ARC-AGI Community Leaderboard
ARC-AGI has gained significant popularity over the past two years, and we've been overwhelmed by the number of researchers and builders who want to showcase their work to the community. The ARC-AGI Community Leaderboard provides a landing spot for these submissions, where the community can review, discuss, and verify results together.
Community Leaderboard submissions must be general purpose and reproducible. Scores are self-reported unless noted otherwise. Results on the ARC-AGI-1 and ARC-AGI-2 semi-private sets are run and verified by ARC Prize. Everything else is scored on a public set and self-reported. We aren't able to verify the authenticity of self-reported scores and won't independently verify submissions except in extraordinary cases, so we encourage the community to explore each submission and validate the results for themselves. We reserve the right to determine what qualifies. For more on how we approach testing, see our testing policy.
To submit your work, head to the ARC-AGI Community Leaderboard repo on GitHub.
| Name | Authors | Benchmark | Score | Cost | Links | |
|---|---|---|---|---|---|---|
Tycho A multimodal model maintains one growing conversation per game and can delegate a falsification-tested executable world model to a builder for planning. | Jens Lehmann, Andrei Aioanei, Sahar Vahdati | ARC-AGI-3Public Demo | 100.0% | $2,986 | 2026-07-29 | |
Retrodict A general LLM agent that logs every frame and requires rule hypotheses to retrodict recorded history before spending live actions. | Ryan Brown | ARC-AGI-3Public Demo | 99.9% | $654 | 2026-07-19 | |
baseline1 Coding agent that builds and verifies an executable Python world model, then plans through it. | Sergey Rodionov | ARC-AGI-3Public Demo | 99.0% | $400 | 2026-07-15 | |
NOOA (NVIDIA-labs OO Agents) A general CodeAct agent that builds reusable NumPy world-model helpers and persists learning through memory or Markdown knowledge files. | Gal Kaplun, Elad Sarafian, Paul Furgale, Ricardo Silveira Cabral, Ron Banner | ARC-AGI-3Public Demo | 85.1% | $332 | 2026-07-09 | |
Polyphony Agent - ARC A coding agent grows a verified per-game Heuristic System for state, dynamics, planning, and action selection as readable Python files. | Ruiyang Yu, Anyang Su, Chenxu Zhao, Tianyu Fu, Shuo Wang, Minghui Wu | ARC-AGI-3Public Demo | 19.8% | $115 | 2026-07-07 | |
OPINE-World Two agents iterate in a CEGIS-style loop: one acts while the other rewrites and adversarially tests an executable game engine used for planning. | David Courtis, Wenhao Li, Scott Sanner, D3M Lab | ARC-AGI-3Public Demo | 78.4% | $1,040 | 2026-07-01 | |
ARM (Abstraction Reasoning Model) A Tiny Recursive Model variant whose H-level replaces spatial compression with 32 learned Perceiver/Set-Transformer-style latent tokens. | Marcel Siegert | ARC-AGI-1Public Train | 40.3% | $3 | 2026-06-23 | |
Continual Harness A self-improving ARC-AGI-3 orchestrator whose Refiner rewrites its base policy, memory, skills, and subagents from trajectory evidence. | Ruirong Feng, Seth Karten, Wenzhe Li, Chengshuai Shi, Joel Zhang, Tersoo Upaa Jr, Kiran Vodrahalli, Chi Jin | ARC-AGI-3Public Demo | 20.5% | $774 | 2026-06-18 | |
Vision - Continual Learning v1 Multimodal agent with continual-learning weights carried across games and levels. | Vansh | ARC-AGI-3Public Demo | 63.1% | $4,788 | 2026-05-18 | |
OpenClaw OpenClaw Harness adapted to play ARC-AGI-3 allowed memory and code execution tools. | ARC Prize Foundation | ARC-AGI-3Public Demo | 5.2% | $2,912 | 2026-05-15 | |
DreamTeam Six fixed agent roles coordinate through a shared file workspace and an executable world model that they build and revise at run time. | Elad Sarafian, Gal Kaplun, Ron Banner, Daniel Soudry, Boris Ginsburg | ARC-AGI-3Public Demo | 38.1% | $18,000 | 2026-05-04 | |
Human Intelligence Harness Maximum human intelligence built into an agent harness. | ARC Prize Foundation | ARC-AGI-3Public Demo | 95.3% | - | 2026-04-14 | |
Noemon's agentic ARC-AGI 2 Solver (Gemini 3.1 Pro) A Reasoner → Validator → Refine loop that learns an instruction set with persistent thought history, self-consistency, and a judge for candidate selection. | Noemon | ARC-AGI-2Public Train | 92.5% | $4 | 2026-04-09 | |
TELL Single-conversation agent that compounds confirmed knowledge in a MEMORY.md file. | Dots Post-train Team | ARC-AGI-3Public Demo | 43.9% | $1,406 | 2026-04-09 | |
a-evolve MAS Evolved Evolved multi-agent orchestrator with 9 learned skills mined from competition logs. | Zhan Shi, Hanqing Lu, Bing He, Yisi Sang, Minhua Lin | ARC-AGI-3Public Demo | 12.3% | $5,300 | 2026-04-09 | |
Read-Grep-Bash Agent A coding agent that uses search and Python scripting over game logs. | Alexis Fox, Junlin Wang, Paul Rosu, Bhuwan Dhingra | ARC-AGI-3Public Demo | 50.2% | - | 2026-03-13 | |
Evolutionary Test-Time Compute with Natural Language Instructions Evolves natural language instructions instead of code. | Jeremy Berman | ARC-AGI-2Semi-Private | 29.4% | $3,648 | 2025-09-16 | |
Efficient Evolutionary Program Synthesis Evolves a growing library of Python programs with an LLM. | Eric Pang | ARC-AGI-2Semi-Private | 26.0% | $476 | 2025-09-01 | |
Tiny Recursive Model (TRM) 7M parameter recursive model with think-act refinement loops. | Alexia Jolicoeur-Martineau | ARC-AGI-2Public Train | 7.8% | $252 | 2025-07-01 | |
Hierarchical Reasoning Model (HRM) Brain-inspired 27M parameter model with iterative refinement. | Sapient Intelligence | ARC-AGI-2Semi-Private | 2.0% | $201 | 2025-06-08 | |
Evolutionary Test-time Compute Genetic algorithm over LLM-generated Python transforms. | Jeremy Berman | ARC-AGI-1Semi-Private | 53.6% | $2,900 | 2024-12-18 | |
Ryan Greenblatt LLM generates and refines thousands of candidate programs per task. | Ryan Greenblatt | ARC-AGI-1Semi-Private | 43.0% | $40,000 | 2024-06-17 |