{"contributors":[{"id":19,"fullName":"François Chollet & Mike Knoop","handle":"fchollet & mikeknoop","affiliation":"Creators of ARC-AGI & ARC Prize","achievement":"Created ARC-AGI benchmark (2019) & Launched ARC Prize (2024)","description":"François Chollet (Creator of Keras) defined the ARC-AGI benchmark in 2019 to measure general intelligence. Mike Knoop (Co-founder Zapier) launched the $1M+ ARC Prize in 2024 to accelerate progress toward AGI.","yearStart":null,"yearEnd":null,"score":"Founders","approach":"Designed ARC as a measure of intelligence based on skill-acquisition efficiency rather than skill itself. Organized competitions to guide research.","uniqueTechnique":"Created a benchmark that remained unsolved for 5+ years and a prize structure that incentivized open-source breakthroughs.","links":{"github":"https://github.com/fchollet/ARC-AGI","twitter":"https://x.com/fchollet","website":"https://arcprize.org"},"teamName":"ARC Prize Foundation","category":"founder","imageUrl":"/arc founders.png","rank":0,"createdAt":"2025-12-30T19:27:42.523Z"},{"id":25,"fullName":"Dries Smit","handle":"DriesSmit","affiliation":"Independent Researcher / MindsAI & Tufa Labs","achievement":"1st Place ARC-AGI-3 Agent Preview 2025 (12.58%)","description":"Rising ARC-AGI researcher who won 1st place in the ARC-AGI-3 Agent Preview 2025 with his RL-based \"StochasticGoose\" agent, scoring 12.58% and leading all teams in efficiency. Also contributed to the MindsAI & Tufa Labs team in ARC Prize 2025, reaching 3rd place on the public ARC-AGI-2 leaderboard with 15.42%. His work focuses on efficient RL agents and test-time adaptation methods for ARC tasks.","yearStart":2025,"yearEnd":2026,"score":"12.58% (1st Place)","approach":"RL-based agent with test-time adaptation, focusing on efficiency.","uniqueTechnique":"StochasticGoose - an efficient RL agent that led all teams in computational efficiency while achieving top accuracy.","links":{},"teamName":"Tufa AI","category":"arc3_preview","imageUrl":"/dries.png","rank":1,"createdAt":"2025-12-30T19:27:42.692Z"},{"id":7,"fullName":"Team NVARC (Jean-François Puget & Ivan Sorokin)","handle":"JFPuget & lytic","affiliation":"NVIDIA - Machine Learning & Kaggle Grandmasters","achievement":"1st Place ARC Prize 2025 (24.03%)","description":"Jean-François Puget (6x Kaggle Grandmaster, ENS Ulm alumni, ML PhD) and Ivan Sorokin (ML Researcher, Kaggle Grandmaster, Math Olympiad 2025 winner) won 1st place in ARC Prize 2025 with 24.03% accuracy as Team NVARC.","yearStart":2025,"yearEnd":2025,"score":"24.03% (1st Place)","approach":"Synthetic-data-driven ensemble of an improved ARChitects-style, test-time-trained model and TRM-based components that reaches ~24% on ARC-AGI-2 under Kaggle contest constraints.","uniqueTechnique":"TBA","links":{"github":"https://github.com/jfpuget","kaggle":"https://www.kaggle.com/cpmpml","twitter":"https://x.com/JFPuget","linkedin":"https://www.linkedin.com/in/lytic/"},"teamName":"NVARC","category":"competition_winner","imageUrl":"/jfPuget3.png,/ivanARC2.png","rank":1,"createdAt":"2025-12-30T19:27:40.935Z"},{"id":1,"fullName":"Alexia Jolicoeur-Martineau","handle":"jm_alexia","affiliation":"Samsung SAIT Montréal, Senior AI Researcher","achievement":"2025 Top Paper Award: \"Less is More: Recursive Reasoning with Tiny Networks\" (TRM)","description":"Lead author of the Tiny Recursive Model (TRM) work on ARC-AGI—a groundbreaking ~7M-parameter, 2-layer recursive network that repeatedly refines an internal reasoning state and answer instead of relying on huge LLMs. TRM became the leading