Part Ⅶ · The interactive frontier NEW

ARC-AGI-3 don't solve the puzzle — learn the game.

The first interactive benchmark in the ARC series (preview 2025; full launch 2026). Instead of predicting a static grid, an agent enters an abstract, turn-based game with no instructions or stated goal—only a screen and the controls currently available. It must explore, model the world, infer what winning means, and plan. Milestone Prize #1 has three winning open-source submissions; the final 2026 competition is still running.

100%

Humans solve

< 1%

Launch-era AI (harness-free)

M1 · 1st

Latest prize checkpoint

135

Environments

Animated official capture of a human playing the LS20 ARC-AGI-3 environment — See the benchmark first

No instructions. Learn by acting.

Make a deliberate move, observe exactly what changed, and treat every action as an experiment. Discovering the controls, rules, and goal is the test—and every extra action costs score.

**Probe one control.** Try an arrow, Space, or a click.

**Model the response.** Track every change in the world.

**Infer and pursue the goal.** Reuse mechanics and minimize actions.

Play LS20 ↗ Explore public games ↗

Official LS20 human-play capture and game: ARC Prize Foundation.

What changed

Not a puzzle. A world.

ARC-AGI-1 and ARC-AGI-2 hand you a few input→output grids and ask for the rule. ARC-AGI-3 hands you nothing — just a screen and a few keys — and asks you to figure out the game by playing it.

Each turn the agent sees a 64×64 grid of 16 colors and submits one action from a tiny set (five keys + Undo + click-a-cell). The world is turn-based and never moves on its own, so the whole challenge is reasoning, not reflexes. Difficulty grows by composition of mechanics you had to discover for yourself. This is the first format change since 2019.

And it is scored on efficiency, not just success. For each completed level, RHAE (Relative Human Action Efficiency) squares the human-to-agent action ratio, caps that level term at 1.15, then applies level weighting and benchmark aggregation — a direct, first-contact human-vs-AI learning-efficiency comparison.

Figure · Relative Human Action Efficiency

The price of an extra action

At twice a human’s actions the efficiency term falls to 25%; at five times, to 4%.

Question answered: How sharply does ARC-AGI-3 punish actions beyond the human baseline?

This curve includes the official 1.15 per-level cap. The per-game completion cap, level weighting, and benchmark aggregation are separate parts of the final score.

Agent actions relative to a human

2× human actions → 25% efficiency

level efficiency = min(1.15, (human actions / agent actions)²)

Anchor values for the curve
Agent actions	0.75×	1×	2×	3×	4×	5×
Efficiency term	115%	100%	25%	11.1%	6.25%	4%

Source: ARC-AGI-3 scoring methodology. Retrieved 2026-07-28.

The direct literature

Papers about ARC-AGI-3 itself.

Only work directly definingThe reference paper is an ARC Prize Foundation preprint dated April 2026 — benchmark design by François Chollet, with co-founders Mike Knoop and Gregory Kamradt and lead environment designer Hunter Henry. or solving ARC-AGI-3 appears here. Broader ARC theory and adjacent methods stay in Part Ⅱ; competition implementations without papers stay in the solutions section below.

Category 1

Reference papers

The benchmark definition, task format, scoring, validation, and human baseline.

1 paper

№03 paper 2026

ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence

Chollet et al. · arXiv · 2026

AgenticBenchmark

The first interactive ARC benchmark: an agent explores an instruction-free game and is scored on action efficiency; launch-era harness-free AI scored below 1%

Its job in the ARC-AGI-3 storyThe problem definition itself: the first interactive, instruction-free ARC benchmark, where an agent dropped into a novel turn-based game must explore, model dynamics, infer the goal, and plan across four pillars, scored by action efficiency relative to a human baseline (RHAE). At launch, calibrated humans solved every environment while harness-free frontier AI scored below 1%.

100%Human baseline

<1%Launch-era AI (harness-free)

12.58%Best preview agent

~135Environments

Read the original ↗Full entry in Part Ⅱ →

Category 2

Solution papers

The smallest set of direct, technically substantive attempts to solve the interactive benchmark.

