WorksComputer Science
Certified Improvements to Best-Known Thomson-Problem Configurations with a Search Pipeline Built with Large Language Model Agents
Abstract
The Thomson problem asks for N unit charges on a sphere with minimum Coulomb energy; for large N the best configurations known are curated in the Cambridge Cluster Database, whose largest-N entries were set in October 2026. We report configurations for N = 912, 1382 and 2502 with energies 0.0036186, 0.0109730 and 0.0614033 below the database values, certified by exact integer arithmetic. All three are numerically strict local minima modulo rotations and keep the defect inventory and scar compositions of the database configurations; the database value for N = 912 had been reached by two different methods. Our search runs unconstrained replica-exchange basin hopping from the database configurations, with a shared pool of minima, recombination of improved regions after relabelling by geometry, and gradient-only polishing; replica-exchange basin hopping alone found the N = 912 and 1382 configurations, the former in one of five runs. We built and ran it with large language model agents (Claude, through a chat interface and the Claude Code agent), and we report every search, including those that failed. Two findings matter for others who search the database. Polishing alone lowers two entries by more than the threshold at which our pipeline reports a candidate, which produced two spurious records until candidates were compared with the polished database energies (the polished baseline). A local-relaxation speed-up passed its exactness test yet found nothing in a controlled comparison in which every arm with full relaxation went below the database value within 15 minutes. We distil three practices for computational discovery with large language model agents: calibrate the search code and verify candidates with code separate from the search, compare candidates with the polished baseline, and test design changes in controlled comparisons with pass criteria fixed in advance.
Introduction
Place unit charges on the unit sphere so that their Coulomb energy
is as small as possible. This is the Thomson problem Thomson [57]. Its global minimum has been proved only for a handful of small Schwartz, [52]; Cohn & Kumar [18]; the closely related task of computing near-optimal points for the logarithmic energy is Smale’s seventh problem Smale, [54]; Saff & Kuijlaars [51]; Borodachov et al. [11], and the minimisers are models of spherical crystals whose topologically required defects organise into “scars” of dislocations Pérez-Garrido et al., [47]; Bowick et al. [13, 12]; Bausch et al. [9]; Wales & Ulker [65]; Wales et al. [66]. Because the number of local minima appears to grow exponentially with Erber & Hockney, [23]; Amore et al. [6], the problem is also a long-standing benchmark for global optimisation Wales, [62]. For large the best configurations known are curated in the Cambridge Cluster Database (CCD; Wales et al., [67]). Its 53 Coulomb entries span ; the current values for the largest sizes, including and , were set by a search method published in October 2026 (hereafter the October 2026 method; Amore & Wales, [5]), which substantially lowered earlier basin-hopping results Wales et al., [66].
At the same time, artificial intelligence (AI) systems are increasingly used to search for mathematical constructions, from reinforcement learning and neural-guided search Wagner, [61]; Fawzi et al. [24]; Charton et al. [16] to evolutionary search over programs written by large language models (LLMs) Romera-Paredes et al., [48]; Novikov et al. [45]. AlphaEvolve was applied to the Thomson problem itself: for up to about 300, a range searched intensively for decades, it reproduced the best-known energies to high precision but did not improve them Georgiev et al., [26]. LLMs and agent teams have also contributed to machine-checked proofs of optimality for and Tran, [59]; Kryvonos et al. [36]; Tooby-Smith & Zughaid [58]. These systems are engineered pipelines in which the model proposes and an automated evaluator scores. A different and widely used mode is a general-purpose LLM assistant or coding agent Jimenez et al., [33]; Anthropic [7] that writes and runs research code while a human decides what to try next. This mode has already touched the CCD: in October 2026, within a day of the problem being posted in a public AI research challenge, two participants working with the Claude Code agent posted configurations up to 0.045 below the CCD value for Few Shots, (2026); Abreto [1]; Ronnie2025 (2026). Such results are reported as leaderboard entries; how reliable this way of working is, and what it takes to turn its output into a claim others can trust, is rarely documented.
We used this mode of working to search the CCD for improvable entries. Claude, used through a chat interface (hereafter the chat assistant), designed and wrote the search and verification code, and the Claude Code agent running on a laptop (hereafter the agent) deployed, calibrated, ran and patched it; we chose the problem and the sizes, set time boxes and stop rules, judged the results and decided what to release (Table 1). Where it matters who acted, we name the chat assistant or the agent; “the AI systems” means both. Our contributions are as follows.
| Role | Responsibilities |
| Authors | choice of problem and sizes; time boxes and stop rules; decisions to continue, stop, or change direction; requirement that every result pass verification by separate code; accepting or rejecting review comments; decisions to release; communication with the CCD maintainers |
| Chat assistant | literature search and reading; experiment design; the search code of every stage (energy kernel, basin hopping, pool and combiner, scheduler and local relaxation) and the verification and release scripts, delivered as packages with instructions for the agent; reading returned data and proposing decisions; drafting reports |
| Agent | deployment and calibration; launching and monitoring long runs; debugging and patches, including the combiner’s relabelling fix; verification runs; status reports; executing approved releases |
| Reviewing AI system | review of the plan (nine issues); a short independent check run at |
Table 1. Division of labour, as recorded in the project notes.
We give configurations for , 1382 and 2502 with energies 0.0036186, 0.0109730 and 0.0614033 below the CCD values, found by a search pipeline that runs unconstrained replica-exchange basin hopping from the CCD configurations with a shared pool, recombination of improved regions and gradient-only polishing (§ 4 and § 5); replica-exchange basin hopping alone, without pool or recombination, found the configurations for and . The improvements are certified by exact integer arithmetic (§ 3), exceed by a factor of more than the agreement between our floating-point evaluations, and were reproduced by a floating-point recomputation written for this paper (hereafter the paper’s verification code), which shares no code with the search. Numerically, all three configurations are strict local minima modulo rotations with the defect inventory and scar compositions of the CCD configurations, so for these sizes the CCD configurations were not the lowest minima with their defect inventory and scar compositions. The CCD value for predates the October 2026 method and was reproduced by it.
We show that comparison with the CCD values is not enough: polishing alone lowers two of the 53 CCD entries by more than the threshold at which our pipeline reports a candidate (hereafter the candidate threshold), which produced two spurious records before candidates were compared with the lower of the CCD value and the energy of the polished CCD configuration (hereafter the polished baseline; § 6.2).
We show, in a controlled comparison whose pass criterion was fixed in advance (according to the project notes), that a local-relaxation speed-up designed by the chat assistant, exact to in its energy accounting, found no configuration below the CCD value in four core-hours, while every arm that included full relaxation did so within 15 wall-clock minutes; measurements on identical moves show that local relaxation leaves the starting minimum less often and, when it does leave, reports energies 0.1 to 0.4 above those of full relaxation from the same move (§ 6.1).
Every search we ran, including the thirteen sizes that did not improve, is reported with its budget (§ 4 and § 5, Sections B and E), together with the division of labour between us and the AI systems. From these results we draw three practices for AI-assisted computational discovery (§ 7).
Scope of the claims. We claim energies lower than the CCD values as listed on 6 and 10 November 2026 (for , on 10 November); to our knowledge no lower energies for these sizes had been published by 10 November 2026 (the CCD page and files were unchanged, the arXiv version of Amore & Wales [5] had not been revised, and we found no other report). We do not claim global optimality. We do not claim that our search method is better than that of Amore & Wales [5]: our search on used about 12 core-hours before the final configuration appeared (10 to 100 times the per-size timings they report) and about 58 in total, and we ran no equal-budget comparison (§ 7.1). In absolute terms the decreases (0.0036, 0.011 and 0.061 for , 1382 and 2502) are comparable to or larger than the energy differences between the distinct low-lying minima our searches found near the CCD configurations ( to ); in relative terms they are about . All three exceed by many orders of magnitude the agreement between our floating-point evaluations and the width of the certified enclosures (below ; § 3). We directed the project throughout; the search that produced the configuration was proposed and run by Claude Code in a later session, after the project, with our approval (Section E). In neither case did an AI system choose the problem or decide what to report.
Related work
The Thomson problem and spherical crystals. Numerical studies of the Thomson problem go back decades ([23, 3]). For large the twelve disclinations required by Euler’s formula are screened by dislocations (scars) whose length grows with ([47, 46, 13, 12]), as observed in colloidal crystals ([9]). Systematic surveys of low-energy configurations and their defect motifs were carried out by Wales & Ulker [65] and Wales et al. [66]; icosahedral and slightly broken icosahedral constructions are discussed by Altschuler & Pérez-Garrido [4] and Roshal et al. [50]. Amore & Wales [5] introduced oscillating-size hybrid basin hopping, which grows and shrinks configurations one charge at a time across a window of , and used it to set the current CCD values for the largest sizes, including and , reporting a few hundred to a few thousand core-seconds per size for the comparisons with earlier results; they note that improved configurations were still being found as their runs continued.