open-source approach on ARC-AGI shortly before the 2025 ARC Prize deadline, though its compute budget exceeded competition constraints.","yearStart":2025,"yearEnd":2025,"score":"~45% ARC-AGI-1, ~8% ARC-AGI-2","approach":"Tiny Recursive Model (TRM)—a compact 2-layer network with recursive refinement that iteratively improves its reasoning state and answer prediction. Trained from scratch without LLM pretraining.","uniqueTechnique":"Demonstrated that ~7M parameters can match or beat models 10,000× larger (DeepSeek R1, o3-mini, Gemini 2.5 Pro) on ARC-AGI-2, proving \"less is more\" for abstract reasoning.","links":{"github":"https://github.com/SamsungSAILMontreal/TinyRecursiveModels","kaggle":"https://www.kaggle.com/code/alexiajm/arc-agi-without-pretraining","papers":["https://arxiv.org/abs/2510.04871"],"twitter":"https://x.com/jm_alexia","website":"https://alexiajm.github.io/2025/09/29/tiny_recursive_models.html"},"teamName":"Samsung SAIL Montréal","category":"top_paper_award","imageUrl":"/alexiaJM4.png","rank":1,"createdAt":"2025-12-30T19:27:40.775Z"},{"id":6,"fullName":"ARChitects 2024 (Franzen, Disselhoff, Hartmann)","handle":null,"affiliation":"Johannes Gutenberg University Mainz, Germany","achievement":"1st Place ARC Prize 2024 (53.5%)","description":"Three-person team (Daniel Franzen, Jan Disselhoff, David Hartmann) won ARC Prize 2024, scoring 53.5% on ARC-AGI-1 private eval with \"The LLM ARChitect,\" using test-time training plus a product-of-experts ensemble over different perspectives of each grid.","yearStart":2024,"yearEnd":2024,"score":"53.5% (1st Place)","approach":"Product of Experts ensemble using multiple perspectives to improve solution selection, combining LLM fine-tuning with test-time training","uniqueTechnique":"Test-time training with augmentation-based validation using depth-first search for token selection","links":{"github":"https://github.com/da-fr/arc-prize-2024","papers":["https://arxiv.org/abs/arc-prize-2024"]},"teamName":"ARChitects","category":"competition_winner","imageUrl":"/ARChitechts.png","rank":1,"createdAt":"2025-12-30T19:27:40.904Z"},{"id":5,"fullName":"ARChitects (Franzen, Disselhoff, Hartmann)","handle":null,"affiliation":"JGU Mainz / Lambda, Inc.","achievement":"2nd Place ARC Prize 2025 (16.53%)","description":"The ARChitects Kaggle team (Daniel Franzen 1*, Jan Disselhoff 1*, David Hartmann 2*) dominated ARC-AGI research with their ARC Prize 2025 Solution Summary, combining a 2D-aware masked-diffusion LLM with recursive self-refinement and LambdaLabs compute support.","yearStart":2025,"yearEnd":2025,"score":"16.53% (2nd Place)","approach":"2D-aware masked-diffusion LLM with recursive self-refinement","uniqueTechnique":"Masked diffusion approach applied to 2D grid reasoning with iterative refinement","links":{"github":"https://github.com/da-fr/arc-prize-2024","papers":["https://arxiv.org/abs/arc-prize-2024"],"website":"https://lambdalabsml.github.io/ARC2025_Solution_by_the_ARChitects/"},"teamName":"ARChitects","category":"competition_winner","imageUrl":"/ARChitechts.png","rank":2,"createdAt":"2025-12-30T19:27:40.883Z"},{"id":2,"fullName":"SOAR Team (Pourcel, Colas, Oudeyer)","handle":null,"affiliation":"LLM4Code / Inria & Sorbonne University","achievement":"2nd Place 2025 Paper Award: \"Self-Improving Language Models for Evolutionary Program Synthesis\" (SOAR)","description":"Authors of SOAR, a self-improving evolutionary program synthesis framework that alternates between LLM-driven evolutionary search for ARC-AGI programs and hindsight fine-tuning on its own search traces, steadily