5 papers

№17 paper 2025

Graph-Based Exploration for ARC-AGI-3 Interactive Reasoning Tasks

Rudakov et al. · arXiv · 2025

Agentic

A training-free graph explorer records every visited state and places 3rd on ARC-AGI-3 by using far more of the interaction budget than model-controlled agents

Its job in the ARC-AGI-3 storyThe working ARC-AGI-3 agent on the shelf: a training-free explorer that keeps an explicit directed graph of hashed frame-states with salience-ranked action edges, reaching 3rd on the Preview Challenge private leaderboard. It is the proof-point that cheap deterministic exploration beats LLM agents (interaction-limited to ~4,000 usable steps vs. ~96,000 allowed) — while deliberately stopping at exploration and never inducing a model of the transitions it records.

3rd placeARC-AGI-3 Preview (private)

≈30/52Levels solved (median, 8h)

12 private levelsOfficial submission

~4k / 96k stepsInteraction budget used

Read the original ↗Full entry in Part Ⅱ →

S02 paper 2026

Hill-Climbing ARC-AGI-3

Fox, Wang, Rosu & Dhingra · ARC-AGI-3 academic partner paper · 2026

AgenticLLM harness

A Duke agentic harness gives Claude Opus 4.6 Python-based access to selectively retrieve and transform its long action history, solving all three public environments near human action counts.

Its job in the ARC-AGI-3 storyA key LLM-harness result. Instead of placing every 64×64 frame in a rolling prompt, the model executes Python to query and compress its recorded history before acting. The harness reaches near-human efficiency on seen public environments but does not transfer, making it both a strong result and a warning about overfitting.

3/3Public environments

Opus 4.6Model evaluated

0.0%TR87 without harness

97.1%TR87 with harness

Read the benchmark discussion ↗

The official benchmark report cites the standalone paper, but this guide does not have a separate public URL for it.

№20 paper 2026

Executable World Models for ARC-AGI-3 in the Era of Coding Agents

Sergey Rodionov · arXiv · AGI-2026

AgenticSynthesis

A coding agent writes a persistent Python simulator, checks it against every observed transition, simplifies it, and plans through it before acting

Its job in the ARC-AGI-3 storyThe first direct executable-world-model system in this corpus: a coding agent maintains a Python simulator, replays observations against it, simplifies the program, and plans through it. Its released public-run artifacts make the model itself inspectable, while private validation remains open.

15 / 25Public games solved

58.12%GPT-5.5 high RHAE

UntestedPrivate validation

Read the original ↗Full entry in Part Ⅱ →

№21 paper 2026

OPINE-World: Programmatic World Modeling with Ontology-error-Prioritized Interactive Exploration for ARC-AGI-3

Courtis et al. · arXiv · 2026

AgenticSynthesis

Two cooperating agents learn an object-centric program online and choose experiments that target uncertainty about the world's object vocabulary

Its job in the ARC-AGI-3 storyAn object-centric extension that treats representation discovery as part of exploration. Two cooperating agents synthesize and test a program, while ontology error directs actions toward uncertainty about the current object vocabulary.

20 / 25Public games solved

78.4Action efficiency

NonePer-game training

Read the original ↗Full entry in Part Ⅱ →

№22 paper 2026

Do Coding Agents Need Executable World Models, Simplification, and Verification to Solve ARC-AGI-3?

Sergey Rodionov · arXiv · 2026

AgenticCritique

A four-treatment ablation asks which parts of executable world modeling matter—and finds verification most reliable, simplification useful, and model strength dominant

Its job in the ARC-AGI-3 storyThe component-evidence paper: four nested agents isolate executable persistence, simplification, and exact replay verification. It finds that verification is the most consistent treatment, while stronger models and greater reasoning effort remain the most robust levers.

4Main treatments

1st in 4 / 4Verification rank

Not yetHeld-out evidence

Read the original ↗Full entry in Part Ⅱ →

What's been tried · updated July 2026

The main reported solutions so far.

The list now begins with the first 2026 milestone podium—The Duck, Reki, and forge—then preserves the 2025 preview and pre-launch harness workThe Preview's first two were StochasticGoose (Tufa Labs) at 12.58%, a CNN+RL action-effect predictor, and Blind Squirrel at 6.71%, a directed state graph. The 2026 milestone instead favored small local vision-language models with memory, structured actions, and—in The Duck's case—a live Python REPL.. These are different checkpoints, so placements are shown before incompatible percentages are compared.