Global optimisation of clusters. Basin hopping ([38, 64]) performs a Monte Carlo walk on the set of local minima. Replica exchange (parallel tempering) runs several replicas at different temperatures and swaps them ([56, 31, 22]); multi-funnel landscapes such as that of the 38-atom Lennard-Jones cluster (Doye et al. [21]) motivate its use. Genetic algorithms with “cut-and-splice” crossover combine halves of parent clusters (Deaven & Ho [20]) and were applied to the Thomson problem soon after (Morris et al. [42]). Our recombination step is closest to cut-and-splice but splices the region in which a parent differs from a common reference configuration, after the kind of permutational alignment that is routine for atomic clusters ([28] (Wales 2026)).
AI for mathematical and scientific discovery. Neural or LLM-guided search has produced new constructions in combinatorics and algorithms ([61, 19, 24, 48, 45, 26] (Chаrton et al. 2024)); LLMs have been used directly as optimisers [70]; and agentic systems aim to automate larger parts of research [40, 27], with benchmarks that measure how well agents carry out machine-learning engineering, AI research and development, replication of published research, and data-driven science [15, 69, 55, 17]. Work on the Thomson problem itself is summarised in § 1. Open platforms now let AI agents compete on verified problems [10] (Few Shots 2026). The closest prior work is the pair of entries described in § 1 (Section D). Our work differs from these in that it targets mostly larger sizes, characterises the new configurations (local optimality and defect structure) beyond the exact energy check that the challenge’s official verifier also performs, and documents the whole project, including the division of labour between us and the AI systems and how each failure was detected. As at , the CCD values for and were recent; the entries achieved a larger relative decrease (about against –). The emphasis on pass criteria fixed in advance follows practice in empirical science and machine learning [44, 30].
Problem, certification, and local optimality
A configuration has energy (1). The energy is invariant under rotations and reflections of the sphere and under relabelling of the points. We write for the energy listed by the CCD (the CCD value) and for the length scale used throughout ( for ; in the new and the CCD configurations for the three sizes we report the smallest pairwise distance is about ). We report energies relative to the CCD value, , call the decrease when it is positive and the relative decrease, and refer to configurations by (“the configuration”). We write for the Riemannian (tangential) gradient of on , whose th block is , where is the identity matrix and is the Euclidean gradient of with respect to , and for the largest absolute value of its entries.
What counts as an improvement. Floating-point evaluation of (1) for sums terms to a total of about , adjacent double-precision numbers near this value are apart (one unit in the last place, ulp), and a plain sum of so many terms can accumulate much larger rounding errors. The decreases we report exceed both by orders of magnitude, but we make the claim independent of floating point altogether.
Let with integers satisfying , so that exactly, write , and let . Since , every term of equals , so for any integer the bounds
hold term by term. Both sums are evaluated exactly in integer arithmetic, because the floor and the ceiling of the square root of a rational number can be computed exactly with integers; we call (2) the integer certificate.
We map each point, given in floating point, to a rational point exactly on the sphere by rounding its stereographic coordinates (projecting from the pole opposite its largest coordinate) to the grid $2^{-29}\mathbb{Z}^2$ and inverting the projection; this moves points by at most . With each term is bounded to within , so the enclosure (2) is narrower than for $N = 2502. We call an improvement certified if the upper bound for the rationalised new configuration is below ; the rationalised configuration is itself a valid configuration, so no rounding argument is needed.
Local optimality. The claim above does not require local optimality, but we check it to characterise the configurations. With the block of Euclidean second derivatives of with respect to and , the Riemannian Hessian on acts on a tangent vector by [2]
Because is rotation invariant, Hess has a kernel of dimension at least three at any critical point. We call a strict local minimum modulo rotations if it is a critical point and the remaining eigenvalues are positive. Our configurations are critical points to within , so this is a numerical statement; it could be made rigorous by an a-posteriori Newton–Kantorovich argument in interval arithmetic, which our main claim does not need. We assemble the matrix of (3) in orthonormal tangent bases, check it against finite differences of the Riemannian gradient (relative error about ), and compute its full spectrum.
Terminology. A relaxation is a local minimisation of from a perturbed configuration; a full relaxation minimises over all points, and a local relaxation (Section 6.1) only over the points nearest the perturbation. Polishing is a much tighter minimisation, to (Section 4). In a low-energy configuration most points have six neighbours in the Delaunay triangulation (coordination number six); points with five or seven are defects, and Euler’s formula forces the number of 5-fold points to exceed that of 7-fold points by twelve. A bound 5–7 pair is a dislocation; a chain of alternating 5- and 7-fold points with one extra 5-fold point is a scar; the defect inventory of a configuration is its numbers of 5- and 7-fold points. Point groups are in Schoenflies notation (: no symmetry; : one -fold axis; : an -fold axis and two-fold axes perpendicular to it; : the rotations of a regular tetrahedron).
Search pipeline and workflow
Figure 1 summarises the pipeline; full parameters are in Section A. The search ran in three stages, detailed at the end of this section: a two-hour screening of four sizes (stage 1), a 9.6-hour search with a shared pool and recombination, mainly on the two sizes that improved (stage 2), and time slices over 11 sizes (stage 3). Each run at a new size starts from the CCD configuration; later stages restart from the best configurations in the shared pool.

Figure 1. The pipeline that produced the configuration, with verification as finally applied (the comparison with the polished baseline was added after the last search stage, Section 6.2; the Hessian analysis was done for this paper). We chose which sizes to run, for how long, and when to stop; the code was written by the chat assistant and deployed, run and patched by the agent (Table 1).
Exploration: unconstrained replica-exchange basin hopping. The search is unconstrained: no symmetry is imposed and all coordinates move. A run is one single-core search process. Each run keeps three replicas at temperatures 0.003, 0.0173 and 0.1 (energy units; geometrically spaced). A step perturbs each replica with a localised move centred on a random point (the move centre): (i) Gaussian displacements of standard deviation , where denotes a uniform distribution, applied to the nearest neighbours of the move centre; (ii) a rotation by rad of the nearest neighbours about the axis through the move centre, which tends to make dislocations glide; or (iii) from stage 2 on, Gaussian displacements with of neighbours. The perturbed configuration is then relaxed in all coordinates by the limited-memory Broyden–Fletcher–Goldfarb–Shanno
Algorithm 1 Recombination of two improved configurations , relative to the reference configuration (the CCD configuration)
Align and to : iterate an orthogonal Procrustes fit ([34]) with nearest-neighbour correspondences, then relabel points by nearest neighbour, resolving conflicts by a minimum-cost assignment ([37]).
Keep a configuration only if its aligned fraction relative to exceeds 0.9.
, enlarged by every point within of a point in this set; likewise.
Form two children: with the points in replaced by those of , and with the points in replaced by those of ▷ region splices.
if then ▷ changes roughly disjoint
Form a third child, normalise ▷ displacement superposition.
Relax and polish each child; insert it into the pool if it is lower than both parents.
quasi-Newton method (L-BFGS) ([39, 14]) on unnormalised coordinates with , and accepted with the Metropolis rule at the replica’s temperature. After each step, neighbouring replicas are swapped with the usual replica-exchange probability ([31]). A relaxed configuration whose energy falls below the lowest energy in the shared pool (the pool best) by more than the polishing trigger of Table 5 is polished and, if new, inserted into the pool. Energies stored in the pool and compared with the baseline (the CCD energy recomputed from the posted coordinates, later corrected as described in § 6.2) are computed with compensated summation ([35, 43]); the fast plain sums are used only inside the walk.
| Component | Setting |
| Length scale | |
| Replicas, temperatures | 3; 0.003, 0.0173, 0.1 (geometric) |
| Moves (stages 2 and 3) | displacement of neighbours, , probability 0.55; rotation of neighbours by , probability 0.35; displacement of neighbours, , probability 0.10 |
| Moves (stage 1) | first two move types with probabilities 0.65 and 0.35 |
| Relaxation | L-BFGS (SciPy implementation, no bounds), memory 30, gradient tolerance , relative function tolerance , at most 30,000 iterations, on unnormalised coordinates |
| Polishing | step of Section 4 (the first Barzilai–Borwein step), capped at and halved when ; stop at (stage 1 polishing: the lowest of three L-BFGS minimisations to gradient tolerance , two of them after Gaussian perturbations of standard deviation ; released configuration: alternated with the second Barzilai–Borwein step ) |
| Polishing trigger | plain-sum energy below the pool best by (stage 1: below the baseline by ) |
| Pool novelty | energy differs by more than from that of every pool member |
| Reseeding | every 30 min, coldest replica from the pool best; other replicas from the four best pool configurations at start |
| Combiner | every 5 min; the ten best pool configurations below the baseline; aligned fraction ; moved threshold ; enlargement radius |
| Candidate threshold | below the baseline (the CCD energy recomputed from the posted coordinates; after stage 3, the polished baseline) |
| ET (controlled comparison; default afterwards) | stop a relaxation after 20 iterations if no point has moved by from the starting minimum |
| Local relaxation (stage 3) | active set: the points nearest the move centre, where is the number of points moved; L-BFGS, memory 20, at most 600 iterations, gradient tolerance ; hybrid arm re-relaxes fully if the local energy change exceeds |
| Random seeds | stage 1: seeds 0 and 1 per size; stage 2: for runs ; controlled comparison: 1000 and 1001 in every arm; probe of smaller sizes: 0, 1, 10, 11 and 12 per size |
| Energy kernel | C, , single-threaded; compensated summation for reported energies |
| Hardware | Apple M4 laptop (4 performance and 6 efficiency cores); up to nine single-core processes |
Table 5. Search parameters (stage 1 had neither the shared pool nor the combiner, and otherwise used the same settings except where noted).