improving the model as a search operator without human-engineered DSLs or curated solution datasets.","yearStart":2025,"yearEnd":2025,"score":"2nd Place 2025 Paper Award","approach":"Self-improving evolutionary program synthesis: use an LLM to sample and refine candidate programs, then convert successful and near-miss search attempts into training data for a hindsight fine-tuning phase that improves the LLM search operator over iterations.","uniqueTechnique":"Shows how to close the loop between evolutionary search and LLM training: SOAR turns search traces into supervised data, leveraging positive transfer across tasks to significantly boost ARC-AGI performance without hand-designed DSLs or solution datasets.","links":{"github":"","papers":[""],"youtube":"https://www.youtube.com/watch?v=9lIuoslCHWI"},"teamName":"SOAR","category":"paper_award","imageUrl":"/julienPourcel.png","rank":2,"createdAt":"2025-12-30T19:27:40.801Z"},{"id":10,"fullName":"Jack Cole (2024)","handle":null,"affiliation":"MindsAI Team","achievement":"MindsAI: Highest score 55.5% on ARC Prize 2024 (ineligible)","description":"Core researcher on Team MindsAI, which achieved the top 55.5% score on the ARC Prize 2024 private evaluation set using heavy test-time training, while remaining ineligible for official prizes because the solution was not open sourced.","yearStart":2024,"yearEnd":2024,"score":"55.5% (Highest, ineligible)","approach":"Test-time training (TTT) and ARC-specific domain knowledge as part of MindsAI's proprietary system.","uniqueTechnique":"Helped pioneer test-time training for ARC-AGI, setting the 55.5% high score on 2024 private eval.","links":{},"teamName":"MindsAI","category":"competition_winner","imageUrl":"/jackcole.jpeg","rank":2,"createdAt":"2025-12-30T19:27:41.039Z"},{"id":27,"fullName":"IsaiahP","handle":"pressman1","affiliation":"MindsAI & Tufa Labs (2025); Tufa Labs (2026)","achievement":"3rd Place ARC Prize 2025 with MindsAI & Tufa Labs","description":"Listed on the MindsAI & Tufa Labs Kaggle team that ARC Prize ranked third in 2025, alongside Jack Cole, Dries Smit, Mohamed Osman and Michael Hodel. Listed on both Tufa Labs teams in 2026.","yearStart":2025,"yearEnd":2025,"score":"12.6% (3rd Place, team)","approach":null,"uniqueTechnique":null,"links":{"kaggle":"https://www.kaggle.com/pressman1","website":"https://www.kaggle.com/competitions/arc-prize-2025/writeups/mindsai-and-tufa-labs-arc-prize-2025-solution"},"teamName":"MindsAI & Tufa Labs","category":"competition_winner","imageUrl":"/news-images/people/isaiahp.webp","rank":3,"createdAt":"2026-10-09T22:46:37.198Z"},{"id":9,"fullName":"Team MindsAI (Jack Cole & Dries Smit)","handle":null,"affiliation":"MindsAI & Tufa Labs","achievement":"3rd Place ARC Prize 2025 (15.42% on public ARC-AGI-2)","description":"Jack Cole (test-time training pioneer) and Dries Smit (RL agent specialist) teamed up for ARC Prize 2025, reaching 3rd place on the public ARC-AGI-2 leaderboard with 15.42%.","yearStart":2025,"yearEnd":2025,"score":"15.42% (3rd Place)","approach":"Test-time training (TTT) combined with RL-based adaptation.","uniqueTechnique":"Combined TTT expertise with efficient RL agent approaches.","links":{},"teamName":"MindsAI","category":"competition_winner","imageUrl":"/jackCole2.png,/dries.png","rank":3,"createdAt":"2025-12-30T19:27:40.993Z"},{"id":3,"fullName":"Isaac Liao","handle":null,"affiliation":"Carnegie Mellon University, Graduate Researcher","achievement":"3rd Place 2025 Paper Award: \"ARC-AGI Without Pretraining\" (CompressARC)","description":"Creator of CompressARC, an