--- title: "ARC-AGI-2" source: "https://systems-analysis.info/eng/ARC-AGI-2" wiki: "systems-analysis.info/eng" article: "ARC-AGI-2" language: "en" categories: - "Category:English" - "Category:Large language models" - "Category:LLM benchmarks" - "Category:Machine learning" - "Category:Technology" revision_id: 60 wiki_created_at: 2026-09-06T22:16:52Z wiki_modified_at: 2026-09-06T22:16:52Z downloaded_at: 2026-09-07T22:20:42Z --- # ARC-AGI-2 **ARC-AGI-2** (**Abstraction and Reasoning Corpus for Artificial General Intelligence 2**) is a [benchmark](https://systems-analysis.info/eng/LLM_benchmarks "LLM benchmarks") for evaluating few-shot abstract reasoning, compositional generalization, and the efficient acquisition of new problem-solving skills by artificial intelligence systems. It was developed by the ARC Prize Foundation as the successor to ARC-AGI-1 and was introduced publicly in March 2025. Its design and human-calibration methodology were described by François Chollet, Mike Knoop, Gregory Kamradt, Bryan Landers, and Henry Pinkard in the paper *ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems*[\[1\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_paper-1). Each ARC-AGI-2 task presents a small number of example transformations between two-dimensional grids of colored cells. The transformation rule is not described in natural language. A human or AI system must infer the latent rule from the examples and apply it to one or more previously unseen input grids. Correct solutions must reproduce the required output dimensions, cell values, object positions, and colors exactly. The public release contains **1,000 training tasks** and **120 public evaluation tasks**. Two additional calibrated sets contain 120 tasks each: a semi-private evaluation set used for controlled model verification and live competition feedback, and a private evaluation set used for final competition scoring. Every evaluation task was solved within two attempts by at least two independent human participants[\[2\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_official-2)[\[3\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_repo-3). ARC-AGI-2 was initially extremely difficult for AI systems. The strongest baselines reported in May 2025 achieved approximately 3% on the semi-private evaluation set, while the winning open-source entry in ARC Prize 2025 reached 24.03% on the private set[\[1\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_paper-1)[\[4\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arc2025_results-4). Progress subsequently accelerated. On 9 July 2026, the official verified-results page reported **92.5%** for GPT-5.6 Sol at its maximum reasoning setting and 85.4% at its high setting[\[5\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-gpt56_sol-5). A verified score from a proprietary model is not equivalent to winning the ARC Prize competition. Competition entries must run offline within Kaggle resource limits, submit two candidate outputs per test input, and be open-sourced to qualify for prizes[\[6\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arc2026_competition-6). ARC-AGI should not be confused with the **AI2 Reasoning Challenge**, a separate multiple-choice science-question benchmark also commonly abbreviated as ARC. ## Background ARC-AGI-2 is based on the theory of intelligence proposed by François Chollet in the 2019 paper *On the Measure of Intelligence*. The paper defines intelligence primarily in terms of the efficiency with which a system acquires new skills, rather than the total number of tasks it can perform after extensive training[\[7\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-measure_intelligence-7). Under this view, a system should not receive strong evidence of general intelligence merely because it has memorized a large number of solutions or absorbed extensive domain-specific knowledge. A more informative evaluation presents previously unseen tasks, supplies only a few demonstrations, and measures whether the system can construct an appropriate new solution efficiently. The original Abstraction and Reasoning Corpus, later called ARC-AGI-1, implemented this idea through small grid-transformation problems. It was designed around three principles: - Every evaluation task should be sufficiently novel to resist direct memorization. - Solving the task should require little or no specialized world knowledge. - Ordinary humans should be able to infer the intended rule from a small number of examples. ARC-AGI-1 became the subject of public competitions between 2020 and 2024. Progress remained limited for several years, but test-time adaptation, synthetic-data generation, program search, and reasoning models produced substantial improvements during 2024[\[8\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arc2024_report-8). These advances exposed limitations in the original benchmark and motivated the development of ARC-AGI-2. ## Limitations of ARC-AGI-1 Addressed by ARC-AGI-2 | ARC-AGI-1 limitation | Consequence | ARC-AGI-2 response | |-------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | **Susceptibility to exhaustive program