LLM-assisted search finds lower-energy Thomson configurations for N = 912, 1382, 2502
Researchers report new best-known arrangements of charges on a sphere, verified with exact arithmetic, using a search pipeline built and run with large language model agents.
In plain words
The Thomson problem asks how to place equal electric charges on a sphere so their mutual repulsion energy is as small as possible. The authors found better arrangements than the current records for three sizes, and a computer algebra check confirms the improvements. The search code was written and operated with help from AI agents, though the results themselves are about the physics, not the AI.
What was announced
The team reports configurations for 912, 1382 and 2502 charges whose energies fall below the values held in the Cambridge Cluster Database, which tracks the best known solutions and had not updated its largest entries since October 2026. The margins are 0.0036186, 0.0109730 and 0.0614033 respectively, and each result was certified using exact integer arithmetic. The authors say all three are strict local minima up to rotation and retain the same defect patterns as the database entries; the 912-charge record had previously been matched by two independent methods.
The search method starts from the database configurations and uses replica-exchange basin hopping with a shared pool of minima, geometric relabelling before recombining improved regions, and a final gradient-only polish. Basin hopping alone produced the two smaller improvements. The pipeline was built and operated with Claude agents, and the authors publish failed searches too. They also flag methodological lessons: polishing can shift energies past the reporting threshold and create false positives unless candidates are checked against a polished baseline, and a faster relaxation variant passed its accuracy test but found nothing in a controlled trial. They recommend verifying candidates with code independent of the search and pre-setting pass criteria for design tests.
What it does not show
- The evidence is the paper's abstract only; the full article, including methods detail and peer review status, is not covered.
- The reported energy margins and certification come from the authors; no independent confirmation is available in the supplied passages.
- The abstract is truncated mid-sentence, so the complete description of the search pipeline is not visible.
For security and IT teams
- Scope
- The work concerns a mathematical physics optimization record, not a security product; the AI agents were used to build and run the search code.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.