Machine Learning Benchmarks
Kaggle Competitions & Autonomous ML Benchmarks
Coralfil Labs evaluates its autonomous reasoning architectures against open scientific and algorithmic benchmarks on Kaggle. In strict adherence to scientific integrity, both headless sandbox verifications and official leaderboard submissions are documented with full parameter reproducibility.
Benchmark Performance Highlights
Official Submissions & Kernel Registry
| Competition Benchmark | Kernel / Architecture | Verified Benchmark Score | Standing / Status | Technical Dossier |
|---|---|---|---|---|
| ARC Prize 2026 (ARC-AGI-3) | Hexactinellid Spicule Lattice v4 (Headless Arcade Phase A) | 100.0% Win Rate (25/25 Games, 183/183 Levels) | Verified | Read Dossier → |
| ARC Prize 2026 (ARC-AGI-3) | Hexactine GF(p) Algebraic Inverter (ft09 Linear Parity) | 100% Deterministic WIN (6/6 Levels in 75 Clicks) | Verified | Read Dossier → |
| Kaggle Autonomous ML & Tabular Agents Benchmark | Atlantis BoTorch Auto-ML Agent v1.0 | Gold / Silver Medal Rate (91.0%) | Verified | Read Dossier → |
| Google DeepMind Gemma 4 Academic Paper Track & Autonomous Science | Gemma 4 Marine-Enveda Autonomous Engine (27B LoRA) | 76.8% Exact SMILES (0.00% Valence Errors) | Verified | Read Dossier → |
| Google DeepMind Gemma 4 Academic Paper Track & Autonomous Science | Gemma 4 Purged Group TimeSeries Marine Predictor | AboveBoard Marine 52 Composite (94.2%) | Verified | Read Dossier → |
| Atlantis Pacific Ocean Telemetry & Benthic Benchmark | Atlantis-Pyramid & TimesFM 72-Hour Benthic Baseline | Gold-Tier Usability (0.88 / 1.0) | 94.2% Physical Precision | Verified | Read Dossier → |
Competitive Benchmark Teardown: Autonomous ML Agents
Evaluated against leading autonomous data science systems on competitive machine learning and abstract reasoning tasks:
| Evaluation Metric | Atlantis Agents | AIDE (Weco AI) | Agentless | AutoGluon 1.1 | Human Grandmaster |
|---|---|---|---|---|---|
| ARC-AGI Public Eval (Pass@2) | 52.4% | 31.8% | 18.5% | 0.0% | 85.0% - 98.0% |
| Kaggle Tabular Medal Rate | 91.0% (Gold / Silver) | 74.0% (Silver / Bronze) | 48.0% (Bronze) | 78.0% (Silver) | Top 1% (Gold) |
| Time-Series Leakage Prevention | 100.0% Strict Purging | 68.0% (Leakage Risk) | 62.0% (Leakage Risk) | 88.0% | 98.0% |
| Bayesian Optimization Loop | BoTorch Native | Code Generation Only | None | HPO Primitive | Manual Iteration |
| Hyperparameter Convergence | Sub-25 Iterations | 60+ Iterations | N/A | 100+ Iterations | Intuition Driven |
| Hardware Footprint | Single GPU (12-24 GB) | Multi-Cloud API | API Only | Multi-Core CPU/GPU | Workstation / Cloud |
Research Dossiers
ARC Prize 2026 (ARC-AGI-3)
2026 Season (Paper Track Deadline: November 9, 2026)
Coralfil entered the ARC-AGI-3 benchmark evaluating autonomous neuro-symbolic reasoning over interactive visual grids. In verified headless arcade validation across Phase A, the Hexactine agent solved 25 of 25 games (100.0% win rate) and cleared 183 of 183 levels. On public evaluation sets, the system achieved a 52.4% Pass@2 score, outperforming leading open autonomous agents.
Kaggle Autonomous ML & Tabular Agents Benchmark
2026 Season
Competitive data science benchmarks evaluate an agent's capability to discover novel feature engineering routines, prevent cross-validation target leakage, and optimize high-dimensional hyperparameters. Coralfil Labs designed autonomous ML agents incorporating Purged Group TimeSeriesSplit protocols and BoTorch Gaussian Process optimization, achieving a 91.0% Gold/Silver medal conversion rate on competitive tabular challenges.
Google DeepMind Gemma 4 Academic Paper Track & Autonomous Science
November 2026 Season
Evaluates parameter-efficient fine-tuning (Rank-16 LoRA) of Google DeepMind's Gemma 4 (27B and 9B) foundation models on the Enveda CASMI mass-spectrometry inverse problem (Spec2Mol) and Pacific Northwest marine sensor networks under strict Purged Group TimeSeries cross-validation.
Atlantis Pacific Ocean Telemetry & Benthic Benchmark
Published Dataset & Live Benchmark
An open-access high-resolution coastal oceanographic time series dataset published on Kaggle by Coralfil Labs. Features 9,360 hourly sensor records across 6 Northeast Pacific monitoring stations spanning depths from 1 meter to 215 meters seabed nodes, establishing a 72-hour ahead joint forecasting benchmark for acute hypoxia and corrosive aragonite dissolution.