The DuckTufa LabsLLM + harness 1st · 2026 Milestone Prize #1

Runs Qwen 3.6 27B FP8 locally as an agent that can inspect rendered images, raw grid arrays, and segmented regions, then write and execute Python in a live REPL before choosing an action.

Why it worksMultimodal inspection, executable scratch work, and continuous eviction of old messages let a small open model keep playing without exhausting its context. Its light, generic harness won the first $37.5k checkpoint.

Where it stopsIt is a milestone winner, not the final 2026 champion. The public report does not show that its improvised code becomes a compact dynamics model that transfers across hidden environments.

RekiLLM agent 2nd · 2026 Milestone Prize #1

Uses Gemma-4-31B as a vision policy: recent frames become labeled images, the model returns structured JSON observations and a short 1–4 action queue, and a reflection memory is refreshed about every ten steps.

Why it worksLegal-action guards, JSON repair, persistent reflection, rare-color click preference, and a learned dead-signature make exploration less wasteful while keeping every feature ablatable.

Where it stopsIts rare-color and button-shape preferences are useful priors, but they are still hand-chosen exploration biases; the final competition must test whether they support broad hidden-environment transfer.

forgeMd Boktiar Mahbub MuradLLM agent 3rd · 2026 Milestone Prize #1

Packages a Gemma-4-31B vision-policy loop as a configurable framework with rendered histories, reflection memory, structured JSON actions, repair, legal-action guards, and optional generator/arbiter machinery.

Why it worksThe framework makes alternate models, prompts, memory strategies, and action policies easy to swap and test; the submitted profile reached the milestone podium with the simpler policy path.

Where it stopsMuch of the advertised generator and arbiter machinery was disabled in the winning profile, so the placement supports the lean vision-policy loop more strongly than the whole framework.

StochasticGooseTufa LabsLearned policy 12.58% · 1st, preview

Trains a CNN+RL model to predict which actions actually change the frame, then runs informed search over the tiny action space guided by those predictions.

Why it worksLearning which actions are state-changing focuses search far better than random play or per-step LLM control, taking the top ARC-AGI-3 Preview competition score.

Where it stopsIt searches the action space rather than building a genuine predictive world model, so it never infers dynamics or the hidden goal.

Blind SquirrelExploration · state graph 6.71% · 2nd, preview

Builds a directed state graph from observed frames and searches over it to cover the environment systematically.

Why it worksAn explicit graph of observed states remembers where it has been and drives methodical coverage, reaching 2nd in the preview competition.

Where it stopsRoughly half StochasticGoose's score; it is informed search over observed states, not modeling of the underlying dynamics.

Just-Explore (graph-based explorer)Rudakov, Shock & CowleyExploration · state graph 3.64% · 3rd, preview · 12 private levels

A training-free algorithmic explorer: a connected-component vision front-end with five salience tiers feeds a directed frame-hash state graph, over which a frontier-driven policy always heads toward the nearest untested action.

Why it worksBecause it never calls a model per step it spends the full ~96,000-step budget productively — where LLM agents stall around ~4,000 interactions — and beats LLM baselines to 3rd on the private leaderboard.

Where it stopsIt explores but never induces a rule for the transitions; state-space blowup caps hard levels, and a reset-marking bug depressed its official score below its demonstrated median.

Duke LRM harnessDuke UniversityLLM + harness all 3 public envs, ~human action counts · no transfer

A harness around a large reasoning model that executes Python to selectively retrieve and transform the recorded action history before each decision.

Why it worksTreating interaction history as a queryable, compressible database beats a naive rolling context window and reaches human-comparable efficiency on the public environments.

Where it stopsThe gains are on seen public environments only and do not transfer to unseen private ones — seen-environment performance is explicitly not AGI progress.

ArcgenticaSymbolicaLLM + harness all 3 public envs, ~human action counts · no transfer

An orchestrator-plus-subagent harness that decomposes play across coordinated LLM roles.