Polishing. At the energy changes near a minimum fall below the resolution of an energy-based line search, and L-BFGS stalls with around . We therefore polish with a gradient-only Barzilai–Borwein iteration ([8]) on , with , and , where normalise maps each point to , until (stage-specific variants are listed in Table 5).
Shared pool. Runs exchange configurations through a shared pool of polished minima; a configuration enters only if its energy differs from that of every pool member by more than . Every 30 minutes each run restarts its coldest replica from the pool best. A pool configuration more than the candidate threshold () below the baseline is reported as a candidate.
Recombination. Independent runs often improve different parts of the sphere. Combining two such configurations requires a point-to-point correspondence, but during a long walk identical points exchange places (a local rotation move, for example, permutes the points of a patch), so labels stop corresponding even where the geometry has barely changed: the stage-1 configuration, whose changes are confined to three small clusters of 22 points, has an aligned fraction (the share of points within of the same-numbered point of the CCD configuration) of only 0.51. A separate process, the combiner, applies Algorithm 1 every five minutes to pairs of the ten best pool configurations below the baseline; Algorithm 1 relabels each configuration by geometry before combining (an early version compared raw labels; § 6). Each child, which we call a combination, is relaxed, polished, and inserted if it is lower than both parents.
Stages and choice of sizes. The project began with a ten-hour attempt at (the attempt; nine runs, about 80 core-hours), starting from the best configuration then posted in the public challenge mentioned above; it found no improvement (Section D). This, together with the observation that the CCD values for the largest sizes were recent and that Amore & Wales [5] reported modest per-size timings, motivated a change of target: rather than ask how to search faster, we asked where further search was most likely to pay off. Stage 1 was a two-hour screening of four sizes (, 2000, 2502, 3552; two unconstrained runs each, plus one symmetry-constrained run at to try out the fallback) with a decision rule fixed in advance (according to the project notes): concentrate compute on any size that produced a configuration below the CCD value, and switch to symmetry-constrained searches if none did.
Stage 2 ran the pool and the combiner for 9.6 hours, in two segments, on and and searched three further sizes with one run each. Stage 3 cycled through 11 sizes whose CCD configurations have no symmetry () in time slices, and a controlled comparison on then tested the design change made for stage 3 (§ 6.1). The machine was a single ten-core laptop (Table 5) running up to nine single-core processes (runs and the combiner). After the project we ran the stage-1 protocol, five runs per size, on the four CCD entries without symmetry just below (, 912, 920 and 932; the probe of smaller sizes, Section E).
Human–AI workflow. Table 1 lists the division of labour as recorded in the project notes (the plans, status reports, decisions and summaries written during the project). Each stage started with the chat assistant, which designed the plan, wrote the code and delivered it as a package together with a written instruction for the agent. The agent installed the package, calibrated it against known energies, launched and monitored the runs, patched bugs (including, with our approval, the combiner’s relabelling step), ran the verification scripts, wrote the status reports, and published the results. We read the reports, discussed them with the chat assistant and decided the next step. The project took about three and a half days (5–8 November 2026). A further AI system, used separately from the Claude sessions and not identified in the project notes (hereafter the reviewing AI system), reviewed the plan before the attempt and raised nine issues, all of which were addressed; it also ran a short independent check at (Section D).
Results
Three certified improvements
Table 2 gives the results. For the certified upper bound for the new configuration, 3060842.1789773580, is 0.0614033 below the CCD value and 0.0614034 below the certified lower bound for the (rationalised) CCD configuration; the corresponding decreases are 0.0109730 and 0.0109730 for and 0.0036186 and 0.0036187 for , and the relative decreases are 2.01 , 1.18 and 9.03 . We verified the values three ways: the integer certificate (2); the paper’s verification code, which uses exactly rounded summation ([53]); and the release verification script, written by the chat assistant during the project and published with the coordinates, run on freshly downloaded CCD files. All agree to . The CCD page and coordinate files downloaded on 10 November 2026 are byte-identical to those used during the search on 6 November, and the file for is byte-identical to the one from which its search started.
Numerically, all three new configurations are strict local minima modulo rotations: apart from the three rotational modes (eigenvalues of magnitude below ), the smallest Hessian eigenvalue is 1.91 (), 8.17 () and 11.52 (). Consistently, adding Gaussian noise of standard deviation 0.05 to every coordinate and re-polishing returned to the same energy (difference 0 at double precision) for three random seeds. No configuration has a rotation axis of order 2–6, an inversion centre, or a mirror plane.
The improvements do not come from polishing. Polishing the CCD configurations with the same method changes their computed energies by at most one ulp (0 for , for , for ), that is, not at all within rounding. For scale, the October 2026 update of the CCD lowered the previous best values for and by 0.0145 and 1.2968 ([5, 66]) and reproduced the earlier value for ([5 Table 1]).
How the configurations were found
. The configuration was found by the probe of smaller sizes, which ran the stage-1 protocol, without pool or combiner, five times on each of four sizes (Section E). One run reached it after 10,835 relaxations (40 wall-clock minutes); the other four runs at this size found nothing below the CCD value. . The configuration was found directly by basin hopping, 66 minutes into one of the two stage-1 runs, and only polished afterwards. About 19 further core-hours in stage 2 (101,646 relaxations, with the combiner) and 7.5 core-hours in stage 3 did not improve it.
. Figure 2 shows the trajectory. In stage 1 both runs found configurations below the CCD value within 25 wall-clock minutes (the first after 1.2 minutes), reaching −0.0149 and −0.0144 after two hours. Superposing the displacement fields of the two best configurations from one run, −0.0144 and −0.0078, and relaxing gave −0.0196 (the sum of the parents’ would be −0.0222). In stage 2 one run found −0.0511 and then −0.0569, configurations in which 251 and 278 points, respectively, have moved by more than 0.15 from their CCD positions. Label exchanges accumulated during the walk had left the −0.0569 configuration with an aligned fraction of only 0.78, below the combiner’s threshold of 0.9, so the combiner never used it. Once the relabelling step was fixed, the combiner produced the final configuration, −0.0614, within three minutes: it spliced into the −0.0569 configuration the region in which a stage-1 configuration (−0.0096) differs from the CCD configuration, realising about half of that configuration’s decrease. After this, the three runs and the combiner found nothing lower in the remaining 7.6 hours. In total about 12 core-hours were spent on before the final configuration appeared (6 in stage 1, including a two-hour symmetry-constrained run that contributed nothing, and 6 in stage 2); including the later stages and the controlled comparison, about 58 core-hours were spent on this size.

Figure 2. Best found for against wall-clock time since the start of stage 1. Markers show the event that produced each new best. The final configuration is a region splice of the stage-2 configuration at with a stage-1 configuration at ; it appeared less than three minutes after the combiner was restarted with the corrected relabelling step of Algorithm 1 (at 5.2 h).
The first step is reproducible; the last one may not be. Getting below the CCD value for is easy with full relaxation: in each of the two stage-1 runs and in each of the three two-run arms of the controlled comparison that included full relaxation (§ 6.1; the arms shared two random seeds but differed in relaxation mode), a configuration below the CCD value appeared within 25 wall-clock minutes. Reaching −0.0614 is a different matter: it depended on a single stage-2 run, and the other two runs, which differed only in their random seeds, found nothing below −0.0196 on their own.