search** | Some tasks could be solved by searching large spaces of predefined transformations without constructing a broadly reusable abstraction. | Tasks vulnerable to known brute-force strategies were removed from the evaluation sets, and new tasks were designed to require more contextual and compositional reasoning. | | **Limited official human data** | Human solvability was known informally, but the private competition tasks lacked a large first-party baseline gathered under standardized conditions. | More than 400 people participated in controlled testing used to select and calibrate the new evaluation tasks. | | **Early human saturation** | Strong human solvers could complete nearly all ARC-AGI-1 tasks, leaving limited resolution across the upper range of human ability. | ARC-AGI-2 includes problems requiring more deliberate reasoning while remaining solvable by ordinary human participants. | | **Unequal subset difficulty** | Scores on the public, semi-private, and private sets could differ because the sets did not represent the same difficulty distribution. | The three ARC-AGI-2 evaluation sets were partitioned using empirical human-performance data so that their mean difficulty differed by approximately one percentage point. | | **Repeated private-set feedback** | Thousands of leaderboard submissions against the same hidden set created a gradual channel for indirect overfitting. | ARC-AGI-2 introduced newly curated semi-private and private sets and separated live feedback from final private scoring. | | **Limited compositional depth** | Many tasks could be summarized by one global transformation such as reflection, translation, recoloring, or object extraction. | ARC-AGI-2 emphasizes interacting rules, sequential transformations, contextual control, and symbols whose meanings are defined within a task. | The new benchmark preserves the basic interface of ARC-AGI-1 so that existing viewers, data loaders, program-synthesis systems, and training pipelines can be reused. The principal change is therefore not the visible data format but the distribution and cognitive structure of the tasks[\[1\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_paper-1). ## Design Goals The ARC-AGI-2 authors identified six principal goals[\[1\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_paper-1): | Goal | Description | |-------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | **Preserve the original principles** | Tasks should remain novel, require only elementary cognitive priors, and be accessible to human reasoning without specialized factual knowledge. | | **Preserve the task format** | Input-output demonstrations and test inputs remain grids of up to 30 by 30 cells using ten discrete values. | | **Reduce brute-force solvability** | Evaluation tasks should provide less reward for exhaustive enumeration of familiar transformations. | | **Establish first-party human baselines** | Task selection and difficulty estimates should be grounded in controlled testing rather than informal assumptions about human ability. | | **Increase score resolution** | The benchmark should differentiate systems across a wider range of reasoning capability instead of clustering most systems near zero or allowing rapid ceiling saturation. | | **Calibrate evaluation subsets** | Public, semi-private, and private evaluation scores should be comparable when a system has not overfit to one of the subsets. | The benchmark is described by its creators as a test of **general fluid intelligence**. This is a theoretical and operational claim about what the tasks are intended to measure; a high score does not by itself establish that a system possesses every capability commonly associated with artificial general intelligence. ## Task Format ### Grids and Cell Values An ARC-AGI-2 grid is a rectangular matrix containing integers from 0 through 9. A viewer normally renders the integers as colors, but the mapping between integers and displayed colors is conventional rather than semantically fixed. Grid dimensions range from **1 × 1** to **30 × 30**. The output grid may have the same dimensions as the input, or it may be larger, smaller, cropped, expanded, rotated, tiled, or otherwise restructured. A solver must infer: - Which cells form meaningful objects. - Which visual or relational properties are relevant. - Which examples illustrate the same latent rule. - Whether several rules operate simultaneously or sequentially. - How the output dimensions are determined. - How objects should be selected, transformed, combined, or discarded. - Whether colors represent literal appearance, object classes, instructions, quantities, or task-specific symbols. The cells should not always be interpreted as ordinary image pixels. A contiguous shape can represent an object, a number, a direction, a template, a container, an instruction, or a symbolic reference whose meaning is defined only by the demonstrations. ### Demonstration and Test Pairs Each task contains two lists: - `train` contains fully observed input-output demonstrations. - `test` contains one or more new inputs whose outputs must be predicted. The official guide states that `train` normally contains between two and ten pairs, typically three. The `test` list normally contains between one and three pairs, typically one[\[9\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arc_guide-9). A simplified task file is: { "train": [ { "input": [ [1, 0], [0, 0] ], "output": [ [1, 1], [1, 1] ] }, { "input": [ [0, 0], [4, 0] ], "output": [ [4, 4], [4, 4] ] }, { "input": [ [0, 0], [6, 0] ], "output": [ [6, 6], [6, 6] ] } ], "test": [ { "input": [ [0, 0], [0, 8] ], "output": [ [8, 8], [8, 8] ] } ] } The `output` field of public test pairs is included in the downloadable data so that researchers can evaluate their systems. In hidden competition tasks, the corresponding output is withheld. ### Absence of Natural-Language Instructions The task does not provide a verbal question such as “rotate the red object” or “count the blue squares.” The examples themselves define the problem. This design reduces dependence on: - Natural-language proficiency. - Encyclopedic knowledge. - Cultural familiarity. - Retrieval from external corpora. - Memorized problem statements. - Prompt interpretation. It also makes the benchmark narrower than a general reasoning evaluation. ARC-AGI-2 does not directly test reading comprehension, factual knowledge, dialogue, tool use, long-form explanation, or linguistic communication. ## Core Knowledge Priors ARC-AGI tasks are designed around a limited collection of elementary priors that most human participants can apply without explicit instruction[\[7\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-measure_intelligence-7)[\[1\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_paper-1). | Prior | Examples in grid tasks | |-----------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------| | **Objectness** | Treating connected or visually grouped cells as persistent entities rather than unrelated pixels. | | **Object persistence** | Recognizing that an object remains the same object after movement, recoloring, rotation, scaling, or partial occlusion. | | **Goal-directedness** | Interpreting a transformation as completing, repairing, sorting, enclosing, connecting, or otherwise changing a scene toward a coherent outcome. | | **Elementary number sense** | Counting objects, comparing quantities, repeating a transformation a specified number of times, or using cardinality as a symbolic value. | | **Basic geometry** | Applying translation, rotation, reflection, scaling, alignment, containment, intersection, and relative position. | | **Topology** | Reasoning about connectivity, holes, boundaries, inside-outside relationships, adjacency, and separation. | | **Symmetry** | Detecting or completing reflective, rotational, translational, or repeated structure. | | **Simple causality** | Inferring that one object acts as a source, obstacle, key, marker, instruction, or trigger for another transformation. | These priors are not literally absent from a solver's prior knowledge. Humans have extensive perceptual and developmental experience with objects, space, quantity, and causality. ARC-AGI attempts to minimize specialized acquired knowledge, not all prior knowledge. ## Dataset Composition | Subset | Tasks | Public availability | Calibration | Primary use | |-----------------------------|-----------|---------------------|---------------------------|-----------------------------------------------------------------------------------------------------------| | **Public training** | **1,000** | Public | Not difficulty-calibrated | Training, task-format familiarization, development of ARC-specific priors, and synthetic-data generation. | | **Public evaluation** | **120** | Public | Human-calibrated | Open research evaluation and local system development. | | **Semi-private evaluation** | **120** | Hidden | Human-calibrated | Controlled verification of remotely hosted models and live competition feedback. | | **Private evaluation** | **120** | Hidden | Human-calibrated | Final competition scoring and high-stakes evaluation with minimal third-party exposure. | ARC-AGI-2 dataset structure The public repository therefore contains **1,120 tasks**. The public training set combines tasks inherited from ARC-AGI-1 with tasks created or selected for ARC-AGI-2. It spans a broad range of difficulty and was not constructed as a representative evaluation sample[\[3\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_repo-3)[\[9\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arc_guide-9). The evaluation subsets contain 360 tasks in total, although only the 120 public tasks can be downloaded directly. The official benchmark page describes the three evaluation subsets as statistically similar in human-facing difficulty, with an expected difference below approximately one percentage point when a system has not overfit to a particular subset[\[2\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_official-2). The semi-private designation indicates that the tasks may be exposed to a limited number of third parties during controlled API or leaderboard evaluations. The private set is intended to remain inaccessible outside the final competition infrastructure. ## Human