Why it worksThe orchestrator/subagent decomposition solves all three public environments at roughly human action counts.

Where it stopsLike all hand-built harnesses it overfits seen environments — bimodality as extreme as 0% harness-free vs ~97% hand-built on one env — and fails to generalize to private ones.

Frontier LLMs (harness-free)major frontier labsLLM agent ~0.1–0.5% · <1% at launch

The strongest general-purpose models served behind their own APIs, driven only by a fixed tool-free system prompt ('you are playing a game; reply with the exact action'), with no ARC-specific scaffolding.

Why it worksThis is the honest developer-aware-generalization baseline the official leaderboard reports — no per-environment engineering hiding the model's true ability.

Where it stopsIt sits at the sub-1% floor versus 100% human, failing at exploration, hypothesis revision, goal inference, and efficient planning under uncertainty.

The solution space

Exploration got smarter. Transfer is still the test.

The milestone agents remember, parse, plan locally, and sometimes write code; the preview agents search cheaply. What remains unproven is a compact, predictive model of hidden dynamicsThe training-free preview explorer is the clean negative case: it records every (state, action) → next_state transition but never induces a rule that explains them. The Duck can improvise Python, but the milestone report does not establish that this becomes a reusable world model across hidden games., an inferred goal, and a learning strategy that transfers to the next unseen environment.

Dominant today

Local vision-language policy + memory + tools

A small local multimodal model sees recent rendered frames, emits structured actions, and carries explicit memory. The Duck adds a live Python REPL, multiple board representations, and context eviction; Reki and forge add reflection, repair, legal-action guards, and short action queues.

e.g. The Duck, Reki, and forge — the top three in ARC Prize 2026 Milestone Prize #1

Tried

Graph-based state exploration (training-free algorithmic search)

A training-free explorer keeps a directed graph of hashed frame-states with tried-action edges, driven by a frontier-coverage policy (untested-local action -> shortest-path to nearest untested state -> relax salience threshold). A connected-component vision front-end with status-bar masking and five visual-salience tiers tames the 64x64 = 4,096-cell click space. It records transitions but never induces a rule for them.

e.g. Rudakov et al. training-free graph explorer (3rd, Preview Challenge); 'Blind Squirrel' directed state graph (2nd, ~6.71%)

Tried

Neural action-effect / RL exploration

A trained CNN+RL model predicts which actions actually change the frame and steers exploration toward state-changing moves. It is still informed search over the action space rather than a reusable model of the dynamics or the goal.

e.g. 'StochasticGoose' (Tufa Labs), 1st in the Jul-Aug 2025 Preview Challenge at ~12.58%

Tried

Frontier LLM as per-step controller (harness-free)

A general-purpose model behind its own API picks one action per turn from a fixed tool-free prompt ('you are playing a game, reply with the action'), with no ARC-specific scaffolding. In the launch evaluation, this official mode was interaction-limited to about 4,000 of the roughly 96,000 allowed steps because every step required a model call.

e.g. Launch-era semi-private harness-free leaderboard: strongest frontier labs at ~0.5% / 0.4% / 0.2% / 0.1%

Tried

LLM + hand-built harness / scaffolding

Wrap a frontier model in per-environment scaffolding: execute Python to selectively retrieve and compress action history, or an orchestrator-subagent design. It solves all three public environments at roughly human-comparable action counts, but overfits to seen environments and does not transfer to the private sets.

e.g. Duke University LRM harness; Symbolica 'Arcgentica' orchestrator-subagent (both solved all 3 public envs; 0% harness-free vs 97.1% hand-built on one env)

Open gap

World-model program synthesis + planning (the main gap)

Induce an explicit program next_frame = f(frame, action) from recorded interaction traces, fit it incrementally, accept it only when it exactly replays every observed transition, then plan lookahead inside it with a verifier-in-the-loop that rejects the model on the first mispredicted transition. This is the Modeling pillar operationalized as verifiable code and is essentially untried on ARC-AGI-3.

e.g. No published ARC-AGI-3 system using this approach yet; the closest analogue is the 2025 refinement loop: propose, verify, and iterate

Open gap

Goal / reward induction from delayed feedback

Infer 'what winning means' with no instructions: a hypothesis search over candidate reward functions or desirable states, driven by intrinsic motivation / empowerment and revised from sparse, delayed terminal (win-frame) feedback. Today's agents instead piggyback on the environment's level-completion signal rather than inferring the goal.

e.g. None deployed; the Goal-Setting pillar and the 'learned goal inference from delayed reward' direction (intrinsic motivation + hypothesis maintenance over candidate goals)

Table · Capability evidence

Capabilities supported by published evidence

Current systems are strongest at exploration; reusable world modeling and goal inference remain unproven.