MDL-based neural \"code golf\" system that trains a single small network per puzzle from scratch and selects the model that minimizes description length, solving ARC-AGI puzzles using only the target puzzle data with no pretraining or external datasets.","yearStart":2025,"yearEnd":2025,"score":"ARC-AGI-1: 34.75% (training), 20% (evaluation); AR","approach":"Minimum Description Length (MDL) compression-driven solver that treats ARC-AGI as a lossless compression problem, performing single-puzzle training of a compact neural network and choosing architectures/hyperparameters that yield the shortest overall description of inputs and outputs.","uniqueTechnique":"Shows that pure compression at inference time can yield intelligent behavior: a single-puzzle-trained neural network that optimizes description length instead of accuracy, achieving non-trivial ARC-AGI-1/2 performance without any pretraining or synthetic data.","links":{"github":"https://github.com/iliao2345/CompressARC","papers":["https://iliao2345.github.io/blog_posts/arc_agi_without_pretraining/ARC_AGI_Without_Pretraining.pdf"],"website":"https://iliao2345.github.io/blog_posts/arc_agi_without_pretraining/arc_agi_without_pretraining.html","youtube":"https://www.youtube.com/watch?v=N9GvFj0cE9s"},"teamName":null,"category":"paper_award","imageUrl":"/isaacliao.png","rank":3,"createdAt":"2025-12-30T19:27:40.827Z"},{"id":28,"fullName":"alijs","handle":"alijs1","affiliation":"Kaggle competitor","achievement":"3rd Place ARC Prize 2024 (40%)","description":"Finished third in ARC Prize 2024 on the ARC-AGI-1 private evaluation as a one-account team (alijs1), and entered both 2026 contests.","yearStart":2024,"yearEnd":2024,"score":"40% (3rd Place)","approach":null,"uniqueTechnique":null,"links":{"kaggle":"https://www.kaggle.com/alijs1","website":"https://www.kaggle.com/competitions/arc-prize-2024/writeups/alijs-3rd-place-solution"},"teamName":"alijs","category":"competition_winner","imageUrl":"/news-images/people/alijs.webp","rank":3,"createdAt":"2026-10-09T22:46:37.206Z"},{"id":26,"fullName":"Lonnie","handle":"lonnieqin","affiliation":"Kaggle competitor","achievement":"4th Place ARC Prize 2025 (6.7%)","description":"Finished fourth in ARC Prize 2025 on the ARC-AGI-2 private evaluation as a one-account team (lonnieqin), and entered both 2026 contests.","yearStart":2025,"yearEnd":2025,"score":"6.7% (4th Place)","approach":null,"uniqueTechnique":null,"links":{"kaggle":"https://www.kaggle.com/lonnieqin","website":"https://www.kaggle.com/competitions/arc-prize-2025/writeups/arc-prize-2025-competition-writeup-5th-place"},"teamName":"Lonnie","category":"competition_winner","imageUrl":"/news-images/people/lonnie.webp","rank":4,"createdAt":"2026-10-09T22:46:37.157Z"},{"id":29,"fullName":"gromml","handle":"gromml","affiliation":"PoohAI (2024)","achievement":"5th Place ARC Prize 2024 with PoohAI (37%)","description":"Listed among the four accounts on PoohAI, which ARC Prize ranked fifth in 2024 on the ARC-AGI-1 private evaluation. Entered both 2026 contests.","yearStart":2024,"yearEnd":2024,"score":"37% (5th Place, team)","approach":null,"uniqueTechnique":null,"links":{"kaggle":"https://www.kaggle.com/gromml","website":"https://www.kaggle.com/competitions/arc-prize-2024/writeups/poohai-5th-place-solution"},"teamName":"PoohAI","category":"competition_winner","imageUrl":null,"rank":5,"createdAt":"2026-10-09T22:46:37.222Z"},{"id":8,"fullName":"Guillermo Barbadillo","handle":"ironbar","affiliation":"Veridas AI","achievement":"5th Place ARC Prize 2025 - 6.53% accuracy (semi-private