Calibration ### Testing Protocol ARC-AGI-2 used a larger first-party human study than its predecessor. Testing was conducted from November 2024 through May 2025 in controlled sessions, principally in San Diego[\[1\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_paper-1). Participants worked individually on computers using a purpose-built grid interface. They received a short tutorial and survey before attempting randomly assigned candidate tasks. Sessions lasted approximately 90 minutes, and no more than 34 participants were tested in the room at one time. The study used monetary compensation consisting of a base payment and a performance bonus. The incentive was intended to encourage serious attempts while retaining participants from the general public rather than only professional puzzle solvers. | Statistic | Reported value | |--------------------------------------------|--------------------------| | Unique participants | **407** | | Testing sessions | **515** | | Unique candidate test pairs | **1,848** | | Recorded attempts | **13,405** | | Solved attempts | **8,277** | | Overall solved-attempt proportion | Approximately **62%** | | Maximum participants tested simultaneously | 34 | | Session duration | Approximately 90 minutes | Human-testing statistics reported in the ARC-AGI-2 paper The candidate pool included newly authored tasks and previously unused reserve tasks from earlier ARC development. ARC-AGI-1 public training tasks were excluded from the human-screening pool. The authors reported no clear statistically significant relationship between performance and participants' occupations, industries, programming experience, mathematics experience, or puzzle background. This was interpreted as evidence that the selected tasks depended more strongly on general problem solving than on one specialized professional skill[\[1\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_paper-1). ### Selection Criterion A candidate task advanced from the human-testing stage only if at least two independent participants solved one or more of its test pairs within their first two attempts. This rule aligns human and AI evaluation conditions. A task was not retained merely because its author knew a solution; it had to demonstrate independent solvability by people who encountered it without prior exposure. After selection, all public, semi-private, and private evaluation tasks underwent two validation layers: - At least two external human testers independently solved the task. - An internal reviewer other than the original author examined and solved the task. The training collection received less exhaustive calibration. Some training tasks were selected primarily to demonstrate useful concepts or priors, and not every training item received the complete evaluation-set testing protocol. ### Difficulty Index and Partitioning For each candidate task, the researchers calculated an empirical difficulty index based on the proportion of participants who produced a fully correct answer. The retained tasks were partitioned among the public, semi-private, and private sets so that their mean human accuracy was similar. Newly authored tasks were preferentially assigned to hidden subsets, while previously public material remained public. The resulting evaluation data had the following reported human characteristics[\[1\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_paper-1): - Final test pairs were solved by approximately **75%** of people who attempted them. - The average participant solved approximately **66%** of the tasks they attempted. - Every retained evaluation task was solved by at least two people within two attempts. - Human solution time averaged approximately 2.7 minutes for successfully completed tasks. The 66% value is an average over individual test takers and assigned tasks. It does not mean that 34% of the benchmark is unsolvable by humans. Across the participant population, every retained task had multiple verified human solutions. ## What Makes ARC-AGI-2 Difficult ARC-AGI-2 tasks generally contain more information than ARC-AGI-1 tasks. They tend to use larger grids, more objects, more relationships, and a greater number of interacting concepts[\[1\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_paper-1). The benchmark particularly emphasizes four forms of compositional generalization. | Challenge type | Description | Typical failure mode | |----------------------------------|---------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------| | **Multi-rule composition** | Several transformations must be applied within the same problem, potentially to different objects or regions. | A system identifies one prominent rule but ignores another required transformation. | | **Multi-step composition** | The output of one transformation becomes the input state for the next transformation. | A system attempts to predict the final arrangement without executing the intermediate states. | | **Contextual rule application** | The applicable operation depends on a contextual marker, object property, location, color, or relation. | A system learns the general operation but applies it to every object instead of selecting the contextually appropriate case. | | **In-context symbol definition** | An object or pattern represents an instruction or value whose meaning is defined by the demonstrations. | A system treats the symbol as a shape to copy rather than as a task-specific semantic variable. | ### Multi-Rule Composition An ARC-AGI-1 problem could often be described by a single rule such as “move all objects downward,” “reflect the shape,” or “extract the unique object.” ARC-AGI-2 tasks are more likely to require several coordinated operations. A solver may need to: 1. Detect a framed region. 