Question answered: Which parts of the ARC-AGI-3 loop does each reported solver actually demonstrate?

Statuses summarize published evidence, not an intrinsic capability score. “Generalize” is an added cross-cutting diagnostic, not one of the benchmark’s four core capabilities.

Evidence coverage across current ARC-AGI-3 solution attempts.
System	Explore	Model	Infer goal	Plan	Generalize
The Duck1st · 2026 Milestone Prize #1	YesMultimodal views, segmentation, executable Python, and a private milestone win show effective first-contact exploration.	PartialPython supports explicit local hypotheses, but no transferable transition model is reported.	PartialThe agent can reason about progress during play; the report does not isolate independent hidden-goal induction.	YesThe REPL loop supports multi-step reasoning, action, observation, and revision.	YesPlaced first on the hidden holdout for 2026 Milestone #1.
Reki2nd · 2026 Milestone Prize #1	YesRare-color clicks, dead-signatures, recent-frame vision, and reflection reduce wasted exploration.	PartialStructured observations and reflection preserve local beliefs, but no reusable dynamics model is reported.	PartialThe policy writes a plan from visual evidence; the milestone report does not isolate goal induction.	YesShort action queues and recurring reflection support plan execution and revision.	YesPlaced second on the hidden holdout for 2026 Milestone #1.
forge3rd · 2026 Milestone Prize #1	YesA local visual policy, structured actions, and persistent memory reached the milestone podium.	PartialThe configurable framework can support code generation, but the winning profile did not establish a transferable model.	PartialReflection can retain hypotheses about progress; independent goal induction remains unmeasured.	YesStructured action selection with repair and guards supports reliable local execution.	YesPlaced third on the hidden holdout for 2026 Milestone #1.
StochasticGoose12.58% · 1st, preview	YesLearned policy predicts state-changing actions and searches from them.	PartialPredicts useful action effects, but does not induce an explicit dynamics model.	UnprovenNo reported mechanism for inferring the hidden win condition.	PartialSearch supplies short-horizon action selection rather than goal-directed planning.	YesRanked first on the private preview without an environment-specific harness.
Blind Squirrel6.71% · 2nd, preview	YesBuilds and searches an observed state graph.	PartialStores observed transitions but does not infer the rule producing them.	UnprovenNo reported hidden-goal induction.	PartialUses graph search to reach unexplored states.	YesRanked second on the private preview without a hand-built game harness.
Just-Explore (graph-based explorer)3.64% · 3rd, preview · 12 private levels	YesFrontier-driven exploration is the system's central mechanism.	PartialRecords state-action transitions but never induces a transition program.	UnprovenNo goal-acquisition component is reported.	PartialPlans paths to the nearest state with an untested action.	YesTraining-free policy was evaluated on private preview levels.
Duke LRM harnessall 3 public envs, ~human action counts · no transfer	PartialLLM queries can choose actions, but exploration depends on a custom harness.	PartialHistory compression supports local reasoning without a reusable learned dynamics model.	PartialCan reason about goals on the seen public environments.	YesAchieved human-like action counts on all three public environments.	NoReported gains did not transfer to unseen private environments.
Arcgenticaall 3 public envs, ~human action counts · no transfer	PartialAgent orchestration supports action selection on seen environments.	PartialSubagents reason over traces, but no reusable transition model is reported.	PartialCan infer progress on the public games with harness support.	YesCompleted all three public environments at roughly human action counts.	NoThe environment-specific harness did not transfer to private games.
Frontier LLMs (harness-free)~0.1–0.5% · <1% at launch	PartialModels act, but stall after relatively few expensive model-gated steps.	UnprovenNo robust world-model induction is demonstrated by the sub-1% results.	UnprovenHidden-goal inference remains a reported failure mode.	PartialCan propose action sequences, but does not revise them reliably under uncertainty.	PartialThe same harness-free protocol applied across hidden environments, but launch-era performance remained below 1%.

