Pipeline mind map
Built for this competition from the overview, rules and official metric notebook (v17), the public record catalogues, and our own runs of the metric on Oct 22, 2026. It follows a geometry search workflow; there is nothing to train.
- 01Bank the known packings
- 1.1Start from a public, credited portfolio
- 1.2Check every size with the official metric
- 1.3Leave out sizes you can't beat
- 02Respect the checker
- 2.1No exact contacts: add a tiny gap
- 2.2Write decimals with enough digits
- 2.3Re-read your CSV and score it locally
- 03Search where points are left
- 3.1Lower part of each band between squares
- 3.2Small n pays most per unit of L
- 3.3Trim big packings to fit smaller n
- 04Learn from proofs
- 4.1n = k² − 1 can't beat the grid (proved)
- 4.2Perfect squares are already 0
- 4.3Credit sources; winners must open-source
Key facts
- Host
- Kaggle (Walter Reade and Ashley Oldacre); sponsor Google LLC. Playground category: no Kaggle points or medals ↗
- Prizes
- $50,000: $12,000, $10,000, $8,000, $5,000 and $5,000 for the final top 5, plus $5,000 each for the leader on Jan 17, 2027 and on Apr 18, 2027 ↗
- Deadlines
- Entry and team-merger deadline Jul 11, 2027; final submission Jul 18, 2027; all at 11:59 PM UTC ↗
- What you submit
- submission.csv with id, n, square_id, x, y, angle (radians). Partial files are allowed; every missing or rejected puzzle is charged the plain grid ↗
- Submissions
- 10 a day; teams of up to 5; merges allowed until Jul 11 if the combined submission count fits ↗
- Data
- No files to download. The rules say no competition data is provided; public research (for example the Squares Project catalogue) counts as external data that everyone can reach ↗
- Winner licence
- Winning submissions and code are licensed CC BY 4.0 or Apache 2.0, and winners hand over reproducible code ↗
- Eligibility
- Kaggle account; 18 or the age of majority, whichever is higher; not resident in a sanctioned region ↗
- Participation (Oct 22, 2026, morning UTC)
- 275 entrants · 74 participants · 71 teams · 209 submissions, about 9 hours after launch ↗
Problem breakdown
For each n the metric takes the corners of all your squares and measures L, the larger of the horizontal and vertical spans. The puzzle scores L/√n − 1, and the leaderboard is the sum over n = 1…1,000. √n is the area bound, so a perfect square n = k² packed as a k × k grid scores exactly 0. Everything else is above zero, and most of the true minimum values are unknown: the record page maintained by David Ellsworth lists best-known packings, only a few of which are proved optimal.
A missing puzzle is charged the plain grid, side ⌈√n⌉. We ran the official metric on a file containing only n = 1 and got 28.937669916550586, the score of a full grid for all 1,000 sizes. So a blank start is not a zero; it is the baseline every public notebook is measured against. You never have to submit a size you can't improve.
Göbel's packing for n = 5 puts four squares in the corners and a fifth turned 45° in the middle, side 2 + √2/2 ≈ 2.7071 instead of 3. We wrote it out with exact touching and ran the official metric: the packing was rejected and n = 5 was charged the grid, total still 28.9377. With every gap widened by 10⁻¹², the same packing passed, L came out as 2.707106781187548 and the total dropped to 28.80668, a gain of 0.131. The checker first tests pairs in ordinary floating point with a 10⁻⁸ margin, then re-checks close pairs with interval arithmetic at up to 500 digits; a contact that is exact on paper becomes an overlap once rounded to decimals. Build every packing with a small clearance.
Chris Deotte's EDA notebook embeds a 1,000-size reference curve from the public portfolio, total 22.6024. We decoded it and compared it with the grid. 543 sizes beat the grid; together they account for 6.34 points of gain. Sizes 101–1,000 provide 4.96 of those points, but the largest gains per size are at small n just above a perfect square: n = 5 (0.131), 10 (0.093), 17 (0.079), 26 (0.074), 37 (0.066) and 50 (0.061).
Split each band between two squares, k² to (k+1)², into a lower and an upper part. In that reference portfolio, none of the 368 sizes in the upper 40% of a band beat the grid, while 528 of the 547 sizes in the lower 55% did. 457 sizes in total are still on the grid, and they hold 7.36 of the 22.60 points (33%). Some of that is locked: s(k² − 1) = k, so the 30 sizes like 8, 15, 24 … 960 cannot beat the grid. Nagamochi's 2005 paper also claimed s(k² − 2) = k, but a September 2026 arXiv paper found a gap in that proof and re-proves only the k² − 1 case. Apart from small sizes proved separately (2, 7 and 14), the k² − 2 sizes such as 23, 34, 47 … 959 now rest on a proof with a known gap; nobody has shown they can beat the grid, though.
A side reduction ΔL at size n is worth ΔL/√n. Shaving 0.001 off L is worth 0.00032 at n = 10, 0.0001 at n = 100 and 0.00003 at n = 1,000. Day-one notebooks differ by about 0.02 in total (22.586 to 22.602 on the metric), so early movement comes from collecting more catalogued packings and polishing many sizes a little, not from one clever layout.
The notebooks we read build mostly on three credited collections: Joshua Levy's Squares Project (machine-readable witnesses up to n = 324), Francisco Couzo's certificate catalogue (up to n = 998) and recent refinements posted to the Squares Project issue tracker. They overlap, so 966 witnesses are not 966 separate discoveries. The packings belong to the original discoverers (Göbel, Friedman, Ellsworth, Cantrell, Bidwell, Schadt and others); credit them in anything you publish.
