Competition award
A final placement and cash award under that season's private evaluation, compute, openness, and eligibility rules.
The story beneath the leaderboard
The search did not disappear. It changed address.
Every strong ARC system closes the same loop: choose a way to see the problem, propose a change, test it against reality, keep what worked, and spend the next unit of compute or action more intelligently.
The core model
The history is not symbolic methods being replaced by neural methods and then by agents. It is a sequence of relocations. Search moves from hand-written programs into latent spaces, model weights, populations of candidate programs, prompt context, persistent memories, and finally actions inside an unknown world. Each era keeps the useful machinery of the previous one while moving the bottleneck.
In static ARC, the demonstration pairs are an unusually strong oracle: a candidate transformation can be replayed exactly and rejected immediately. Icecuber exploited that oracle with a hand-built language. The 2024 systems kept exact verification but learned better proposal distributions and adapted them to each task.
ARC-AGI-2 made the rules more compositional, so a single proposal was rarely enough. The characteristic 2025 move was refinement: generate, execute, diagnose the mismatch, and feed the failure back into the next proposal. Search stopped being a one-shot ranking problem and became a conversation with a verifier.
ARC-AGI-3 removes even the demonstration oracle. The agent must create evidence by acting, decide which observation is worth buying, infer what state changes mean, and remember discoveries without drowning in history. The problem has therefore moved from solving a supplied specification to learning the specification itself.
Make the grid searchable
Icecuber and the symbolic lineageReplace an open-ended visual puzzle with a finite language of objects, transforms, and compositions.
Kept: Exact replay against demonstrations.
New bottleneck: The hand-written grammar decides what can ever be discovered.
Make the solver task-local
the ARChitects, MIT TTT, MindsAIAdapt a learned model to the current task instead of asking one frozen model to solve every task.
Kept: Augmentation, candidate verification, and ensembling.
New bottleneck: A few demonstrations can support several plausible rules; local fitting can choose the wrong one.
Make failure productive
NVARC and refinement-loop systemsTurn a failed candidate into information for the next candidate rather than simply discarding it.
Kept: Task-time adaptation and exact execution feedback.
New bottleneck: More refinement can still polish the wrong abstraction, and cost rises with every loop.
Make exploration deliberate
StochasticGoose, Blind Squirrel, Just-ExploreChoose actions for information value, record visited states, and avoid spending the action budget on repeats.
Kept: Search, but now the search tree is generated by interaction.
New bottleneck: Novelty detection is not yet a transferable model of dynamics or a theory of the goal.
Make the agent a tool-using investigator
The Duck, Reki, forgeLet a multimodal model inspect several representations, write code or structured actions, reflect, and manage its own working memory.
Kept: Explicit state, verification, repair, and iterative search.
New bottleneck: The harness is competent at local play; robust transfer to unseen dynamics and goals remains unproved.
Conceptual figure · the common machine
Six verbs hiding inside almost every winning system
The implementation changes radically; the epistemic loop does not. Static solvers receive evidence as demonstration pairs. Interactive agents must manufacture evidence through action.
-
1
See
What is the state made of?
Static ARCObjects, colors, regions, symmetries, or learned features.
Interactive ARCRendered frames, raw grids, segments, events, and action history.
-
2
Hypothesize
What rule might explain it?
Static ARCA DSL program, Python transform, latent program, or task-tuned prediction.
Interactive ARCA claim about controls, dynamics, hazards, progress, or the hidden goal.
-
3
Choose
Which candidate or action is worth trying?
Static ARCEnumerate, sample, mutate, ensemble, or rank.
Interactive ARCPrefer information-gaining, legal, non-repeating, goal-directed actions.
-
4
Test
What did reality say?
Static ARCReplay every demonstration pair and measure exact mismatch.
Interactive ARCObserve the next frame, reward, death, novelty, or lack of change.
-
5
Retain
What should survive the attempt?
Static ARCA surviving program, tuned weights, a population archive, or a vote.
Interactive ARCA state graph, reflection note, compact rule, reusable tool, or revised plan.
-
6
Transfer
What can be reused without overfitting?
Static ARCPrimitives, augmentations, proposal models, and refinement operators.
Interactive ARCA general way to learn dynamics and infer goals—not a script for one seen game.
Atlas synthesis from the official competition reports and the ARC-AGI-3 Technical Report. This is a comparative model, not an organizer-defined taxonomy.