eval)","description":"Continued development of multi-technique approach, achieving 5th place in ARC Prize 2025 with strong public Kaggle leaderboard performance.","yearStart":2024,"yearEnd":2025,"score":"6.53% (5th Place) / 11.94% (Public Kaggle)","approach":"Omni-ARC approach combining multiple techniques for comprehensive puzzle solving","uniqueTechnique":"Multi-method ensemble combining neural and symbolic approaches","links":{"kaggle":"https://www.kaggle.com/guillermobarba","twitter":"https://x.com/guille_bar"},"teamName":"Veridas","category":"competition_winner","imageUrl":"/guillermo.png","rank":5,"createdAt":"2025-12-30T19:27:40.957Z"},{"id":20,"fullName":"Simon Strandgaard (neoneye)","handle":"neoneye","affiliation":"Independent developer & ARC community contributor","achievement":"ARC community datasets and tools curator; long-time maintainer of ARC-related GitHub resources","description":"Independent developer who has collected, curated, and published many ARC-related datasets and tools on GitHub, making it dramatically easier for researchers and hobbyists to explore ARC puzzles, human interaction traces, and alternative task formats. His work has been a major inspiration and reference point for the ARC Explainer project and the broader community.","yearStart":null,"yearEnd":null,"score":"Community datasets & visual tooling widely used ac","approach":"Open-source curation and tooling for ARC: maintains multiple repositories that package ARC variants, derived datasets, and interactive history logs in researcher-friendly formats, along with viewers and utilities for inspecting tasks and human behavior.","uniqueTechnique":"Treats ARC as a community dataset and tooling problem as much as a benchmark: systematically archives human interaction histories, alternative encodings, and visualization tools so others can build on shared resources instead of re-implementing infrastructure.","links":{"github":"https://github.com/neoneye","papers":["https://github.com/neoneye/ARC-Interactive-History-Dataset"],"website":"https://neoneye.github.io"},"teamName":null,"category":"researcher","imageUrl":"/simonS.png,/simonS1.png","rank":null,"createdAt":"2025-12-30T19:27:42.549Z"},{"id":24,"fullName":"Dr. Jeremy Budd","handle":"jez2718","affiliation":"University of Birmingham, Assistant Professor","achievement":"ARC Research Presenter - Inverse Problem Formulation","description":"Assistant Professor at the University of Birmingham conducting research on ARC-AGI from the perspective of inverse problems. Dr. Budd earned his PhD from Delft University of Technology in January 2022 (cum laude) and holds degrees from Cambridge University. His research blends applied analysis and data science, focusing on graph-based machine learning for image processing, segmentation, and reconstruction methods. He presented his ARC work to the ARC Discord community, contributing fresh theoretical perspectives to the field.","yearStart":2025,"yearEnd":null,"score":"Theoretical Research Contribution","approach":"Inverse problem formulation combined with graph-based machine learning for ARC-AGI reasoning","uniqueTechnique":"Applied analysis and data science perspective on ARC as an inverse problem","links":{"website":"https://jeremybudd.com/"},"teamName":"University of Birmingham","category":"researcher","imageUrl":"/jbudd.png","rank":null,"createdAt":"2025-12-30T19:27:42.661Z"},{"id":22,"fullName":"Eric Pang","handle":null,"affiliation":"Amazon (Machine Learning Engineer; previously Quora), University of Waterloo Math/CS","achievement":"ARC-AGI-2 SoTA: Efficient