Part Ⅳ · The chronological spine
Seven years of ARC-AGI.
ARC-AGI began in 2019 with a new argument about how to measure intelligence. After five years without a decisive machine breakthrough, new methods broke through in 2024—only for ARC-AGI-2 and ARC-AGI-3 to raise the bar again. Every milestone links to its source in Part Ⅱ.
2019 · The measure
A benchmark built to measure the thing scale hides.
Nov 2019
On the Measure of Intelligencepaper
François Chollet defines intelligence as skill-acquisition efficiency, argues that a fair test must control for prior knowledge, and introduces ARC: a few-shot, Core-Knowledge benchmark where every task is new. This becomes the reference point for later work.
№01 · full entry in Part Ⅱ →2020 · The template
A solo searcher wins, and sets a pattern that holds for three years.
2020
The first Kaggle competition — Icecuber winscompetition
A solo entrant wins with a hand-built DSL of ~142 grid operations, composed through exhaustive search and filtered by an exact-match verifier. With no learning, the DSL + search + verifier template leads for three years while large models stay below 5%.
№07 · full entry in Part Ⅱ →2020 solution write-ups ↗2024 · The break
Test-time adaptation, then o3 — the record moves at last.
Jun 2024
50% with GPT-4o — sampling as an axismilestone
Ryan Greenblatt samples thousands of GPT-4o programs per task, keeps those that replay every demonstration pair, and votes. The result shows that sample count against a verifier is a tunable axis of accuracy and anticipates the refinement loop.
№10 · full entry in Part Ⅱ →Nov 2024
Test-time training formalizedpaper
MIT shows that fitting a fresh adapter to each task's demonstrations can produce a roughly 6× gain over a fine-tuned baseline and reach average-human level when combined with program synthesis. Per-task adaptation defines the 2024 season.
№08 · full entry in Part Ⅱ →2024
The ARChitects win the ARC Prizecompetition
Test-time training on an open 8B model plus a product-of-experts over task symmetries takes the 2024 Kaggle competition — entirely offline, at about two cents per task. The winners report records 53.5% for the prize winner and 55.5% for an ineligible non-open submission.
№09 · full entry in Part Ⅱ →Official 2024 results ↗Dec 2024
o3 clears 87.5% — the step-functionmilestone
OpenAI's o3, searching over chains-of-thought with a learned evaluator, posts the first human-competitive score on ARC-AGI-1 — at tens to thousands of dollars a task. It is a proprietary benchmark record, not an open-source prize win, and it forces a redesign.
№06 · full entry in Part Ⅱ →Official o3 analysis ↗2025 · The reset
A harder, cost-scored benchmark, and the first interactive preview.
2025
ARC-AGI-2 reopens the gappaper
A harder benchmark adds compositional tasks with multiple interacting rules and in-context symbols, plus cost per task as a scored axis. Launch-era frontier-model scores fell to the low single digits, reopening the human–machine gap.
№02 · full entry in Part Ⅱ →ARC-AGI-2 report ↗Jul–Aug 2025
The ARC-AGI-3 Preview Challengecompetition
A 30-day preview of the interactive benchmark. StochasticGoose (12.58%) predicts which actions change the frame; Blind Squirrel (6.71%) builds a state graph. The results establish informed exploration as the first useful baseline.
Official preview learnings ↗2025
The refinement loop — ARC Prize 2025 reportpaper
The official account of the first ARC-AGI-2 season names the year's theme: propose, verify against the demos, feed back, iterate. NVARC wins at 24.03%; the 85% grand prize goes unclaimed; cheap-and-small beats costly-and-large.
№05 · full entry in Part Ⅱ →Official 2025 results ↗2025
ARC without pretraining — CompressARCpaper
A different approach solves each task by minimizing description length, with no pretraining or external program search. Its ~20–34% result tests the claim that MDL may underlie what search-and-adapt methods approximate.
№16 · full entry in Part Ⅱ →2026 · The frontier
The interactive benchmark lands, essentially unsolved.
Dec 2025
Graph-based exploration takes 3rdpaper
The first published ARC-AGI-3 agent is a training-free explorer that keeps a directed graph of hashed frame-states and heads toward the nearest untested action. It uses more of the interaction budget than model-controlled agents, but explores without inducing a transition model.
№17 · full entry in Part Ⅱ →2026
ARC-AGI-3 — the interactive benchmarkpaper
The first format change since 2019: an agent enters a turn-based game with no instructions or stated goal, using only the controls currently available. RHAE scores action efficiency; humans solve every included environment, while the report's harness-free launch baselines sit below 1%.
№03 · full entry in Part Ⅱ →ARC-AGI-3 technical report ↗2026
ARC Prize 2026 — $2M on Kagglecompetition
Two tracks run on Kaggle — ARC-AGI-3 and a final ARC-AGI-2 track — with a $2M pool. The ARC-AGI-3 season publishes two paid milestones before the final awards.
Official 2026 schedule ↗Jul 2026
The Duck wins ARC-AGI-3 Milestone #1milestone
Tufa Labs' small local coding agent takes first, with Reki second and forge third. The podium shifts from preview-era graph search toward vision-language policies with memory, structured actions, and—in The Duck—a live Python REPL. The final competition remains open.
Official milestone report ↗The shape of it
Every time the machines catch up, the bar moves.
Across the three competitions, the pattern repeats: a method closes the gap, then a redesigned benchmark opens a harder one. Icecuber's searchSolo, Icecuber's hand-built DSL of ~142 grid operations — greedily composed into a DAG, no learning at all — cleared ~20% of the private set; pooling every 2020 entry into one ensemble reached only ~49%. That was the ceiling of pure search. yielded to test-time trainingMIT's recipe fits a fresh LoRA adapter to each task's own demonstrations at inference — up to a 6× accuracy lever, and ~62% once ensembled with program synthesis, matching average-human performance., then o3o3 scored 75.7% at high efficiency and 87.5% using ~172× the compute — at roughly $20 to thousands of dollars a task, against ~$5 for a human. Chollet's own verdict: not AGI.. o3 prompted ARC-AGI-2; ARC-AGI-2 led to the interactive ARC-AGI-3.
Figure · The climb and the resets
The frontier score across all three generations
Each reset drops the best-known score back toward zero against a near-100% human baseline.
Five eras, one idea
Hand-built search
Icecuber
A solo DSL + brute-force search wins the first Kaggle. No learning; the template holds for 3 years.
Test-time adaptation
MindsAI · MIT · ARChitects
Per-task fine-tuning and augmentation-ensembling drive the private record from 33% to 55.5%.
The o3 break
OpenAI
Guided search over chains-of-thought clears 87.5% — near-human, but at thousands of dollars a task.
The refinement loop
ARC-AGI-2 field
Propose → verify → feed back → iterate. Cheap small models beat costly big ones. Top score: 24%.
Agentic exploration
ARC-AGI-3 preview + 2026 milestones
Interactive worlds. Exploration progresses from state graphs to multimodal policies, memory, and executable tools; transferable world modeling and goal induction remain open.
The techniques change, but the recurring pattern is adaptation followed by verification. ARC-AGI-3 removes the clean verifier: there are no demonstration pairs or stated goals, only a world to probe.
→ Who won each round, in Part Ⅸ · Winning Solutions