Two observations qualify the role of recombination. First, decreases do not add: none of the six pairwise displacement superpositions tried in stage 1 produced a within of the sum of the parents’ , and when the localised changes found in the stage-1 configurations (five locations for each size) were transplanted one at a time into the CCD configuration, three of five lowered the energy for (but applying them in sequence retained only one) and none of five did for . Second, recombination was not needed to reach the −0.0196 configuration: basin hopping from the CCD configuration reached a minimum with the same energy (to ) in two arms of the controlled comparison that share a random seed and differ only in early termination (ET) of relaxations returning to the starting minimum (§ 6.1). Of the 147 combinations formed in stage 2 after the fix (114 for , 33 for ), 10 were lower than both parents; one of these was the new best, and the other nine reproduced configurations already in the pool (one of them a second copy of the final configuration). The combiner is therefore best viewed as a cheap way to merge progress made in different regions, not as the engine of the search; isolating its contribution would require the ablation discussed in § 7.1.
What changed structurally
Figure 3 shows where the new configurations for and differ from the CCD configurations. After optimal rotation and relabelling, 279 of the 2502 points move by more than 0.15 and 14 by more than 0.5 (maximum 0.89 ); for , 50 points move by more than 0.15 and 3 by more than (maximum 1.75 ), along an arc between two scars, consistent with dislocation glide. The defect inventory is unchanged in all three cases (Table 2). For , both configurations consist of twelve scars, each a 5–7–5 chain, and 24 isolated dislocations; for , both consist of four eleven-defect scars, seven five-defect scars, one isolated 5-fold point, and two isolated dislocations; for , both consist of nine five-defect and three three-defect scars. The coordination number of 71 points () and 24 points () changes, so for these two sizes the improvements are rearrangements of dislocations within the same defect inventory and the same scar compositions. For all three sizes, the CCD configurations were not the lowest minima with their defect inventory and scar compositions. Whether this holds for other entries is open; eight of the eleven sizes searched with full relaxation from the CCD configuration did not improve (§ 5.4).
| Energy | Certified bound | Decrease () | 5-/7-fold points | |||||
| 912 | CCD | 400657.6165027 | 1.68 | 33 / 21 | ||||
| new | 400657.6128841 | 0.0036186 | 1.91 | 33 / 21 | ||||
| 1382 | CCD | 926576.1030330 | 8.23 | 48 / 36 | ||||
| new | 926576.0920600 | 0.0109730 | 8.17 | 48 / 36 | ||||
| 2502 | CCD | 3060842.2403807 | 11.32 | 48 / 36 | ||||
| new | 3060842.1789774 | 0.0614033 | 11.52 | 48 / 36 |
Table 2. New configurations and the CCD configurations they improve. Energy: CCD value (rows CCD) or double-precision value with exactly rounded summation (rows new). Certified bound: integer bounds (2) for the rationalised configurations (lower bound for the CCD configurations, upper bound for the new ones), rounded outwards at the tenth decimal. and : smallest and largest Hessian eigenvalues after removing the three rotational modes (whose eigenvalues have magnitude below ). All six configurations have point group at tolerance and at the CCD’s tolerance of .

Figure 3. Where the new configurations for and differ from the CCD configurations. Left: Mollweide projection (centred on the displaced region) of the CCD configuration; points displaced by more than in the new configuration are coloured by displacement (in units of ), hollow markers are defects of the CCD configuration and filled markers those of the new configuration. Middle and right: close-ups of the Delaunay triangulation around the points displaced by more than , with every point coloured by displacement. Correspondence between the configurations is obtained by optimal rotation and a minimum-cost assignment.
Complete accounting of the searches
Table 6 in Section B lists every size we searched, with its budget and outcome. Unconstrained basin hopping with full relaxation, started from the CCD configuration, improved two of the seven sizes on which it was run in the project ( and , but not , , , , or ). The five sizes that did not improve received 4 to 10 core-hours of such search each, against about 31 () and 58 () core-hours in total for the two that did, so the comparison between them is not balanced. Four further sizes (, , , ) were searched only with the flawed local-relaxation variant described next and remain untested with full relaxation. The total compute of the project, including combiner processes, was about 290 core-hours; short tests run by the chat assistant in its own sandbox (minutes each) and the reviewing AI system’s check (Section D) are not included. The probe of smaller sizes improved one of its four sizes, (Section E); in all, unconstrained basin hopping with full relaxation from the CCD configuration improved three of the eleven sizes on which it was run.
| Stage | Processes × h | Relaxations | Best | Notes | |
| Preliminary | 960 | 9 × 10 | 1.88 × | – | five strategies; about 80 core-hours used |
| 1 (screening) | 1382 | 2 × 2 | 25,839 | −0.0110 | both runs below the CCD value, the first after 1.5 min |
| 1 (screening) | 2000 | 2 × 2 | 13,164 | – | lowest distinct minimum +0.0030 |
| 1 (screening) | 2502 | 2 × 2 | 8,796 | −0.0149 | both runs below the CCD value, the first after 1.2 min |
| 1 (screening) | 3552 | 2 × 2 | 4,785 | – | lowest distinct minimum +0.0006 |
| 1 (screening) | 2502 | 1 × 2 | – | – | symmetry-constrained (, , , ); best +0.459 |
| 2 (pool) | 1382 | 2 × 9.6 | 101,646 | −0.0110 | no change |
| 2 (pool) | 2502 | 3 × 9.6 | 52,367 | −0.0614 | −0.0511 in the first segment |
| 2 (pool) | 2192 | 1 × 9.6 | 21,128 | – | |
| 2 (pool) | 3182 | 1 × 9.6 | 11,358 | – | |
| 2 (pool) | 4352 | 1 × 9.6 | 6,499 | – | |
| 2 (pool) | all | 1 × 9.6 | – | −0.0614 | combiner; 149 combinations, two of them before the relabelling fix; one new best |
| 3 (time slices) | 11 sizes | 8 × 10 | 203,520 | – | local relaxation for (§ 6.1) |
| 3 (time slices) | all | 1 × 10 | – | – | combiner; 240 combinations; nothing lower |
| Controlled comparison | 2502 | 8 × 2 | – | −0.0196 | four arms, pass criterion fixed in advance (Table 3) |
| Probe of smaller sizes | 900 | – | 116,226 | – | lowest distinct minimum +0.0001 |
| Probe of smaller sizes | 912 | – | 117,852 | −0.0036 | one of five runs (Section E) |
| Probe of smaller sizes | 920 | – | 113,817 | – | lowest distinct minimum +0.0004 |
| Probe of smaller sizes | 932 | – | 109,965 | – | lowest distinct minimum +0.0019 |
Table 6. Every size searched, with budget and outcome.
Failures and how they were detected
A correct speed-up that disabled the search
In the eight-hour second segment of stage 2 the largest size, , managed only 5,181 relaxations, because every move was followed by a relaxation of all points. For stage 3 the chat assistant proposed local relaxation, predicting relaxations tens of times faster for the largest sizes: after a move, relax only the 80–260 points nearest to it and freeze the rest. The design is sound as arithmetic: frozen–frozen interactions are unchanged, so the energy change can be computed exactly, and the implementation passed its exactness test (the accumulated energy over 40 consecutive moves matched a full recomputation to ). We approved its deployment on the strength of this exactness test, without a speed benchmark or a controlled comparison; we count this as a failure of the human–AI workflow, not only of the design. Stage 3 then ran for 79.8 core-hours over 11 sizes (about 72 of them with local relaxation) and found nothing (Table 7).
| 1382 | 2000 | 2192 | 2502 | 3182 | 3552 | 4352 | 5792 | 5872 | 9911 | 9913 | |
| 1382 | 2000 | 2192 | 2502 | 3182 | 3552 | 4352 | 5792 | 5872 | 9911 | 9913 | |
| Core-hours | 7.5 | 7.9 | 7.5 | 7.5 | 6.9 | 7.0 | 7.0 | 8.0 | 7.0 | 7.0 | 6.5 |
| Relaxations () | 42.1 | 31.9 | 26.8 | 24.5 | 17.8 | 16.2 | 13.2 | 11.0 | 9.2 | 5.5 | 5.2 |
| Relaxation | full | local | local | local | local | local | local | local | local | local | local |
Table 7. Stage 3 (time slices of 1.5–2 h cycling over 11 sizes, eight runs, about ten hours).