Evolutionary Program Synthesis with Grok-4 and DreamCoder-inspired library learning","description":"Developed an open-source, DreamCoder-inspired evolutionary program synthesis system that builds a reusable library of Python programs across ARC-AGI tasks, reusing concepts via accuracy-based heuristics and LLM-guided search to break the performance–cost frontier versus frontier models.","yearStart":2025,"yearEnd":2025,"score":"Outperforms frontier models on ARC-AGI-1 & ARC-AGI","approach":"Efficient evolutionary program synthesis with Grok-4 that iteratively expands a shared program library, selects high-scoring programs using accuracy heuristics and score-weighted sampling, and conditions new generations on the best existing programs.","uniqueTechnique":"DreamCoder-inspired DREAM library learning across tasks, using a single evolving program library and low test-time compute (10 LLM calls per task) to outperform frontier models on ARC-AGI-1/2 while breaking the performance–cost Pareto frontier.","links":{"website":"https://gist.github.com/inspiredlabs/a3fc232eba4b9754ad4a8234b85b8d34"},"teamName":null,"category":"researcher","imageUrl":"/ericpang.jpeg","rank":null,"createdAt":"2025-12-30T19:27:42.614Z"},{"id":4,"fullName":"Jeremy Berman","handle":"jerber888","affiliation":"Independent Researcher","achievement":"New SOTA: 79.6% on ARC v1, 29.4% on ARC v2","description":"Achieved record-breaking scores using evolutionary test-time compute with Claude Sonnet 3.5, pioneering natural language programming as an alternative to code-based approaches.","yearStart":2025,"yearEnd":null,"score":"79.6% SOTA","approach":"Evolutionary Test-Time Compute with Claude Sonnet 3.5 generating Python transform functions (v1) and plain English instructions with Grok-4 (v2). Uses up to 500 candidate functions with 31-36 dynamic prompts per challenge.","uniqueTechnique":"Pioneered using natural language as a programming medium instead of code; evolutionary approach with diversity preservation","links":{"twitter":"https://x.com/jerber888","website":"https://params.com/@jeremy-berman/arc-agi","substack":"https://jeremyberman.substack.com/"},"teamName":null,"category":"researcher","imageUrl":"/jberARC.png","rank":null,"createdAt":"2025-12-30T19:27:40.859Z"},{"id":23,"fullName":"Paul Fletcher-Hill","handle":null,"affiliation":"Independent Researcher","achievement":"Runner-Up Paper Award - Mini-ARC","description":"Achieved 41% on subset of 114 puzzles using small 67M parameter transformer models trained exclusively on ARC puzzles.","yearStart":2024,"yearEnd":2024,"score":"41% (subset)","approach":"Small transformer models with test-time training and refinement, without search, language models, or program synthesis","uniqueTechnique":"2D positional embedding scheme for grid-based reasoning with minimal parameters","links":{"papers":["https://www.paulfletcherhill.com/mini-arc.pdf"],"website":"https://www.paulfletcherhill.com/arcprize"},"teamName":null,"category":"paper_award","imageUrl":null,"rank":null,"createdAt":"2025-12-30T19:27:42.635Z"},{"id":21,"fullName":"Ryan Greenblatt","handle":null,"affiliation":"Redwood Research","achievement":"42-43% on ARC-AGI-Pub leaderboard","description":"Achieved strong results using LLM-guided program synthesis with GPT-4o generating ~8,000 Python programs per task, deterministically verified against demonstrations.","yearStart":2024,"yearEnd":2024,"score":"42-43%","approach":"LLM-guided program synthesis with specialized few-shot prompts for different problem types","uniqueTechnique":"LLM-based error identification