The impact of Large Language Models (LLMs) on theoretical research is substantial and goes far beyond hype. Across research fields, frontier models yield concrete results, but at the same time raise questions regarding: the conventional scientific process (e.g. verification and peer-review); early career scientists; funding; and more. To address these challenges, QuSoft hosted an extended special symposium with various speakers on September 4th.
LLMs and theory’s rapidly moving goalpost
The symposium addressed the overarching question: “will it continue to be possible to do frontier theoretical research without access to frontier LLMs?” It seems nescessary to adjust scientific methodologies, but also the academic community needs to address the moving goalpost of scientific progress, which goes faster-than-humane. Researchers across mathematics and physics actively test the newest agentic AI systems against real problems.
At QuSoft, we want a community-lead conversation, to bring front-line experiences together and in focus. Below you can read the summary of each presentation—and when available the slides (PDF download) and video recording (YouTube link).
Tobias Osborne: “Large language models: The industrialization of theoretical science”
Tobias Osborne directly challenged the feelings of hopelessness that seem to spread lately in the community “I want to prove wrong the people who say that humans will be useless, by trying to prove them true and failing”.
He unveiled the agentic power of LLMs in theory research, stressing that researchers should try learning to apply these models at the state-of-the-art, frontier, level in order to form an informed opinion on whether they wish to use them or not and why. In the meantime, the community has to prioritize understanding how to deal with the growing impact of these machines. Since these systems are non-deterministic and limited by context windows that induce “amnesia”, managing them requires careful error-correction loops and a scaffolding system to obtain verifiable results. More on the methods shared can be found in Tobias’ workflow, published here.
Besides the powerful, practical indications on how to approach the matter operationally, Osborne shared his insights on the impact of LLMs. Showing the sudden breakdown of correlation between AI-assisted homework and unassisted exam performance, he called for an end to academic taboos around AI use. Across education and research, he urged departments to form dedicated emergency committees to rewrite social norms openly. Perhaps the advent of LLMs will mean that the standards of research ambitions can and should rise, this can lead to more successful research and faster solution of important problems.
Download his presentation here.
David Holmes: “AI in mathematics research: what do we value?”
David’s talk started with the results which appeared during summer of 2026, which represent an historical turning point of machine capabilities in mathematical tasks, which lead to the “Leiden Declaration on Artificial Intelligence and Mathematics”.
Current models struggle to keep “in mind” the big-picture and do not offer genuinely new interesting mathematical definitions, moreover they do not actually learn: one has to wait for the next model to see an appreciable improvement. LLMs do excel at verifying established results or working with old problems, formalizing problems and executing localized proof steps. This creates a paradox: while AI checking human-written code and papers offers promising usefulness, the enormous influx of AI-generated papers threatens to overwhelm actual humans. In other words, they have a lot of potential generating sloppy proofs/results (which is the mathematical version of AI slop). Holmes raised fundamental questions about value and responsibility: asking what is valuable in view of all the limitations, regarding AI-aided work? Moreover he urged that researchers should not force others to read heavily AI-generated results, unless the conceptual insights justify the attention. Ultimately, mathematics retains a vital role in developing the theoretical frameworks required to understand and control these systems in the first place.
Download his presentation here.
Krystal Guo: “Digesting quantum information”
Krystal Guo broke down mathematical problem-solving into three distinct phases: generation, verification, and digestion. While machines are increasingly capable of handling parts of generation and verification, Guo focused on the third: the deep conceptual “digestion”. Here researchers extract intuitive principles and structural insights from a proof. She illustrated her take on the matter through in-depth insights in a recent work on quantum walks and state transfer on Cayley graphs (a critical method in quantum hardware where states must be transported across graph networks), Guo demonstrated that combining the use of agentic AI systems with standard research “digestion” can lead to interesting results. Correlations which emerge across several research outputs are now available way more easily than before. The path forward lies in leveraging automation and combine that with our capacity to synthesize and understand what the results mean, looping between the two to create unprecedented efficiency.
Download her presentation here.
Carsten Dominik: “Industrialization in astronomy”
Carsten Dominik shifted the perspective to large-scale physical science, highlighting how the sheer volume of data in modern astrophysics demands an evolution in theoretical workflows. As experimental and observational pipelines scale exponentially, classical analytical techniques must interface directly with automated model generation to keep pace with incoming data streams. The integration of intelligent scaffolding allows theorists to process, model, and evaluate massive astrophysical datasets at scales that were previously intractable, bridging the gap between raw observational feeds and high-level theoretical interpretation.
Download his presentation here.
Nicolas Renaud (SURF): “Infrastructure, open models, and sovereignty”
Representing SURF, the Nicolas Renaud turned to the infrastructural at the base of it all, sharing the advancement around Dutch computing infrastructure in academia, and touching on concepts such as hardware autonomy and open-source models. He addressed the strategic risk we as Europeans face, relying entirely on proprietary US infrastructure. The presentation detailed ongoing national efforts to build an independent European AI stack. For each layer in the stack he informed the audience of the ongoing efforts at Surf. From national supercomputing resources like Snellius and dedicated GPU clusters in Groningen to localized deployments of models like GPT-NL and open tools (such as VS Code, LibreChat, and LangFlow), Nicolas emphasized that maintaining high-level theoretical capabilities in Europe requires sovereign, open-source AI infrastructure that keeps research data and institutional funding within local domain controls.
(no PDF presentation yet)