2. Crop or isolate that region. 3. Identify objects outside it. 4. Rescale the objects. 5. Match them to openings with corresponding shapes. 6. Insert them in a specified order. A candidate solution that explains only some training pairs is insufficient. Every operation must generalize consistently across all demonstrations and test inputs. ### Multi-Step Composition Some tasks require repeated state updates. The correct position or orientation of the next object depends on the result of placing the previous object. Such a task cannot always be solved reliably by recognizing a static visual analogy. The solver may need to simulate an algorithm, retain an intermediate state, and stop only after a condition is satisfied. ### Contextual Control Flow A transformation may be selected by a cue such as: - The color of an outline. - The position of a marker. - The orientation of an arrow-like object. - Whether a container is open or closed. - The number of cells in a reference object. - Which object is unique within a local group. The same visible object may therefore require different treatment in different examples. This resembles conditional control flow in a program. ### Symbolic Interpretation Some objects stand for something other than their literal visual structure. A small pattern may encode: - A target color. - A number of repetitions. - A direction. - A sorting order. - A mapping between object classes. - A transformation to apply elsewhere. The symbol's meaning is local to the task and cannot be recovered from a global dictionary. It must be induced from the current demonstrations. ### Resistance to Simple Search Program synthesis remains a valid approach to ARC-AGI-2, but the task distribution was designed to reduce the effectiveness of enumerating short combinations of familiar geometric primitives. A larger description length, deeper transformation sequence, and context-sensitive rule increase the search space. A successful search system therefore benefits from learned intuition, abstraction discovery, probabilistic guidance, or task-specific adaptation rather than undirected enumeration alone. ## Evaluation Methodology ### Exact-Match Scoring For each hidden test input, a solver may submit **two candidate output grids**. A test input passes if either candidate matches the reference output exactly. Exact matching requires agreement on: - Output height. - Output width. - Every cell value. - Every object position. - Every color or symbol. - Empty and background cells. A task is conventionally described as solved when all required test outputs are produced correctly under the benchmark protocol[\[3\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_repo-3)[\[6\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arc2026_competition-6). A parser-safe description of the task rule is: `test_input_pass = 1 if either of the two candidate grids exactly matches the reference grid; otherwise 0` The aggregate score is the proportion of accepted test cases or tasks under the implementation used by the relevant evaluation harness: `accuracy = accepted evaluation units / all evaluated units` Published reports should identify the scoring implementation, particularly when tasks contain more than one test input. ### Pixel Correctness Some researchers additionally report **pixel correctness**, the proportion of output cells assigned the correct value. This diagnostic can distinguish a nearly correct transformation from an unrelated output. Pixel correctness is not the official headline metric. A grid with one incorrect cell fails exact-match evaluation, while a grid with the wrong dimensions cannot be considered a valid exact solution even if many visible cells overlap. The strict rule reflects the intended task: infer and execute the complete transformation rather than approximate the visual appearance. ### Two-Attempt Rule The two-output allowance applies to both human and AI evaluation. It reduces the effect of a single ambiguous decision without permitting unrestricted search through many candidate answers. The metric should not be confused with pass@2 as commonly used in code-generation benchmarks. In ARC-AGI-2, both candidate grids are part of the defined answer protocol for one test input. Allowing two guesses can benefit systems that generate diverse hypotheses, but both candidates must be created within the same computational and competition constraints. ### Efficiency Reporting ARC-AGI-2 introduced an explicit efficiency dimension. The official leaderboard reports monetary cost per task