Because “nothing” is also the expected outcome on sizes that are simply hard, the null result alone was uninformative. We therefore ran a controlled comparison on , where full relaxation had reliably found configurations below the CCD value within minutes, and fixed the pass criterion in advance (according to the project notes, in a file that is not independently timestamped): an arm passes if it finds, within two hours, a configuration at least below the CCD value; passing arms are ranked by best , then by the number of distinct such configurations, then by the time to the first one. The four arms were “full” (full relaxation), “full+ET” (full relaxation with ET: a relaxation is stopped after 20 iterations if no point has moved by ), “hybrid” (local relaxation, followed by full relaxation whenever the local result left the starting minimum, that is, changed the energy by more than ), and “local” (local relaxation); all four used the same two random seeds. Figure 4 and Table 3 show the outcome: the local arm found nothing in four core-hours, while the three arms that included full relaxation passed, full and full+ET both reaching . Local relaxation was switched off, and by the ranking fixed in advance (six distinct configurations below the CCD value against four) full+ET became the default setting for ; no further searches were run. Two runs per arm is a small sample; the contrast rests on its size. The two stage-1 runs at and the three arms that included full relaxation (§ 5) each found a configuration below the

Figure 4. Controlled comparison of relaxation modes (, starting from the CCD configuration; four arms, two runs of two hours each, separate pools, no combiner): best in each arm’s pool against wall-clock minutes. Arms: “full”, full relaxation; “full+ET”, full relaxation with ET; “hybrid”, local relaxation followed by full relaxation when the local result leaves the starting minimum; “local”, local relaxation. Pass criterion, fixed in advance according to the project notes: a configuration at least below the CCD value within two hours.
| Arm | first (wall-clock min) | best after 2 h |
| full | 4.3 | |
| full+ET | 1.5 | |
| hybrid | 13.8 | |
| local | – | 0.0000 |
| Relaxation of the same 183 moves | left starting minimum (%) | time per relaxation (s) |
| local, points | 3.3 | 1.45 |
| local, 800 points | 7.1 | 17.51 |
| full (2502 points) | 7.7 | 2.11 |
Table 3. Outcome of the controlled comparison (left: wall-clock minutes to the first configuration with , earlier of the two runs, a dash meaning none within two hours; best after two hours) and relaxation of the same 183 random moves applied to the CCD configuration (right: fraction of relaxations that left the starting minimum; mean time per relaxation). CCD value after 1.2 to 27.7 core-minutes (wall-clock time to the first such configuration times the number of runs); the local arm found none in 240.
A follow-up measurement points to the mechanism. We applied 183 random moves to the CCD configuration and relaxed each perturbed configuration three ways (Table 3). With active points, only 3.3% of relaxations left the starting minimum (ended at an energy differing from the start by more than ), against 7.7% for full relaxation; the outcomes are paired, and in 8 moves only full relaxation left the starting minimum, never the reverse (exact McNemar test, ). Enlarging the active set to 800 points restored the rate but made relaxation eight times slower than full relaxation; and local relaxation with at most 260 active points was only 1.45 times faster than full relaxation (1.45 s against 2.11 s at ; the predicted speed-up of tens of times was for the largest sizes). This deficit alone is too small to explain the null result: per core-second, local relaxation left the starting minimum at 0.62 times the rate of full relaxation, which, if the time to a first configuration below the CCD value scales inversely with this rate, would stretch even the slowest of the times above to about 45 core-minutes, far below the local arm’s 240. The more likely explanation is a bias in the energies. When the local relaxation did leave its starting minimum, its energy exceeded the energy after full relaxation of the same move by 0.1 to 0.4, which is 5 to 60 times the decreases of 0.007 to 0.02 the search was looking for, and by more than 40 in two relaxations that did not converge. Because both the Metropolis rule and the pool-insertion test used these local energies, they could not see improvements that only appear once the surrounding lattice relaxes; this is consistent with the long-range elastic field of a dislocation, which a frozen boundary cannot accommodate. The hybrid arm, which removes the bias for moves that leave the starting minimum, partly recovered. The general lesson is that the exactness test verified that the code computed what it claimed to compute, and the proxy it optimised (relaxations per hour) improved; neither says anything about whether the search still finds better minima (Mannheim & Garrabrant, 2018).
Polishing dividends and spurious records
During stage 3 the pipeline reported configurations below the CCD values for and . The ease of these results raised suspicion: they appeared without any accepted move. The CCD configurations for these sizes are not converged, and polishing them alone lowers their energies by more than the candidate threshold: they have () and 0.18 (), and polishing lowers their energies by and while moving no point by more than . A pipeline that polishes its own candidates but compares them with an unpolished baseline will report this decrease, the polishing dividend, as a discovery. We changed the baseline to the lower of the CCD value and the energy of the polished CCD configuration (the polished baseline), which removed both spurious records; for , 1382 and 2502, polishing the CCD configuration changes its energy by no more than one ulp. Figure 5 in Section C shows the dividend for every CCD entry.

Figure 5. Convergence of the 53 CCD entries. (a) of the posted coordinates. (b) Energy decrease obtained by polishing each CCD configuration (the iteration of § 4, to ), relative to the energy recomputed from the posted coordinates; grey triangles, drawn at an arbitrary floor, mark entries whose energy changed by at most 2 ulp. Only and change by more than 2 ulp (1.50 and 1.23 ). Dotted line in (a): the polishing target ; dashed line in (b): the candidate threshold . CCD values agree with the recomputed energies to within (they are rounded to seven decimals); the values for every entry are provided as supplementary material.
Other errors
Table 4 lists the other errors that occurred. Of the six, five originated in code, a status report or an interpretation from the AI systems, and one (the symmetry hypothesis) came from us. Three were caught by calibration or by anomalies in logged statistics, two by a deliberate test or by the reviewing AI system, and one by checking a report against the run logs; none was found by reading the code. One of them, the combiner’s reliance on raw point labels, blocked recombination for two hours and had to be fixed before the final configuration could be produced.
| Error | Origin | Detected by | Consequence and fix |
| Aggressive floating-point compiler optimisation removed the compensation terms of the summation routine (absolute energy error at ) | chat assistant | calibration against the official verifier’s scores of three published configurations, before any search | optimisation disabled for that routine only; search kernels unaffected |
| Duplicate detection by rounding energies to , below their noise, in a genetic-algorithm variant at | chat assistant | 94% of offspring identical to the start, none accepted | tolerance band of ; run restarted |
| The stage-2 combiner compared raw point labels, which the walk permutes (aligned fraction ), so the best configurations were never combined; the stage-1 superposition script had relabelled points by geometry | chat assistant | combiner tried one pair in two hours | relabelling by geometry (Algorithm 1) |
| The stage-2 status report first stated 78 candidate pairs and 20 superpositions, ignoring the combiner’s limit of ten configurations | AI systems | checking the report against the run logs | corrected to 54 pairs and 6 superpositions |
| Interpretation that the decreases came from independent local rearrangements | AI systems | transplanting each change separately (Section 5) | claim withdrawn |
| Hypothesis that “lower symmetry is better”, inferred from a sequence of records at | authors | reviewing AI system: and are not nested | claim withdrawn |
Table 4. Errors not covered in Sections 6.1 and 6.2: origin, how each was detected, and its consequence.
Discussion
Why this worked here. For and , we conjecture that the main reason is the choice of sizes, not algorithmic novelty; our data are consistent with this but do not establish it. The CCD values for the largest sizes were recent; their authors report per-size timings of a few hundred to a few thousand core-seconds in their comparisons with earlier results (the compute behind the CCD values is not stated) and reported that improvements were still appearing [5]. A two-hour screening with a decision rule fixed in advance identified two “soft” sizes cheaply, and further compute was concentrated there. The case fits the same pattern: the CCD value there, also from the October 2026 method, fell within a day of being posted as a public challenge, while our ten-hour attempt against the challenge record that replaced it, already lowered by others with the same tools, found nothing. These contrasts differ in more than the compute behind the CCD value, however (size, symmetry, and search code all differ). To remove the differences in symmetry and search code, we ran the stage-1 protocol, with the same code, on the four CCD entries without symmetry just below (, 912, 920 and 932): five runs and about relaxations per size, 4 to 25 times the number of stage-1 relaxations per size at the large sizes (Section E). One size improved: after 10,835 relaxations, one run at found the configuration of § 5; the other 19 runs found nothing below the CCD value, and at each size they found the same lowest distinct minimum (the lowest minimum other than the CCD configuration). According to Amore & Wales (2026, Table 1), the CCD value for predates the October 2026 method, which reproduced it, so an entry that two searches with different methods had reached was still improvable: recency of the CCD value is not necessary for an improvement. At the same time, the probe improved one of four sizes, , in only one of its five runs, against two of four sizes in stage 1, each in both of its runs and with fewer relaxations per run. Counts this small do not show that improvements are easier to find at larger , and the two sets of sizes differ in the origin of their CCD values as well as in , so our data do not separate the effect of size from that of the compute behind each CCD value. The ingredients of the search, basin hopping with replica exchange and crossover-like recombination, are standard [64, 22, 20, 42].