DemonstratedPartial evidenceNot yet demonstratedNot demonstrated

Sources: ARC-AGI-3 Technical Report · ARC-AGI-3 Preview: 30-Day Learnings · ARC Prize 2026: ARC-AGI-3 Milestone Prize #1.

The tradeoffs that shape it

Interaction budget vs per-step cost

cheap deterministic acting (~96,000 steps) expensive per-step LLM reasoning (~4,000 steps)

Every LLM-controlled step requires a model call, so these agents reach ~4,000 of roughly 96,000 allowed steps while cheaper algorithmic explorers can use much more of the budget.

Harness-free generalization vs hand-built scaffolding

harness-free general model (official leaderboard) hand-built per-environment harness (community leaderboard)

A hand-built harness reaches 97.1% on a seen environment but 0% harness-free and does not transfer to private environments, so seen-environment performance is a different claim from developer-aware generalization.

Search/explore vs model-and-infer-goal

record and search transitions (coverage of the state space) induce a reusable dynamics model and infer the goal

Milestone agents add memory, visual parsing, local planning, and code execution, but no published result yet establishes a reusable dynamics model and inferred goal that transfer across hidden environments.

Where the frontier actually is

Harness-free launch models scored 0.10–0.50%, while calibrated humans solved every included environment. The 2025 preview showed that informed exploration helps; the July 2026 milestone added local vision-language policies, memory, structured actions, context management, and executable Python. Public evidence still does not establish transferable world modeling or goal induction.

Two constraints shape the field. Model-controlled agents can use only a fraction of the interaction budget, which favors cheap acting plus selective reflection. Hand-built harnesses can solve seen public environments near human action counts but have not transferred to hidden ones. The main open problem is therefore a world model and goal representation learned from interaction that survives the next unseen environment.

Getting started

Enter the competition.

The competition is liveARC Prize 2026 runs two Kaggle tracks — the primary ARC-AGI-3 track and a final ARC-AGI-2 track — over a $2M pool, with ARC-AGI-3 as the continuing focus.. This overview gives the current schedule, submission format, and scoring target. For implementation details, see the Field Manual.

Prize

ARC Prize 2026 · in progress

Milestone #1 awarded The Duck, Reki, and forge. Milestone #2 closes September 30; final ARC-AGI-3 submissions close November 2 and awards are scheduled for December 4.

Format

An agent inside a notebook

You don't submit predicted grids. Kaggle reruns a generated IPython notebook containing an agent that plays live: it calls the ARC-AGI-3 API, receives a frame, returns one action, and repeats.

Scored on

Action efficiency (RHAE)

Not accuracy. Your action count per level is compared to a human's, squared, and capped — first-contact efficiency against a 100% human baseline.

Evaluation Trust Lens

A result ends where its evidence ends.

Exact replay, unseen states, unseen public games, and private environments answer different questions. Moving down the table requires new evidence; a stronger adjective cannot substitute for it.

Evidence surface	What it proves	What it does not prove
Observed transitions	The model can replay evidence already seen	It predicts unseen states
Unseen states in one game	Within-environment dynamics generalize	The representation transfers
Unseen public games	Some cross-game adaptation	Resistance to public-set tuning or contamination
Semi-private / private games	Held-out benchmark transfer	Broad agentic intelligence outside ARC

The open questions

What to answer next.

The highest-leverage unanswered questions for ARC-AGI-3, ranked. The ★ marks questions about synthesizing a world-model program and accepting it only after it replays every recorded transitionThe benchmark validator also re-executes recorded traces under both win and loss conditions before an environment is trusted..

1

★ Can a reusable world model be formed from a few dozen actions?

At launch, calibrated humans solved every environment while harness-free frontier AI scored below 1%. What learning procedure can fit an accurate, reusable next_frame = f(frame, action) program from only a few dozen actions, under a budget capped at 5× the calibrated human action count?
2

How do you infer the goal with no instructions and no reward until the win frame?

