← All Cracked Math projects
PM-W03NOT STARTEDWOWMATHPYTHONIMPACT ★★★★☆

Hunt a Packomania or AlphaEvolve packing record

Lagrange multipliers, exact Hessians and basin hopping, against packing records that AlphaEvolve just moved.

EST.50 h
REWARD+600 XP
AFTERPackomania + 4

ML-ARENA BENCHMARKS

PROOF: LEADERBOARD
METRICINDUSTRYMINE
Objective value vs the current Packomania record or published AlphaEvolve value for each n
GOAL · Match published records on >= 20 instances; >= 1 strict improvement accepted by Packomania, or a verified beat of a published AlphaEvolve packing value
Packomania records (E. Specht); AlphaEvolve report (arXiv 2506.13131), ~2.635 sum of radii for n = 26—
no results yet
YOUR RESULTS

Your recorded results appear here.

Brief

Packomania is a site that tracks the best known packings of circles. Formulate circle-packing problems as constrained nonlinear programs (optimise a nonlinear objective while the circles stay inside the container and do not overlap). The problems are equal circles in a square or in a circle, and AlphaEvolve's maximise-sum-of-radii of n circles in a unit square. Solve them with your own augmented-Lagrangian method, which turns the constraints into penalty and Lagrange-multiplier terms and solves a sequence of unconstrained problems. Its inner solver is a Newton/trust-region step with exact Hessians (a Newton step limited to a region where the quadratic model can be trusted). Put all of this inside a basin-hopping outer loop: randomly perturb the best solution and re-optimise, to escape local optima. Certify feasibility to 1e-12 with exact rational checks (exact fraction arithmetic instead of floats).

VICTORY CONDITIONS

  1. Match published records on >= 20 instances; >= 1 strict improvement accepted by Packomania, or a verified beat of a published AlphaEvolve packing value (Objective value vs the current Packomania record or published AlphaEvolve value for each n).
  2. A public repository with a README that shows your result and how to reproduce it.

Terms in plain words

Augmented Lagrangian
A constrained-optimisation method that adds both Lagrange-multiplier terms and a quadratic penalty for constraint violation, then updates the multipliers between solves.
Trust region
A safeguard that only accepts a Newton step within a radius where the local quadratic model is reliable, shrinking or growing that radius as it goes.
Basin hopping
A global search that repeatedly perturbs the current best solution and runs a local optimiser, keeping the result if it is better.
AlphaEvolve
Google DeepMind's system that uses language models to search for better algorithms and constructions; its report included new circle-packing values.
Exact rational check
Verifying that no circles overlap or leave the container using exact fractions, so rounding cannot hide a tiny violation.

Why it rates ★★★★☆

A dated, externally verified record on an AlphaEvolve-era benchmark is a strong, legible achievement.