The question of whether frontier theoretical research will soon require access to Large Language Models is no longer hypothetical, according to a recent symposium hosted by QuSoft on September 4th. Researchers are actively testing these powerful AI systems against real-world problems, prompting a critical examination of established scientific processes and the potential for a rapidly shifting landscape.
Tobias Osborne, speaking at the event, directly challenged concerns about human obsolescence, stating “I want to prove wrong the people who say that humans will be useless, by trying to prove them true and failing.” QuSoft, the Dutch research centre for quantum software at CWI and the University of Amsterdam, convened the symposium to foster a community-led conversation about the growing impact of these machines on the future of scientific discovery.
LLMs Industrialize Theory: Osborne’s Agentic Systems & Verification
QuSoft hosted a symposium on September 4th specifically addressing the impact of large language models on theoretical research, signaling a focused effort to understand this rapidly evolving intersection of artificial intelligence and scientific inquiry. The central question posed, “will it continue to be possible to do frontier theoretical research without access to frontier LLMs?”, highlights a growing concern that access to these powerful AI systems may become a prerequisite for work.
This raises questions about equitable access and potential barriers to entry for researchers lacking the necessary resources. Osborne’s work focuses on understanding how to effectively apply these tools, not simply accepting or rejecting them outright, and forming informed opinions about their utility.
Addressing the practical challenges of utilizing LLMs, Osborne detailed the need for careful error-correction loops and scaffolding systems to ensure verifiable results, given the non-deterministic nature of these systems and the limitations imposed by context windows that can induce a form of “amnesia”. He shared a workflow detailing methods for managing these complexities, published alongside his presentation.
Beyond the operational aspects, Osborne also shared insights into the broader impact of LLMs, raising concerns about “epistemic integrity, formalization, and the influx of sloppy proofs,” a point echoed in discussions surrounding the Leiden Declaration on Artificial Intelligence and Mathematics. QuSoft intends to foster a community-led conversation, bringing together front-line experiences to address the challenges of this rapidly changing landscape and questions of “infrastructure, open models, and sovereignty” in the age of AI-assisted research.
I want to prove wrong the people who say that humans will be useless, by trying to prove them true and failing.
Holmes’ “Leiden Declaration” Addresses AI-Generated Mathematical Proofs
However, these same models demonstrate proficiency in verifying existing results, formalizing problems, and executing localized proof steps, creating a paradox where AI-assisted code and paper review offers utility while a surge of AI-generated papers threatens to overwhelm human researchers. Holmes described this as a potential influx of the mathematical equivalent of unreliable AI output. He then raised fundamental questions regarding value and responsibility, asking what constitutes valuable work given these limitations within AI-aided research.
Holmes urged researchers to avoid burdening colleagues with heavily AI-generated results unless the underlying conceptual insights warrant the effort, emphasizing that mathematics itself remains crucial for developing the theoretical frameworks needed to understand and control these increasingly complex systems. Krystal Guo, speaking at the symposium, further delineated mathematical problem-solving into three phases, generation, verification, and digestion, observing that while machines excel at the first two, the deep conceptual “digestion” remains a uniquely human capacity. “Large language models: The industrialization of theoretical science,” as the symposium framed the issue, presents both opportunity and challenge for the future of mathematical discovery.
Large language models: The industrialization of theoretical science.
Guo’s Three-Phase Model: AI-Assisted “Digestion” of Quantum Information
Krystal Guo delineated mathematical problem-solving as a three-phase process, generation, verification, and “digestion”, with the latter receiving focused attention at the September 4th symposium. While automated systems increasingly manage generation and verification tasks, Guo’s work centers on the extraction of intuitive principles and structural insights from proofs, a process she terms “digestion.” Her research, detailed in a recent study of quantum walks and state transfer on Cayley graphs, demonstrates that combining agentic AI systems with conventional research methods can yield notable results.
Guo illustrated how correlations emerging across multiple research outputs are now more readily accessible, accelerating the pace of discovery in quantum hardware where states must be efficiently transported across graph networks. This enhanced ability to identify patterns builds on QuSoft’s existing strengths in quantum algorithms and complexity, areas of research dating back to 1996 when CWI initiated quantum computing research within its Algorithms and Complexity group.
The symposium also addressed the need for revised academic norms in light of AI’s growing influence. Tobias Osborne urged departments to establish emergency committees to openly rewrite social norms, advocating for an end to taboos surrounding AI use in both education and research.
Jonas Helsen received a €1.5 million European Research Council Starting Grant for research on quantum error correction, demonstrating continued investment in the field, while Freek Witteveen was selected for QDNL’s Top Talent Initiative. “Perhaps the advent of LLMs will mean that the standards of research ambitions can and should rise,” Osborne stated, suggesting a potential for increased research success and faster problem-solving.
Dominik Links LLMs to Scalable Astrophysical Data Analysis
Carsten Dominik detailed a critical need for automated model generation in astrophysics, driven by exponentially scaling experimental and observational data pipelines. He presented how classical analytical techniques are increasingly unable to keep pace with incoming data streams, necessitating a shift towards workflows integrating automation, as shared during the September 4th symposium. Dominik’s work focuses on applying LLMs not to replace theoretical physicists, but to augment their capacity for data digestion and hypothesis formation within the vastness of astronomical datasets.
The centre’s research is organised into five lines, one of which is quantum algorithms and complexity, providing a foundation for developing the necessary infrastructure to manage and verify LLM-generated models, ensuring the reliability of results. QuSoft’s history, tracing back to 1996 when CWI initiated quantum computing research in its Algorithms and Complexity group, demonstrates a long-term commitment to tackling computationally intensive problems.
QuSoft’s recent hiring of Stefano Polla as Assistant Professor on September 1, 2026, to establish a research group developing quantum algorithms for chemistry, further demonstrates its expansion into areas that will benefit from these advancements.
Source: https://qusoft.org/2026/09/10/frontier-llms-promise-the-industrialization-of-theoretical-research/