The Goal-Setting pillar provides no ground-truth signal until the terminal win frame. An agent must propose and revise hypotheses about desirable states from sparse, delayed feedback, even when early exploration looks indistinguishable from progress.
3

★ Does an exactly-replaying synthesized program generalize to unseen states?

A program that replays every recorded transition may still fail on unseen states. Does exact-replay acceptance predict held-out generalization, or merely confirm that the program memorized the observed transitions?
4

★ What replaces the missing execution oracle for the refinement loop?

The 2025 refinement loop relied on exact matches against demonstration pairs. ARC-AGI-3 removes that oracle because the goal is hidden. Can replaying observed transitions provide enough feedback to drive incremental program synthesis and repair?
5

What is the right compressed, queryable representation of interaction history?

64×64×16-color frames accumulate fast and a naive rolling window exhausts a model's context, making history management a first-order bottleneck — the strongest observed results came from harnesses that selectively retrieve and compress the action history (treating it as a queryable database), and cheap graph-explorers use ~96,000 allowed steps where LLM agents are interaction-limited to ~4,000 model-gated ones. The open question is which learned summaries preserve planning-relevant state.
6

How do you attribute the launch-era <1% across exploration, modeling, goal, and planning?

The single RHAE score cannot show whether an agent failed at exploration, world-model formation, goal inference, or planning. Diagnostics that isolate each pillar would help researchers choose which subproblem to attack first.
7

How do you measure genuine harness-free generalization instead of overfit scaffolding?

The same environment can yield 0% harness-free and ~97% with a hand-built harness, while systems that solved all three public environments failed on unseen private ones. How can evaluation distinguish general reasoning from answers encoded in the scaffolding?
8

How do you separate rational exploration cost from wasteful brute force under RHAE?

RHAE's squared penalty (the squared h/a level term, capped at 1.15) punishes wasted actions, but good exploration necessarily spends actions to reduce uncertainty — so the open question is whether there is a principled decomposition of "exploration actions a rational agent must spend" versus "brute-force waste," and whether the human baseline already bakes that tradeoff in correctly. It matters because the metric is the only definition of intelligence the benchmark exposes.

3 of these ask how to build and verify a world model from interaction: record transitions, synthesize a next-frame program, and test it by replaying the history exactly.

↩ Back to Part Ⅱ · The Atlas of 24 works

No instructions. Learn by acting.

Not a puzzle. A world.

The price of an extra action

Papers about ARC-AGI-3 itself.

Reference papers

ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence

Solution papers

Graph-Based Exploration for ARC-AGI-3 Interactive Reasoning Tasks

Hill-Climbing ARC-AGI-3

Executable World Models for ARC-AGI-3 in the Era of Coding Agents

OPINE-World: Programmatic World Modeling with Ontology-error-Prioritized Interactive Exploration for ARC-AGI-3

Do Coding Agents Need Executable World Models, Simplification, and Verification to Solve ARC-AGI-3?

The main reported solutions so far.

Exploration got smarter. Transfer is still the test.

Local vision-language policy + memory + tools

Graph-based state exploration (training-free algorithmic search)

Neural action-effect / RL exploration

Frontier LLM as per-step controller (harness-free)

LLM + hand-built harness / scaffolding

World-model program synthesis + planning (the main gap)

Goal / reward induction from delayed feedback

Capabilities supported by published evidence

The tradeoffs that shape it

Where the frontier actually is

Enter the competition.

ARC Prize 2026 · in progress

An agent inside a notebook

Action efficiency (RHAE)

A result ends where its evidence ends.

What to answer next.

★ Can a reusable world model be formed from a few dozen actions?

How do you infer the goal with no instructions and no reward until the win frame?

★ Does an exactly-replaying synthesized program generalize to unseen states?

★ What replaces the missing execution oracle for the refinement loop?

What is the right compressed, queryable representation of interaction history?

How do you attribute the launch-era <1% across exploration, modeling, goal, and planning?

How do you measure genuine harness-free generalization instead of overfit scaffolding?

How do you separate rational exploration cost from wasteful brute force under RHAE?