News · 2026-09-07

Discovery Loop gets ten AI-assisted circle-packing candidates accepted by Packomania

Discovery Loop produced ten circle-packing candidates accepted by the Packomania record table after using Claude Fable 5.1 to revise a solver program. The result is notable not because an agent solved mathematics by itself, but because the project pairs a bounded objective, an independent verifier, public records and a modest reported cost.

Key facts

• The paper reports ten accepted candidates for N=101–103, 105–109, 111 and 114.
• Packomania’s live table records Wes Sander and discovery-loop as the source.
• The run used Claude Fable 5.1 via Claude CLI on a Core i7-13700KF PC with 32 GB RAM.
• It ran roughly eight hours overnight, cost $27.72 and covered 15 iterations.

Circle packing asks how to arrange a fixed number of identical circles in a container as efficiently as possible. A tiny improvement can be a legitimate mathematical result, but it is also easy to overstate. Discovery Loop’s design makes the claim unusually inspectable. The model is not asked to emit a finished geometric layout. It receives the current solver, a score board and a history of ideas, then writes a complete replacement solver. The system runs that program, checks candidate output and feeds the score back into the next cycle.

This is an instance of program synthesis: the model improves the code that searches the space instead of searching every arrangement itself. Think of it as hiring a mathematician to redesign a factory’s jig rather than asking them to hand-assemble every part. A better jig can produce better candidates repeatedly, and the verifier determines whether a claimed improvement is physically valid.

The validation detail is the important one. The repository says its circle-packing plugin fetches live Packomania records, invokes an independent verifier and performs a stricter release check. The relevant problem implementation uses zero tolerance for checking and adds tighter margins for release candidates. The paper reports six parallel evaluations for the published run, while the public default is three workers.

A correction protects the story from becoming too magical. Every one of the ten record-setting rows first improved in iteration 0, the seed solver. Later iterations improved overall solver performance and total score; they did not individually generate the ten headline deltas. That distinction is exactly why public logs and a table matter. The project is still an impressive example of using a model as a search-and-engineering collaborator, but it is not evidence of a self-improving system repeatedly discovering each record from scratch.

Packomania’s update log gives the external acceptance: on September 2 it said Wes Sander and MoltFire had found ten new candidates, and on September 3 it recorded the new candidates for the listed N values. That is stronger than a benchmark score because it is a domain maintainer’s public record. It is also narrower: circle packing is a highly structured optimisation problem with an exact verifier.

The caveat is therefore central. A verified objective changes everything. This method does not show that an agent can autonomously make reliable discoveries in fields where the target is vague, feedback is delayed, experiments are expensive or validation is contested. It shows what happens when evaluation is crisp and an AI can modify the program that conducts the search.

Why it matters: this is the pattern to watch for credible agentic science—closed-loop iteration, explicit cost, reproducible artifacts and independent validation—not a leaderboard claim detached from a real acceptance mechanism.

Primary source, verified: read the paper → (arXiv 2609.05093)

Ground Truth. (2026, September 7). Discovery Loop gets ten AI-assisted circle-packing candidates accepted by Packomania. Ground Truth. https://groundtruth.day/news/discovery-loop-gets-ten-packomania-circle-packing-candidates-accepted.html

@misc{groundtruth:discovery-loop-gets-ten-packomania-circle-packing-candidates-accepted,
  title  = {Discovery Loop gets ten AI-assisted circle-packing candidates accepted by Packomania},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/discovery-loop-gets-ten-packomania-circle-packing-candidates-accepted.html}
}

Topics: research · agents · program-synthesis · mathematics · verification