lean repository to support independent verification and concerns a particular case covered by the official problem statement. The validity of the proof remains under scientific review.