What you get
There are no files to download. The official metric is a public Kaggle notebook you can copy and run offline (it needs mpmath and pandas). It writes submission_details.csv with the L it accepted for every n, so you can see which sizes were rejected.
- Official metric (v17)
- Parses decimal strings, measures L at 60 significant digits, runs a fast floating-point overlap test with a 10⁻⁸ margin and escalates unclear pairs to interval arithmetic at 50, 100, 200 and 500 digits
- Coordinates
- Plain decimal literals, scientific notation allowed. Symbolic forms like sqrt(2) are rejected and that puzzle is charged the grid
- Angles
- Radians, one per square
- Size of a full file
- Sum of n for n = 1…1,000 is 500,500 rows; public notebooks submit only the 543 or so improved sizes
submission.csv
id,n,square_id,x,y,angle 0,1,0,0.5,0.5,0 1,2,0,0.5,0.5,0 2,2,1,1.5,0.5,0 ... # omit any n you can't beat: it is scored as the grid
From the competition data / overview page ↗
Metric: Sum over n = 1…1,000 of L/√n − 1
For each puzzle, L is the side of the smallest axis-aligned square around all your squares, and the score is how far L is above √n as a fraction of √n. Lower is better. A rejected or missing puzzle is charged the grid, ⌈√n⌉/√n − 1, so a bad attempt never costs you more than leaving it out.
The metric computes every puzzle's score directly from your file, with no hidden labels, so the only uncertainty is whether the checker accepts each puzzle. The overview says the highest leaderboard score on Jan 17 and Apr 18, 2027 wins each milestone, with ties going to the earlier submission. ↗
Score of a file with no improved puzzles (our run of the official metric): 28.9377 ↗
Public baselines
Public notebooks we read on Oct 22, 2026 through Kaggle's public API (votes from Kaggle's list that morning). Scores are the authors' own runs of the official metric; we didn't submit any of them. The reference curve we analysed is the one embedded in Chris Deotte's notebook.
| Notebook | ▲ | LB | Why open it |
|---|---|---|---|
| Squares in Squares | Visual EDA, Metric & Gallery | 1 | 22.6024 (reference curve) | Read this first. It walks through the metric, explains why tilted squares beat the grid at n = 5, and shows the exact-contact trap with pictures. It also carries the 1,000-size reference curve we decoded for the numbers above. |
| Squarecraft | Beyond the Grid | 5 | 22.7417 (title) | A CPU-only build from 150 published constructions: it trims big packings to fit smaller n and uses sparse linear programming to tighten them, then checks every size with the metric. The clearest code to learn the trim-and-reuse idea. |
| A Square Deal | Research to Verified Packings | 3 | 22.602120846 | Compares the three public collections, merges them per size with credits and validates the written CSV with the unmodified metric. Good for seeing how much each source adds. |
| Squares in Squares | CUDA Packing Search | 9 | 22.585624489618 | The best number on day one. Searches tilted 'crystal' rows with four corner staircases on a GPU and lets each row slide on its own. Replays on CPU by default; you only need CUDA to rerun the search. |
| 22.590879 | CPU Window Closure + Contact Polish | 1 | 22.590879560380 | Finds, for a large packing, the smallest window that still holds n whole squares (a sweep with a segment tree), then polishes contacts. Shows how much is left in reusing big packings for smaller n. |
Most-voted notebooks right now
Refreshed daily from the Kaggle API (last 2026-11-16). Not reviewed by an editor; the table above is the reviewed list.
- Squares in Squares - Per-Puzzle Leaderboard ▲151
- Squares in Squares Living Atlas of Tight Packings ▲86
- Square Packing Lab | CPU/CUDA/MPS ▲68
- Squares in Squares | CUDA Packing Search ▲64
- LB 22.1945500 · Sacred Geometry of 1000 Squares ▲61
- 🧩 Squares in Squares | Deep EDA & Analogies ▲59
- Squares in Squares | Neighbour Transfer + Squeeze ▲58
- Squares 101: the Starter Kit (13,100+ downloads) ▲58
Timeline
- Start date
- First $5,000 milestone: leaderboard leader at 11:59 PM UTC
- Second $5,000 milestone
- Entry and team-merger deadline
- Final submission deadline, 11:59 PM UTC
Step-by-step plan
- Day 1: score nothing, on purpose
Copy the metric notebook, run it on a CSV with only n = 1 and confirm you get 28.9377. Then add Göbel's n = 5 with a 10⁻¹² gap and watch the total drop by 0.131. Now you know the checker and the file format work on your machine.
- Day 1–2: start from credited packings
Fork one of the portfolio notebooks above, or load the Squares Project witnesses yourself. Submit only the sizes that beat the grid, keep a per-size table of L and source, and score the written CSV, not your in-memory arrays.
- Week 1: pick a band, not a size
Choose a range such as n = 101–324 and list, for each size, the gap between the current L and the next possible improvement. Work in the lower part of each band (just above a perfect square), where the reference portfolio shows almost all gains.
- Week 2–4: one search method
Try one method end to end: trimming larger packings to smaller n, simulated annealing on centres and angles, or a parametric family of tilted rows. Always finish with a clearance step and the official metric; a single overlap throws the whole size back to the grid.
- Before each milestone
Jan 17 and Apr 18 pay the leader at that moment. Keep a validated best file ready, and resubmit it after each improvement rather than at the last minute; ties go to the earlier submission.
- Through Jul 18
Share code only on Kaggle's public pages for this competition, keep your sources credited, and remember winners must release their code under CC BY 4.0 or Apache 2.0.