where the model provider's pricing permits such a calculation[\[2\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_official-2). The intended interpretation is that a system solving a task through a concise abstraction provides stronger evidence of adaptive intelligence than a system obtaining the same answer through an extremely expensive exhaustive search. Relevant efficiency variables include: - Model API cost. - Number of reasoning tokens. - Number of sampled candidate solutions. - Test-time fine-tuning cost. - Search iterations. - CPU and accelerator time. - Memory consumption. - Wall-clock latency. - Cost of external verifiers and model ensembles. Monetary cost is convenient but imperfect. It depends on provider pricing and may not reflect the actual energy, hardware, research, or training resources used. ## = Public and Hidden Evaluation The three evaluation tiers serve different purposes. | Evaluation tier | Main advantage | Main risk or limitation | |------------------|------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------| | **Public** | Complete reproducibility, per-task analysis, and rapid local development. | Researchers can inspect the tasks directly and overfit their methods through repeated feedback. | | **Semi-private** | Lower leakage risk and controlled verification of commercial or remotely hosted systems. | Some third parties or model providers may encounter the tasks during evaluation. | | **Private** | Strongest protection against direct benchmark-specific development. | Independent researchers cannot inspect the tasks or fully reproduce the grading process. | Public-evaluation results remain useful for debugging and open comparison, but they become less informative when a system was tuned repeatedly against the same 120 tasks. ## Initial Baselines The ARC-AGI-2 paper reported the following semi-private results as of 14 May 2025[\[1\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_paper-1): | System | ARC-AGI-1 | ARC-AGI-2 | |----------------------------------------|-----------|-----------| | o3-mini, high reasoning | 34.5% | **3.0%** | | o3, medium reasoning | 53.0% | **3.0%** | | ARChitects, ARC Prize 2024 | 56.0% | 2.5% | | o4-mini, medium reasoning | 41.8% | 2.4% | | Icecuber, ARC competition 2020 | 17.0% | 1.6% | | o1-pro, low reasoning | 23.3% | 0.9% | | Claude 3.7 with an 8K reasoning budget | 21.2% | 0.9% | Initial ARC-AGI-2 semi-private baselines The authors cautioned that scores below approximately 5% did not provide a stable capability signal. At that range, solving one or two tasks could result from incidental pattern matching, a narrow heuristic, or evaluation noise. The contrast with ARC-AGI-1 showed that the second dataset was not merely a larger sample of the original difficulty distribution. Methods that solved a substantial fraction of ARC-AGI-1 frequently returned to near-zero performance on ARC-AGI-2. ## ARC Prize 2025 The first public competition centered on ARC-AGI-2 was conducted through Kaggle in 2025. Submissions ran without Internet access on four NVIDIA L4 GPUs and received a 12-hour wall-clock budget. Each submission processed 120 semi-private tasks and 120 private tasks[\[1\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi2_paper-1). The competition attracted[\[10\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arc2025_report-10): - **1,455 teams**. - **15,154 competition entries**. - **90 research-paper submissions**. The highest private-set results were[\[4\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arc2025_results-4): | Place | Team | Private score | |-------|----------------|---------------| | 1 | **NVARC** | **24.03%** | | 2 | the ARChitects | 16.53% | | 3 | MindsAI | 12.64% | | 4 | Lonnie | 6.67% | | 5 | G. Barbadillo | 6.53% | ARC Prize 2025 private evaluation results No competition entry reached the 85% threshold required for the principal completion prize. The NVARC system combined several approaches, including a test-time-trained ARC solver and components based on recursive reasoning. Its result was substantially higher than the initial May 2025 baselines while remaining far below average human performance. ## Refinement Loops The ARC Prize 2025 technical report identified the **refinement loop** as the central methodological development of the competition[\[10\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arc2025_report-10). A refinement loop generally performs the following operations: 1. Generate a candidate rule, program, or output. 2. Execute or test it against the demonstration pairs. 3. Measure how and where it fails. 4. Modify the candidate using the feedback. 5. Repeat until a stopping criterion or resource limit is reached. 