A new paper on arXiv (2610.24519) examines why large language models struggle with research-level physics reasoning. The authors identify a failure mode that is not about missing knowledge or poor computation, but about how models represent the physical world itself.
The abstract notes that models often silently alter given quantities and fabricate information while reasoning. This suggests they are not maintaining a consistent model of the problem, leading to errors even when they know the relevant physics.
To address this, the paper proposes an 'axiomatic world modeling' approach. While the abstract does not detail the method, the framing implies a structured way to keep the model's representation of the world consistent with the stated problem. This summary is based solely on the paper's abstract.