Practices for AI-assisted computational discovery. Each of the three practices below is tied to an error in this project: calibration caught the summation bug before any search (verification by separate code caught no error; the final claim rests on it); comparison with the polished baseline removed two spurious records after their ease had raised suspicion; and a controlled comparison, run only after stage 3 had spent about 72 core-hours on local relaxation, diagnosed that change (run before adoption, it would have saved those hours).
Calibrate the search code, and verify candidates with code separate from the search. Before any search, the energy code had to reproduce the official verifier’s scores of three published configurations, which exposed the summation bug (Table 4); for the large- stages, the search code had to reproduce the CCD values to . Every candidate was then recomputed by the release verification script, which shares no code with the search, and the integer certificate lets anyone check the final claim on the released rational coordinates.
Compare candidates with the polished baseline. Otherwise a polishing dividend masquerades as a discovery (Figure 5).
Test design changes in controlled comparisons with pass criteria fixed in advance, before deploying them, and record the criteria somewhere timestamped; ours were recorded only in the project notes. A change that is correct, fast and plausible can still disable the search. For , only the controlled comparison distinguished “this size is hard” from “this change broke the search”; for the other sizes searched with the flawed variant the question remains open.
In our experience, the AI systems’ speed changed what was cheap: the chat assistant could write a new search variant or verification script, and the agent could deploy it, within a single working session, which made it practical to build controls and checks that we might otherwise have skipped. In this project the scarcest resource seemed to be judgment about which results to trust.
What was specific to working with LLMs. Several failures, and one weakness of the verification, had a distinctly LLM-assisted character: a regression between code packages delivered at different times (relabelling by geometry in stage 1, raw labels in the stage-2 combiner); a test that checked arithmetic exactness rather than whether the search still worked (§ 6.1); an AI-written status report with wrong counts (Table 4); an unverified speed prediction (“tens of times faster” for the largest sizes, measured 1.45 times at ); and verification code written by the same model family as the search (§ 7.1). The reviewing AI system, in turn, caught a reasoning error of ours (Table 4), which suggests, as a hypothesis, that a different AI system should critique the plan and the conclusions. Agent benchmarks score outcomes against fixed graders [15, 69, 55, 17]; open-ended discovery has no grader, and failures of the kinds above are not exposed by an outcome grader.
Open questions. Whether recombination helps at all when decreases are so far from additive, and if so why, is open: one splice supplied the last 7% of the decrease, but we ran no ablation; elastic interactions between rearranged regions are a natural starting point. It is also open whether the many CCD sizes on which we found nothing are close to optimal or were simply under-searched by our budgets.
Limitations
We ran no equal-budget comparison. We did not run the October 2026 method of Amore & Wales [5], or plain basin hopping, with the same budget on the same sizes, so we cannot say whether our pipeline is more efficient than theirs; our budget on was larger than the per-size timings they report.
There were few independent runs. The success at depended on one run that found the and configurations; runs differing only in the random seed did not. The success at came from one of five runs. The controlled comparison used two runs per arm, with the same two seeds in every arm, on a laptop whose performance and efficiency cores run at different speeds.
We ran no ablations. We did not separately remove the pool, the combiner, or replica exchange, so we cannot attribute the decrease to individual components.
The success rate is low and the effect small. Three of the eleven sizes searched with full relaxation from the CCD configuration improved; the relative decreases are of order .
The CCD maintainers have not yet confirmed or posted the configurations; ongoing searches by others may since have found lower energies.
The release verification script was written by the chat assistant, and the paper’s verification code and the certification script with Claude Code. They share no code with the search, and the certificate is short integer arithmetic that can be audited by hand; we release the rational coordinates it uses, but the scripts are not independent of the model.
The evidence on AI-assisted research comes from one project on one problem, and our account of the AI systems’ contribution is qualitative: we did not measure how much time the AI systems saved or what fraction of the chat assistant’s proposals were adopted, and we cannot say what we would have achieved on the same budget without them, or with standard tools (Wales, 2026).
Conclusion
We reported configurations for the Thomson problem with , and whose energies are certifiably below the CCD values, found on a laptop by a search pipeline that the AI systems wrote and ran under our direction; its first stage alone, replica-exchange basin hopping without pool or recombination, found the configurations for (in one of five runs) and . The AI systems made the search fast to build and easy to extend; what made its output a claim others can trust were the three practices of Section 7: calibration and verification by code separate from the search, ending in an integer certificate that anyone can re-run on the released rational coordinates; comparison with the polished baseline; and controlled comparisons with pass criteria fixed in advance (the one we ran showed that a correct speed-up had disabled the search).
Broader Impact Statement
This work concerns a mathematical optimisation benchmark and poses no direct risk of harm. Its indirect relevance is to how AI-assisted results are reported. Fast generation of plausible code and plausible results makes it easier to publish claims that are wrong; the verification practices and complete reporting used here are intended to counter that, and we encourage similar disclosure of failed attempts and of the human–AI division of labour in other AI-assisted work.
Use of AI systems
The search, analysis and visualisation code used during the project was written by Claude (Anthropic) through the claude.ai chat interface and deployed, run and patched by the Claude Code agent, under our direction (Table 1); the reviewing AI system, used separately, reviewed the plan and ran a short independent check. The project notes do not record the model versions or the identity of the reviewing AI system. The paper’s verification code, the other analysis code written for this paper, and a first draft of this manuscript were produced with Claude Opus 5.5 through Claude Code, which also proposed and ran the probe of smaller sizes with the project’s search code (Section E); all of it was checked and revised by the authors, who take full responsibility for the content.
Data and code availability
Coordinates of the and configurations, the release verification script, which shares no code with the search, and the release history are public at https://github.com/jesse87wang-bit/thomson-records and archived on Zenodo (https://doi.org/10.5281/zenodo.23996279; [68]). The search code, run logs, pool contents, controlled-comparison data, the coordinates and certificate of the configuration, and the analysis scripts of this paper are provided as supplementary material (Section F).
Acknowledgments
We thank David J. Wales and colleagues for maintaining the CCD and making the coordinates of every entry public, which made this work possible.
Parameters of the search
Complete accounting of the searches
Table 6: Every size searched, with budget and outcome. “Processes × h” counts single-core processes: runs and, where marked, the combiner. Relaxations: L-BFGS minimisations after a move (in all coordinates except in the local-relaxation runs of stage 3). values are rounded to four decimals (Table 2 gives those of the new configurations to seven decimals). In the Best column, a dash means that no configuration below the baseline was found (for the baseline is the challenge record from which the search started, 0.0449 below the CCD value) or, in stage 3, that no size improved; in the other columns, a dash means that the quantity is not reported. Lowest distinct minimum: the lowest minimum found other than the starting configuration. Preliminary: the attempt (Section D). Stage 2 ran in two segments, of 1.6 and 8 hours; stage 3 is reported per size in Table 7. The probe of smaller sizes was run on other hardware (per size, two runs on a laptop and three on a server); its budget is given in relaxations (five runs per size).
Table 7: Stage 3 (time slices of 1.5–2 h cycling over 11 sizes, eight runs, about ten hours). Local relaxation was active for . No size improved; for and 2502 the starting point was the pool from stage 2.
Polishing dividends of all CCD entries
The preliminary attempt at
The project began as an entry to the Few Shots challenge (Few Shots, 2026), whose problem p007, posted on 27 October 2026, is the Thomson problem for , scored by an official verifier in exact integer arithmetic with the CCD configuration as baseline. The CCD lists a -symmetric configuration with energy 444368.4373480 for this size, set by Amore & Wales [5]Table 3), who lowered the previous value [50] by 0.382. Within a day of the problem’s posting, two participants working with the Claude Code agent submitted lower configurations: one with symmetry (444368.4040805), found with a symmetry-preserving cut-and-splice genetic algorithm [1], and one with symmetry (444368.3924062), found by replica exchange in the -symmetric subspace [49], 0.0449 below the CCD value. These are not our results and we make no claim about them; we report them because they are the closest prior use of the same tools on the CCD, and because they were our starting point.
Starting from the configuration, nine runs used five strategies for up to ten hours: replica-exchange basin hopping constrained to symmetry (the method behind the challenge record, as a control), -preserving cut-and-splice recombination, unconstrained replica exchange, a targeted scan of scar compositions motivated by elasticity theory [9], and adding and removing a point. In about 80 core-hours (1.88 relaxations) none found a lower configuration: relative to the starting configuration, the symmetry-constrained runs stalled within minutes at +0.0190, and the lowest distinct minimum of the unconstrained runs was +0.0328, which short independent tests by the reviewing AI system also reached. We concluded that this record, freshly lowered by other participants, was a poor target for further search with our budget, and moved to the large- entries.