and rectification; ARC Prize estimates 85% achievable with ~100M programs per task","links":{},"teamName":"Redwood Research","category":"researcher","imageUrl":null,"rank":null,"createdAt":"2025-12-30T19:27:42.587Z"},{"id":15,"fullName":"Jean-François Puget (2024 Paper)","handle":"JFPuget","affiliation":"Machine Learning at NVIDIA, 6x Kaggle Grandmaster","achievement":"Runner-Up ARC Prize 2024 Paper Award - A 2D nGPT Model For ARC Prize","description":"Authored A 2D nGPT Model For ARC Prize, a 2D-aware nGPT-style model tailored to ARC-AGI grids that was recognized as a runner-up paper in the ARC Prize 2024 paper awards.","yearStart":2024,"yearEnd":2024,"score":"Runner-Up (Paper Award)","approach":"2D-aware nGPT-style model that tokenizes ARC grids as 2D structures and uses an autoregressive transformer to predict output grids.","uniqueTechnique":"Emphasizes explicit 2D spatial structure in tokenization and attention so the model can reason over local and global patterns on ARC grids.","links":{"kaggle":"https://www.kaggle.com/competitions/arc-prize-2024/discussion/545844","papers":["https://github.com/jfpuget/ARC-AGI-Challenge-2024/blob/main/arc.pdf"]},"teamName":"NVIDIA","category":"paper_award","imageUrl":"/jfPuget2.png","rank":null,"createdAt":"2025-12-30T19:27:42.422Z"},{"id":14,"fullName":"Clément Bonnet","handle":null,"affiliation":"Independent Researcher","achievement":"3rd Place Paper Award - Latent Program Network","description":"Created Latent Program Network (LPN) that builds test-time search directly into neural models, searching through compact latent space without pre-defined DSLs.","yearStart":2024,"yearEnd":2024,"score":"Paper Award (3rd)","approach":"Latent Program Network combining symbolic adaptability with neural scalability","uniqueTechnique":"Test-time search through latent program space doubled performance on out-of-distribution tasks","links":{"github":"https://github.com/clement-bonnet/lpn","papers":["https://arxiv.org/abs/2411.08706"]},"teamName":null,"category":"paper_award","imageUrl":null,"rank":null,"createdAt":"2025-12-30T19:27:42.396Z"},{"id":13,"fullName":"Ekin Akyürek","handle":"akyurekekin","affiliation":"MIT CSAIL (now OpenAI)","achievement":"2nd Place Paper Award - 53.0% on public validation, 61.9% when ensembled","description":"MIT CSAIL researcher (now at OpenAI) who demonstrated that test-time training gave 6x improvement over base fine-tuned models.","yearStart":2024,"yearEnd":2024,"score":"53.0% (61.9% ensemble)","approach":"Test-time training (TTT) temporarily updating model parameters during inference","uniqueTechnique":"Demonstrated TTT as mechanism for improving LLMs' reasoning capabilities, also improving BIG-Bench Hard from 50.5% to 57.8%","links":{"github":"https://github.com/ekinakyurek/marc","papers":["https://arxiv.org/abs/2411.07279"],"twitter":"https://x.com/akyurekekin","website":"https://ekinakyurek.github.io/"},"teamName":"MIT CSAIL","category":"paper_award","imageUrl":null,"rank":null,"createdAt":"2025-12-30T19:27:42.370Z"},{"id":12,"fullName":"Kevin Ellis","handle":"ellisk_kellis","affiliation":"Cornell University (Associate Professor)","achievement":"1st Place Paper Award (co-author), creator of DreamCoder","description":"Associate Professor at Cornell and creator of DreamCoder program synthesis system, combining program synthesis with inductive reasoning.","yearStart":2024,"yearEnd":2024,"score":"Paper Award (1st)","approach":"Program synthesis and library learning, combining induction and transduction","uniqueTechnique":"DreamCoder system for learning program