Skip to content
← All news

News2 min read

Two Kaggle finishes in the top 3%

Ten agent workflows on Kaggle, with every result sorted by what kind of rank it is. Two closed with official final ranks in the top 3%. Twelve more top-5% results are late-submission estimates, and six live public boards had us in the top 5% on 15 August.

Still running

Fourteen of these competitions were open on 15 August. Public ranks move, and the private board is the only one that counts. Everything below is a snapshot with a date on it, and the follow-up says how the open boards closed.

Kaggle is the setting we keep coming back to, because it is one of the few places an agent loop can be wrong in public. Thousands of humans work on the same problem, the deadline does not negotiate, and the private leaderboard arrives after every decision has already been made.

What was run ​

Built on Humanize, we designed roughly ten distinct agent workflows and pointed all of them at live competitions. They differ in the things we think actually matter on a long run:

  • how context is managed across hours of work,
  • how several skills are orchestrated inside one run,
  • how much of an end-to-end machine-learning engineering pipeline the agents own,
  • whether a second agent reviews the first one's work, and
  • whether attempts run in parallel or in sequence.

Three kinds of rank ​

A Kaggle result can be one of three different claims, and we count them apart.

Official final
2in the top 5%, both in the top 3%

An exact place on a closed board.

The private leaderboard, after the deadline. This is a finish.

Late estimate
12in the top 5%, five in the top 1%

A real score, placed afterwards.

Submitted after the competition closed and placed against the frozen final board. It is real as a score. It is not a rank, a medal, or evidence of having competed.

Live public
6in the top 5%, four in the top 3%

A position that can still move.

The public leaderboard of a competition still open. It is official, and it is a weather report.

Best result per competition · Team Radar · 15 August 2026
Completed19 competitions
2
0
12
Ongoing14 competitions
0
6
0
Each competition counts once, by its best tracked result. Switch to see the top 1%.

So the honest headline is two finishes, not fourteen: Predicting Student Health Risk closed at 67th of 3,355 (top 2.00%) and ROGII Wellbore Geology at 137th of 6,125 (top 2.24%). All five top-1% results among the completed competitions are late estimates.

Every result, by kind ​

Each table is sortable. Click a row to pin it.

Completed · official ranks (final, or public where no private rank was recorded)
Predicting Student Health Risk673,3552.00%
ROGII - Wellbore Geology Prediction1376,1252.24%
Predicting Stellar Class4472,81615.87%
The 2026 NeuroGolf Championship4822,96316.27%

The two below the top 5% are public-board ranks for competitions that had closed, where no private rank was recorded.

Completed · late-submission estimates (not ranks)
NeurIPS 2025 - Google Code Golf Championship11,1420.09%
Critical temperature of superconductors14940.20%
Helios Corn Futures Climate Challenge11530.65%
Electricity consumption11500.67%
CMI - Detect Behavior with Sensor Data192,6570.72%
Hash Code Archive - Photo Slideshow Optimization1891.12%
Salary Prediction for Job Postings1831.20%
Hash Code Archive - Drone Delivery21301.54%
Hotel booking demand74061.72%
Information Retrieval Agent 202582223.60%
Scintillation Detector Signal Types4984.08%
March Machine Learning Mania 2025841,7274.86%
March Machine Learning Mania 2022 - Men’s479305.05%
Santa Gift Matching Challenge404289.35%
Predicting F1 Pit Stops3943,02213.04%
Ongoing · live public board, 15 August
Biohub - Cell Tracking During Development82,3780.34%
Predicting Smartphone Addiction111,9030.58%
Solar Filament Segmentation Challenge 202672962.36%
Titanic - Machine Learning from Disaster26210,7112.45%
Predicting Soil Grain Size Distributions from Images3833.61%
LLM Classification Finetuning92164.17%
Kaggriculture3084,5816.72%
Spaceship Titanic1591,6359.72%
UMUD Challenge: Muscle Architecture in Ultrasound Data1514910.07%
The Pokémon Company - PTCG AI Battle Challenge Simulation9556,82913.98%
CUHK-X Competition Small Model Track3119915.58%
RSNA Knee Abnormality Detection3001,59118.86%
Hyperspectral Object Detection Challenge 2026104025.00%
ARC Prize 2026 - ARC-AGI-23751,49425.10%

Two of the live top-3% positions need a footnote. Titanic is a getting-started practice competition. And Biohub, eighth of 2,378 that day, has since closed at 188th of 3,947: still a top-5% finish, and a smaller one.The Team Radar export behind these tables leaves out a fixed list of 12 competitions, so the counts are over the 19 completed and 14 ongoing that remain. The follow-up post counts every competition.

For comparison, Codex on GPT-5.5 at xhigh reasoning effort generally peaked at around the top-5% level across most of the same competitions. Same models, a different arrangement, and the arrangement is worth the tail of the distribution.

What we think is doing the work ​

Our preliminary read is that the gains come primarily from the diversity of reasoning strategies that multi-agent collaboration makes available. A single strong agent narrows early and commits. Several agents arranged to disagree keep more than one line of attack alive long enough to be tested against a validation split that has not been quietly bent to fit.

That is a hypothesis with a number attached, not a conclusion.

Further analysis is ongoing. The most controlled version of it so far is HMA on MLE-bench: two agents taking turns, against each of them alone.

The audit · the live leaderboard · HMA

Read next

News78.2% MLE-bench medal rate in 6 hoursTwo coding agents taking turns over one workspace, each starting fresh, beat both of them working alone on 75 MLE-bench tasks: 78.2% any-medal against 72.4% and 68.0%, in the same six hours. Self-reported, and not on the official leaderboard.Read it NewsBiohub final: #188 of 3,947, top 5%In August an agent workflow sat 8th on the Biohub cell-tracking public leaderboard. The final private rank is 188th of 3,947, the top 4.76%: lower, but an official finish. Plus the Kaggle tally when only final ranks count.Read it News#1 on Lean-Eval v1 first solvesLean-Eval froze its first 128-problem release and moved everything older to an archive. On v1 our GPT-5.6 flow has the most first solves, 30, and is one behind the leader on total; on the archive it has 170 of 171.Read it