The probe of smaller sizes
The probe ran the stage-1 protocol (stage-1 settings of Table 5) with the project’s search code, unchanged, on the four CCD entries without symmetry just below ; every run started from the CCD configuration.
Table 8 lists the outcomes. At each size, the runs that did not go below the CCD value found the same lowest distinct minimum.
| Origin of CCD value | CCD value | Relaxations | Runs below CCD value | Best | |
| 900 | October 2026 method | 390084.7556201 | 116,226 | 0 | +0.0000591 |
| 912 | earlier, reproduced | 400657.6165027 | 117,852 | 1 | −0.0036186 |
| 920 | earlier, reproduced | 407784.8388324 | 113,817 | 0 | +0.0004066 |
| 932 | earlier, reproduced | 418594.2339500 | 109,965 | 0 | +0.0018764 |
Table 8. The probe of smaller sizes: five runs per size with the stage-1 protocol. Origin of the CCD value according to [5 Tables 1 and 3]: set by the October 2026 method, or set earlier and reproduced by this method. Relaxations: total over the five runs. Best : of the lowest distinct minimum (Table 6) found in the five runs.
The checks of the configuration are reported with those of the other two in § 5.
Reproducibility
Files. The released coordinate files are identical to those analysed here; the leading digits of their hashes under the Secure Hash Algorithm with 256-bit digests are 19a0007473c235d2 () and 8e2ee8176c6f1d46 (). The CCD coordinates were downloaded from https://www-wales.ch.cam.ac.uk/~wales/CCD/Thomson/site/s_1/dat/.
Certificates. The rational coordinates used in the integer certificate (2) are provided for , 1382 and 2502: each point is an integer quadruple with , together with the energy bounds; anyone can re-check the bounds with exact integer arithmetic.
Scripts. The supplementary scripts produce the integer certificates; the energies, gradients, Hessian spectra, defects and displacements; the polishing dividends of all CCD entries; the tests in § 6.1; the checks of the configuration; and Figure 2 to Figure 5. They need Python with NumPy, SciPy and Matplotlib ([29, 60, 32]) and a C compiler.
References
- [1]Abreto. Few Shots p007 entry thomson960-d2-ga: a D₂-symmetric configuration of 960 points, verifier score 444368.4040805305. https://github.com/few-shots/few-shots/pull/15, 2026. Verified 27 October 2026, 20:50 UTC; agent Claude Code. Accessed 10 November 2026.
- [2]P.-A. Absil, R. Mahony, and R. Sepulchre. Optimization Algorithms on Matrix Manifolds. Princeton University Press, Princeton, NJ, 2008.DOI
- [3]E. L. Altschuler, T. J. Williams, E. R. Ratner, R. Tipton, R. Stong, F. Dowla, and F. Wooten. Possible global minimum configurations for Thomson’s problem of charges on a sphere. Physical Review Letters, 78(14):2681–2685, 1997.
- [4]Eric Lewin Altschuler and Antonio Pérez-Garrido. Defect-free global minima in Thomson’s problem of charges on a sphere. Physical Review E, 73(3):036108, 2006.DOI
- [5]Paolo Amore and David J. Wales. Efficient searches for low-energy structures in clusters with thousands of particles: Application to the Thomson problem. The Journal of Chemical Physics, 165(16):164104, 2026. Also arXiv:2610.08122.arxiv.org/abs/2610.08122
- [6]Paolo Amore, Victor Figueroa, Enrique Diaz, Jorge A. López, and Trevor Vincent. Exploring the energy landscape of the Thomson problem: local minima and stationary states. arXiv preprint arXiv:2506.08398, 2025.arxiv.org/abs/2506.08398
- [7]Anthropic. Claude Code. https://www.anthropic.com/claude-code, 2025. Agentic coding tool; accessed November 2026.
- [8]Jonathan Barzilai and Jonathan M. Borwein. Two-point step size gradient methods. IMA Journal of Numerical Analysis, 8(1):141–148, 1988.DOI
- [9]A. R. Bausch, M. J. Bowick, A. Cacciuto, A. D. Dinsmore, M. F. Hsu, D. R. Nelson, M. G. Nikolaides, A. Travesset, and D. A. Weitz. Grain boundary scars and spherical crystallography. Science, 299(5613):1716–1718, 2003.DOI
- [10]Federico Bianchi, Yongchan Kwon, Aneesh Pappu, and James Zou. Harnessing the collective intelligence of AI agents in the wild for new discoveries. arXiv preprint arXiv:2606.10402, 2026.arxiv.org/abs/2606.10402
- [11]Sergiy V. Borodachov, Douglas P. Hardin, and Edward B. Saff. Discrete Energy on Rectifiable Sets. Springer Monographs in Mathematics. Springer, New York, 2019.DOI
- [12]Mark Bowick, Angelo Cacciuto, David R. Nelson, and Alex Travesset. Crystalline order on a sphere and the generalized Thomson problem. Physical Review Letters, 89(18):185502, 2002.DOI
- [13]Mark J. Bowick, David R. Nelson, and Alex Travesset. Interacting topological defects on frozen topographies. Physical Review B, 62(13):8738–8751, 2000.DOI
- [14]Richard H. Byrd, Peihuang Lu, Jorge Nocedal, and Ciyou Zhu. A limited memory algorithm for bound constrained optimization. SIAM Journal on Scientific Computing, 16(5):1190–1208, 1995.
- [15]Jun Shern Chan, Neil Chowdhury, Oliver Jaffe, James Aung, Dane Sherburn, Evan Mays, Giulio Starace, Kevin Liu, Leon Maksin, Tejal Patwardhan, Lilian Weng, and Aleksander Mądry. MLE-bench: Evaluating machine learning agents on machine learning engineering. In The Thirteenth International Conference on Learning Representations (ICLR), 2025.arxiv.org/abs/2410.07095
- [16]François Charton, Jordan S. Ellenberg, Adam Zsolt Wagner, and Geordie Williamson. PatternBoost: Constructions in mathematics with a little help from AI. arXiv preprint arXiv:2411.00566, 2024.arxiv.org/abs/2411.00566
- [17]Ziru Chen, Shijie Chen, Yuting Ning, Qianheng Zhang, Boshi Wang, Botao Yu, Yifei Li, Zeyi Liao, Chen Wei, Zitong Liu, Vishal Dey, Mingyi Xue, Frazier N. Baker, Benjamin Burns, Daniel Adu-Amprah, Xuhui Huang, Xia Ning, Song Gao, Yu Su, and Huan Sun. ScienceAgentBench: Toward rigorous assessment of language agents for data-driven scientific discovery. In The Thirteenth International Conference on Learning Representations (ICLR), 2025.arxiv.org/abs/2410.05080
- [18]Henry Cohn and Abhinav Kumar. Universally optimal distribution of points on spheres. Journal of the American Mathematical Society, 20(1):99–148, 2007.DOI
- [19]Alex Davies, Petar Veličković, Lars Buesing, Sam Blackwell, Daniel Zheng, Nenad Tomašev, Richard Tanburn, Peter Battaglia, Charles Blundell, András Juhász, Marc Lackenby, Geordie Williamson, Demis Hassabis, and Pushmeet Kohli. Advancing mathematics by guiding human intuition with AI. Nature, 600:70–74, 2021.DOI
- [20]D. M. Deaven and K. M. Ho. Molecular geometry optimization with a genetic algorithm. Physical Review Letters, 75(2):288–291, 1995.pubmed.ncbi.nlm.nih.gov/10059656
- [21]Jonathan P. K. Doye, Mark A. Miller, and David J. Wales. The double-funnel energy landscape of the 38-atom Lennard-Jones cluster. The Journal of Chemical Physics, 110(14):6896–6906, 1999.DOI
- [22]David J. Earl and Michael W. Deem. Parallel tempering: Theory, applications, and new perspectives. Physical Chemistry Chemical Physics, 7(23):3910–3916, 2005.DOI
- [23]T. Erber and G. M. Hockney. Equilibrium configurations of N equal charges on a sphere. Journal of Physics A: Mathematical and General, 24(23):L1369–L1377, 1991.DOI
- [24]Alhussein Fawzi, Matej Balog, Aja Huang, Thomas Hubert, Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Francisco J. R. Ruiz, Julian Schrittwieser, Grzegorz Swirszcz, David Silver, Demis Hassabis, and Pushmeet Kohli. Discovering faster matrix multiplication algorithms with reinforcement learning. Nature, 610:47–53, 2022.DOI
- [26]Bogdan Georgiev, Javier Gómez-Serrano, Terence Tao, and Adam Zsolt Wagner. Mathematical exploration and discovery at scale. arXiv preprint arXiv:2511.02864, 2025.arxiv.org/abs/2511.02864
- [27]Juraj Gottweis, Wei-Hung Weng, Alexander Darvin, Tao Tu, Petar Sirkovic, et al. Accelerating scientific discovery with Co-Scientist. arXiv preprint arXiv:2502.18864, 2025.DOI
- [28]Matthew Griffiths, Samuel P. Niblett, and David J. Wales. Optimal alignment of structures for finite and periodic systems. Journal of Chemical Theory and Computation, 13(10):4914–4931, 2017.DOI
- [29]Charles R. Harris, K. Jarrod Millman, Stéfan J. van der Walt, et al. Array programming with NumPy. Nature, 585:357–362, 2020.DOI
- [30]Peter Henderson, Riashat Islam, Philip Bachman, Joelle Pineau, Doina Precup, and David Meger. Deep reinforcement learning that matters. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018.arxiv.org/abs/1709.06560
- [31]Koji Hukushima and Koji Nemoto. Exchange Monte Carlo method and application to spin glass simulations. Journal of the Physical Society of Japan, 65(6):1604–1608, 1996.DOI
- [32]John D. Hunter. Matplotlib: A 2D graphics environment. Computing in Science & Engineering, 9(3):90–95, 2007.DOI
- [33]Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. SWE-bench: Can language models resolve real-world GitHub issues? In International Conference on Learning Representations (ICLR), 2024.arxiv.org/abs/2310.06770
- [34]W. Kabsch. A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A, 32(5):922–923, 1976.