6. Rank the remaining candidates and submit one or two outputs. Refinement can occur in several spaces: | Refinement type | Example | |---------------------------------------|-----------------------------------------------------------------------------------------------------| | **Program space** | Mutating or recombining candidate transformation programs. | | **Natural-language hypothesis space** | Asking a reasoning model to criticize and rewrite its description of the task rule. | | **Output space** | Iteratively correcting a generated grid using comparison with known demonstrations. | | **Weight space** | Fine-tuning a model separately for the current task using augmented versions of its demonstrations. | | **Ensemble space** | Combining predictions from models, geometric perspectives, or independently generated hypotheses. | The method differs from producing one frozen-model response. It allows a system to acquire task-specific skill during inference, which is closer to the adaptation capability ARC-AGI was designed to measure. ## Progress Through July 2026 Performance increased rapidly after the 2025 competition. | Date | System or event | Evaluation condition | Reported score | |---------------|--------------------------------|----------------------------------------|----------------| | May 2025 | o3-mini, high reasoning | Initial semi-private baseline | 3.0% | | November 2025 | NVARC | ARC Prize 2025 private competition set | 24.03% | | July 2026 | GPT-5.6 Sol, low reasoning | Official verified evaluation | 42.5% | | July 2026 | GPT-5.6 Sol, medium reasoning | Official verified evaluation | 67.1% | | July 2026 | GPT-5.6 Sol, high reasoning | Official verified evaluation | 85.4% | | July 2026 | GPT-5.6 Sol, maximum reasoning | Official verified evaluation | **92.5%** | Selected milestones in ARC-AGI-2 performance The GPT-5.6 Sol results reveal a strong dependence on inference-time resources[\[5\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-gpt56_sol-5): | Reasoning setting | ARC-AGI-1 | ARC-AGI-2 | |-------------------|-----------|-----------| | Low | 74.5% | 42.5% | | Medium | 92.5% | 67.1% | | High | 97.0% | 85.4% | | Extra High | 97.5% | 90.0% | | Maximum | 96.5% | **92.5%** | The non-monotonic ARC-AGI-1 values illustrate that a larger inference budget does not guarantee a better score on every finite benchmark. Sampling variation, alternative hypotheses, and task-specific failure modes can change which tasks are solved. The 92.5% result exceeded the average 66% human score reported for the calibrated evaluation tasks. It does not imply that the model solved tasks more efficiently than humans, that it matched the best human participants, or that it possesses general competence outside the ARC grid domain. The result also does not automatically satisfy the ARC Prize competition objective. The official verified leaderboard can include closed-weight and remotely hosted systems, whereas prize-eligible Kaggle submissions must run under the specified offline hardware conditions and must be open-sourced. ## Principal Solution Approaches ### Discrete Program Search A program-search system defines primitive operations such as: - Extract connected components. - Detect bounding boxes. - Reflect or rotate a grid. - Recolor selected objects. - Translate an object. - Crop a region. - Tile a pattern. - Count objects or cells. - Sort objects by size or position. - Draw lines between selected points. - Combine masks. It then searches for a composition that maps every training input to its corresponding output. A well-designed domain-specific language can express common ARC concepts compactly. Its principal limitation is combinatorial explosion: deeper and more context-sensitive tasks require searching an extremely large number of candidate programs. ### Neural Test-Time Adaptation Test-time adaptation trains or updates a model separately for each task. Because only a few demonstration pairs are available, systems generate additional examples through: - Rotations and reflections. - Color permutations. - Object translations. - Input-output inversion where valid. - Synthetic tasks produced by known transformation programs. - Crops and scale changes. - Alternative object orderings. The model is then optimized on these augmented examples before predicting the hidden output. This approach produced major ARC-AGI-1 improvements during 2024. ARC-AGI-2 makes adaptation harder because its rules are more compositional and may not remain valid under naive augmentation. ### Synthetic Task Generation The public dataset is small compared with ordinary neural training corpora. Many systems therefore generate large collections of ARC-like tasks. Synthetic data can be produced by: - Sampling programs from a transformation language. - Mutating existing tasks. - Recombining objects and rules. - Using language models to propose task generators. - Applying random geometric transformations. - Creating paired grids from procedural simulators. Synthetic training is useful only when its distribution transfers to the evaluation tasks. A model can achieve high accuracy on generated tasks while learning biases absent from human-authored ARC-AGI-2 problems. ### Reasoning Models and Code Generation A multimodal or language model can convert the grids into a textual representation, propose a rule, write code implementing the rule, and execute the code against demonstrations. A typical workflow is: 1. Parse the colored grids into objects and relationships. 2. Describe candidate transformations. 3. Generate a Python program. 4. Run it on all training inputs. 