In a striking example of "AI-assisted science," researchers have used Large Language Model (LLM) agents to break long-standing records in a classic physics problem. By combining a sophisticated search pipeline with rigorous mathematical certification, they found new, lower-energy configurations for the Thomson problem—a feat that has traditionally required specialized expertise and months of manual algorithmic tuning.
TL;DR
Researchers utilized LLM agents (Claude) to discover and certify new best-known configurations for the Thomson problem at $N = 912, 1382,$ and $2502$. Beyond the physics results, the paper establishes a rigorous "Human-AI" workflow that uses automated agents for rapid iteration while maintaining scientific integrity through independent verification and exact arithmetic.
What Problem Does This Paper Address?
The Thomson problem, first posed in 1904, asks for the minimum-energy arrangement of $N$ unit charges on a sphere. While it sounds simple, it is a notoriously difficult global optimization challenge. As $N$ increases, the number of local energy minima grows exponentially, making it a standard benchmark for testing optimization algorithms.
The current "gold standard" for these configurations is the Cambridge Cluster Database (CCD). Improving upon the CCD entries usually requires deep domain knowledge in computational geometry and physics. The authors sought to determine if LLM agents could not only write the necessary complex search code but also manage the "research loop"—deploying, debugging, and refining strategies—to find better solutions than the world's best-known human-curated results.
Methods
The authors built a search pipeline using a combination of standard and novel techniques:
- Replica-Exchange Basin Hopping: A global optimization strategy that uses multiple "replicas" of the system at different temperatures to jump between energy valleys and escape local traps.
- Geometric Recombination: A "combiner" process that takes two improved configurations, aligns them geometrically, and splices them together to create a "child" configuration that might inherit the best features of both parents.
- High-Precision Polishing: Using gradient-only Barzilai–Borwein iterations to drive configurations to extremely tight local minima ($|g|_\infty < 10^{-8}$).
- LLM Orchestration: The authors used Claude (via a chat interface) to design the experiments and write the code. A specialized agent, Claude Code, was then used to deploy the code on a laptop, monitor runs, and apply patches.
- Integer Certification: To ensure the improvements weren't just floating-point noise, the authors used exact integer arithmetic to provide rigorous upper and lower bounds on the energy of their new configurations.
Datasets
The study primarily used the Cambridge Cluster Database (CCD) as its baseline. Specifically, they targeted the largest $N$ entries, some of which had been set as recently as October 2026. They also used the official verifier from a public AI research challenge (Few Shots) to calibrate their initial code.
Key Insights and Ideas
- The "Polishing Dividend": The authors discovered that some database entries weren't fully converged. Simply "polishing" an existing entry could lower its energy. To claim a true discovery, they established a "polished baseline"—comparing their results not just to the database value, but to the database value after it had been locally minimized to the same precision.
- Speed-ups can be Sabotage: A "local relaxation" speed-up designed by the AI passed its mathematical exactness tests but failed to find any improvements in practice. It turned out that by "freezing" distant charges, the algorithm introduced a bias that prevented it from finding the long-range rearrangements needed for better minima.
- Verification is the Scarcest Resource: The ease of generating code with LLMs makes it tempting to skip checks. The authors argue that the most critical part of AI-assisted discovery is independent verification code that shares no logic with the search code.
Key Results
The team successfully found and certified three new best-known configurations. The energy drops, while small in relative terms ($\sim 10^{-8}$), are significant in the context of this problem and exceed the precision of previous records by orders of magnitude.
For $N=2502$, the search trajectory (shown below) highlights how different techniques—basin hopping, superposition, and splicing—contributed to the final record-breaking result.
Limitations and Future Work
- Success Rate: The pipeline only improved 3 out of the 11 targeted sizes. This suggests that many CCD entries are already very close to optimal or that the search budget was too small for certain $N$ values.
- Ablation Studies: Because the search was conducted in a "live" research setting, the authors didn't perform full ablation studies to see exactly how much the "combiner" or "replica-exchange" contributed individually.
- Human Oversight: The project still required significant human judgment to choose which sizes to target and when to abandon failing strategies. Future work could explore more autonomous "meta-agents" that can handle these high-level decisions.
Summary updated