- [35]William Kahan. Further remarks on reducing truncation errors. Communications of the ACM, 8(1):40, 1965.DOI
- [36]Liudmyla Kryvonos, Lukas Liehr, and Mitchell A. Taylor. Energy minimization for eight points on the sphere. arXiv preprint arXiv:2609.22077, 2026.arxiv.org/abs/2609.22077
- [37]H. W. Kuhn. The Hungarian method for the assignment problem. Naval Research Logistics Quarterly, 2 (1–2):83–97, 1955.
- [38]Zhenqin Li and Harold A. Scheraga. Monte Carlo-minimization approach to the multiple-minima problem in protein folding. Proceedings of the National Academy of Sciences, 84(19):6611–6615, 1987.pubmed.ncbi.nlm.nih.gov/3477791
- [39]Dong C. Liu and Jorge Nocedal. On the limited memory BFGS method for large scale optimization. Mathematical Programming, 45(1–3):503–528, 1989.
- [40]Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The AI scientist: Towards fully automated open-ended scientific discovery. arXiv preprint arXiv:2408.06292, 2024.arxiv.org/abs/2408.06292
- [42]J. R. Morris, D. M. Deaven, and K. M. Ho. Genetic-algorithm energy minimization for point charges on a sphere. Physical Review B, 53(4):R1740–R1743, 1996.pubmed.ncbi.nlm.nih.gov/9983695
- [43]Arnold Neumaier. Rundungsfehleranalyse einiger Verfahren zur Summation endlicher Summen. ZAMM – Zeitschrift für Angewandte Mathematik und Mechanik, 54(1):39–51, 1974.DOI
- [44]Brian A. Nosek, Charles R. Ebersole, Alexander C. DeHaven, and David T. Mellor. The preregistration revolution. Proceedings of the National Academy of Sciences, 115(11):2600–2606, 2018.DOI
- [45]Alexander Novikov, Ngân Vũ, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog. AlphaEvolve: A coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131, 2025.arxiv.org/abs/2506.13131
- [46]A. Pérez-Garrido and M. A. Moore. Symmetric patterns of dislocations in Thomson’s problem. Physical Review B, 60(23):15628–15631, 1999.DOI
- [47]A. Pérez-Garrido, M. J. W. Dodgson, and M. A. Moore. Influence of dislocations in Thomson’s problem. Physical Review B, 56(7):3640–3643, 1997.
- [48]Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. Mathematical discoveries from program search with large language models. Nature, 625:468–475, 2024.DOI
- [49]Ronnie2025. Few Shots p007 entry c3-replica-exchange-scars: a C₃-symmetric configuration of 960 points, verifier score 444368.3924069173. https://github.com/few-shots/few-shots/pull/17, 2026. Verified 28 October 2026, 00:35 UTC; agent Claude Code, with several declared models. Accessed 10 November 2026.
- [50]D. S. Roshal, A. E. Myasnikova, and S. B. Rochal. Slightly broken icosahedral symmetry advances Thomson problem. Physics Letters A, 379:372–376, 2015.arxiv.org/abs/1408.3473
- [51]Edward B. Saff and Arno B. J. Kuijlaars. Distributing many points on a sphere. The Mathematical Intelligencer, 19(1):5–11, 1997.DOI
- [52]Richard Evan Schwartz. The five-electron case of Thomson’s problem. Experimental Mathematics, 22(2):157–186, 2013.DOI
- [53]Jonathan Richard Shewchuk. Adaptive precision floating-point arithmetic and fast robust geometric predicates. Discrete & Computational Geometry, 18(3):305–363, 1997.DOI
- [54]Steve Smale. Mathematical problems for the next century. The Mathematical Intelligencer, 20(2):7–15, 1998.
- [55]Giulio Starace, Oliver Jaffe, Dane Sherburn, James Aung, Jun Shern Chan, Leon Maksin, Rachel Dias, Evan Mays, Benjamin Kinsella, Wyatt Thompson, Johannes Heidecke, Amelia Glaese, and Tejal Patwardhan. PaperBench: Evaluating AI’s ability to replicate AI research. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pp. 56843–56873. PMLR, 2025.arxiv.org/abs/2504.01848
- [56]Robert H. Swendsen and Jian-Sheng Wang. Replica Monte Carlo simulation of spin-glasses. Physical Review Letters, 57(21):2607–2609, 1986.pubmed.ncbi.nlm.nih.gov/10033814
- [57]J. J. Thomson. On the structure of the atom: an investigation of the stability and periods of oscillation of a number of corpuscles arranged at equal intervals around the circumference of a circle; with application of the results to the theory of atomic structure. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 7(39):237–265, 1904.DOI
- [58]Joseph Tooby-Smith and Alex Zughaid. A warrant for the N = 8 Thomson problem, in Lean 4. https://github.com/jstoobysmith/Thomson-N-8-Warrant, 2026. Accessed 10 November 2026.
- [59]Hung Tran. A Lean proof of the Thomson problem for seven electrons. Vals AI blog, https://www.vals.ai/blogs/thomson-n7-lean-proof, 2026. Posted 2026-09-28.
- [60]Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nature Methods, 17:261–272, 2020.DOI
- [61]Adam Zsolt Wagner. Constructions in combinatorics via neural networks. arXiv preprint arXiv:2104.14516, 2021.arxiv.org/abs/2104.14516
- [62]David J. Wales. Energy Landscapes: Applications to Clusters, Biomolecules and Glasses. Cambridge University Press, Cambridge, 2003.
- [64]David J. Wales and Jonathan P. K. Doye. Global optimization by basin-hopping and the lowest energy structures of Lennard-Jones clusters containing up to 110 atoms. The Journal of Physical Chemistry A, 101(28):5111–5116, 1997.DOI
- [65]David J. Wales and Sidika Ulker. Structure and dynamics of spherical crystals characterized for the Thomson problem. Physical Review B, 74(21):212101, 2006.DOI
- [66]David J. Wales, Hayley McKay, and Eric L. Altschuler. Defect motifs for spherical topologies. Physical Review B, 79(22):224115, 2009.DOI
- [67]David J. Wales et al. The Cambridge Cluster Database: Riesz energy database, s = 1 (Thomson problem). https://www-wales.ch.cam.ac.uk/~wales/CCD/Thomson/site/s_1/index.html, 2026. Accessed 2026-11-06 and 2026-11-10 (identical page content).
- [68]Tao Wang. Lower-energy configurations for the Thomson problem: N = 2502 and N = 1382. Zenodo dataset (concept DOI 10.5281/zenodo.23996279), version 1.1.0; DOI: 10.5281/zenodo.24014200; code and coordinates at https://github.com/jesse87wang-bit/thomson-records, 2026.
- [69]Hjalmar Wijk, Tao Lin, Joel Becker, Sami Jawhar, Neev Parikh, Thomas Broadley, Lawrence Chan, Michael Chen, Joshua Clymer, Jai Dhyani, Elena Erlicheva, Katharyn Garcia, Brian Goodrich, Nikola Jurkovic, Megan Kinniment, Aron Lajko, Seraphina Nix, Lucas Sato, William Saunders, Maksym Taran, Ben West, and Elizabeth Barnes. RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pp. 66772–66832. PMLR, 2025.arxiv.org/abs/2411.15114
- [70]Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. In The Twelfth International Conference on Learning Representations (ICLR), 2024.arxiv.org/abs/2309.03409