libraries through wake-sleep algorithm","links":{"twitter":"https://x.com/ellisk_kellis","website":"https://www.cs.cornell.edu/~ellisk/"},"teamName":"MIT & Cornell","category":"paper_award","imageUrl":null,"rank":null,"createdAt":"2025-12-30T19:27:42.344Z"},{"id":11,"fullName":"Wen-Ding Li","handle":null,"affiliation":"Cornell University (PhD student with Kevin Ellis)","achievement":"1st Place Paper Award - Combined team achieved 47.5% on public leaderboard","description":"Cornell PhD student who investigated whether it's better to infer latent functions (induction) or directly predict outputs (transduction) for ARC tasks.","yearStart":2024,"yearEnd":2024,"score":"47.5% (Paper Award)","approach":"Combined induction and transduction using neural models trained on synthetic variations of Python programs","uniqueTechnique":"Discovered that induction and transduction excel at different types of ARC tasks despite same training data and architecture","links":{"papers":["https://www.cs.cornell.edu/~wdli/"],"website":"https://wending.dev/"},"teamName":"MIT & Cornell","category":"paper_award","imageUrl":null,"rank":null,"createdAt":"2025-12-30T19:27:42.317Z"},{"id":16,"fullName":"MindsAI Team","handle":null,"affiliation":"MindsAI Research Lab","achievement":"Highest score of 55.5% on private evaluation set (ineligible for prize)","description":"Pioneered test-time training (TTT) for ARC-AGI beginning in 2023, inspiring many subsequent approaches. Chose not to open source solution.","yearStart":2023,"yearEnd":2024,"score":"55.5% (Highest)","approach":"Test-time training using Salesforce T5 series model pretrained on public eval set and synthetic data, further fine-tuned at test time","uniqueTechnique":"First team to successfully apply TTT to ARC-AGI (2023), increased SOTA from 33% to 55.5%","links":{},"teamName":"MindsAI","category":"researcher","imageUrl":null,"rank":null,"createdAt":"2025-12-30T19:27:42.443Z"},{"id":18,"fullName":"Michael Hodel","handle":null,"affiliation":"ETH Zurich (Master's student in Computer Science)","achievement":"Winner ARCathon 2022, set world record of 39% in 2024","description":"Created one of the best ARC-AGI domain-specific languages (DSLs) to date, won ARCathon 2022 and continued improving ARC solutions.","yearStart":2022,"yearEnd":2024,"score":"39% (Record 2024)","approach":"Domain-specific language design for ARC puzzle solving","uniqueTechnique":"Optimized DSL and program search process for better performance","links":{},"teamName":null,"category":"pioneer","imageUrl":null,"rank":null,"createdAt":"2025-12-30T19:27:42.493Z"},{"id":17,"fullName":"icecuber","handle":"icecuber","affiliation":"Independent","achievement":"1st Place Kaggle ARC 2020 - 20% (solved 20 of 100 tasks)","description":"First significant success on ARC, establishing DSL-based program search as foundation for subsequent solutions. Showed at least 20% of tasks solvable with just 3-4 transformations.","yearStart":2020,"yearEnd":2020,"score":"20% (1st Place)","approach":"Brute-force program search with DSL containing 142 hand-crafted unary functions on grids; greedy composition stored in directed acyclic graph (DAG)","uniqueTechnique":"Limited search to depth 3-4, demonstrating efficiency of shallow program composition. Written in C++17 and Python.","links":{"github":"https://github.com/top-quarks/ARC-solution","kaggle":"https://www.kaggle.com/competitions/abstraction-and-reasoning-challenge/discussion/154597"},"teamName":null,"category":"pioneer","imageUrl":null,"rank":null,"createdAt":"2025-12-30T19:27:42.465Z"}],"total":29}