THE SIGNAL IN ONE SENTENCE

A language model repeatedly changed an optimization algorithm, but only mathematically verified improvements were allowed to survive.

01

WHAT ACTUALLY CHANGED

Researchers introduced Discovery Loop, a system that uses a language model to iteratively modify optimization programs. It begins with a seed solver, reviews a scoreboard and ideas from earlier attempts, proposes a change, and sends the candidate to an independent verifier.

The experiment targeted variable-radius circle packing, where a program tries to arrange differently sized circles inside a bounded space as efficiently as possible. Packomania maintains the best-known solutions for these problems across many values of N.

Within 15 iterations, Discovery Loop improved the best-known solutions for ten N values between 101 and 114. The reported gains ranged from 2.4 to 5.4 percent, and the total language-model cost was $27.72.

Packomania independently accepted the new results. The public repository includes the implementation and solutions, allowing others to inspect how the loop proposed, tested, retained, and discarded candidate programs.

02

WHY THIS MATTERS

This is a narrow result, which is part of what makes it persuasive. The system did not declare itself a general scientist. It improved a specific optimization record under rules that can be checked exactly.

The architecture separates creative proposal from judgment. The model can improvise unusual program changes because fluent explanations earn it no credit. Every candidate must produce a result the verifier accepts, and failures leave the scoreboard unchanged.

That pattern could make automated search useful far beyond expensive frontier-lab demonstrations. Give a relatively cheap model a bounded problem, durable experimental memory, and a judge it cannot charm. Then let it be weird until something measurable improves.

FIG. 065LET THE MODEL PROPOSE, LET THE VERIFIER DECIDE
1SEED SOLVER→
2MODEL MODIFIES→
3RUN CANDIDATE→
4VERIFY RESULT→
5KEEP IMPROVEMENT
Creative proposals move forward only when an independent measurement shows that the solution actually improved.

03

WHERE IT COULD HELP

  • Search for better packing and scheduling algorithms
  • Evolve optimization code using inexpensive model calls
  • Explore routing and resource-allocation strategies
  • Keep only improvements that pass an independent verifier

KEEP A HAND ON THE WHEEL

Ten accepted records in one optimization family do not establish general scientific autonomy. The verifier must accurately represent the real objective, and a system can still exploit weaknesses in an incomplete evaluation even when every submitted result is technically valid.

04

TERMS WORTH KEEPING

SOURCES AND VERIFICATION STATUS

This article was written from the materials below. Product claims and dates were checked against those sources on September 7, 2026.

PUBLICATION RECEIPT: Revision 1. Approved by Zak and published September 7, 2026.

THE PUBLICATION ENGINE

WANT A SIGNAL OF YOUR OWN?

We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.

WORK WITH US