5. Compare predicted and reference outputs. 6. Diagnose inconsistencies. 7. Revise the program. 8. Generate the final test outputs. Direct prompting without execution or refinement historically performed poorly. Code execution provides a precise feedback signal and supports iterative correction. ### Ensembles and Perspective Transformations Ensemble systems combine solvers with different representations or inductive biases. One solver may specialize in object transformations, another in cellular patterns, and another in program induction. A grid can also be evaluated from several transformed perspectives: - Original orientation. - Horizontal or vertical reflection. - Rotations. - Color-normalized representation. - Object-centered crops. - Transposed coordinates. Predictions are transformed back and combined. Agreement among independent perspectives can help identify robust candidates, although a large ensemble increases inference cost. ### Small Recursive Networks Research associated with ARC Prize 2025 showed that competitive results do not necessarily require a very large pretrained model. Tiny recursive or recurrent systems can repeatedly update an internal representation, train at test time, and solve a subset of ARC tasks with millions rather than billions of parameters[\[11\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-trm_paper-11). Such systems support the argument that iterative computation and task-specific adaptation can be more important than parameter count. Their performance still depends heavily on generated training data, optimization procedure, and benchmark-specific representation. ### Hybrid Neuro-Symbolic Systems Hybrid systems use neural models to guide symbolic search. The neural component may propose: - Relevant objects. - Candidate relations. - A probable transformation family. - A partial program sketch. - Useful program-search branches. - A ranking over generated candidates. The symbolic component then executes exact operations and verifies them against the demonstrations. This division attempts to combine flexible pattern recognition with reliable discrete computation. ## ARC Prize 2026 ARC Prize 2026 retained a separate ARC-AGI-2 competition track. Its stated objective is to reach **85% accuracy** on the private evaluation set within Kaggle efficiency limits[\[6\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arc2026_competition-6). The announced ARC-AGI-2 prize pool is: | Prize category | Amount | Purpose | |-----------------|-----------------|------------------------------------------------------------------------------------------------------------| | Progress prizes | **US\$275,000** | Awards for the highest-ranked eligible solutions. | | Grand Prize | **US\$275,000** | Evaluation of the highest-scoring open solution write-up across several research and engineering criteria. | | Bonus Prize | **US\$150,000** | Awarded to the first eligible solution reaching at least 85% on the private set. | | Total | **US\$700,000** | — | Competition requirements include: - Submission as a Kaggle notebook. - No Internet access during evaluation. - Two candidate outputs for every test input. - Compliance with the announced hardware and runtime limits. - Public release of code and methods for prize eligibility. The 85% bonus is defined by these competition conditions. A closed commercial model exceeding 85% on an official verified evaluation does not itself constitute a prize-eligible Kaggle submission. ## Relationship to ARC-AGI-3 ARC-AGI-2 evaluates **static** reasoning. All information needed to infer the task is contained in the demonstrations and test input. The solver does not take exploratory actions that alter the problem environment. ARC-AGI-3, introduced in 2026, extends the series to interactive environments in which an agent must explore, infer goals, discover transition rules, remember earlier observations, and plan action sequences[\[12\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi3_paper-12). | Property | ARC-AGI-2 | ARC-AGI-3 | |-------------------|-------------------------------------------------------------------|-----------------------------------------------------------------| | Interaction | Static input-output prediction | Repeated observation and action | | Goal | Infer a transformation and construct output grids | Discover the environment's objective and achieve it | | Feedback | Demonstration pairs before submission | State changes and rewards during interaction | | Main capabilities | Abstraction, rule induction, composition, symbolic interpretation | Exploration, planning, memory, goal acquisition, and adaptation | | Output | One or two candidate grids | A sequence of actions | | Efficiency | Cost or compute per task | Actions and resources relative to human baselines | ARC-AGI-3 complements rather than replaces ARC-AGI-2. A system can be strong at static pattern induction while remaining weak at exploration and long-horizon planning, or vice versa. ## Interpretation of Scores An ARC-AGI-2 score measures performance on a narrow but deliberately challenging class of few-shot grid transformations. A higher score provides evidence that the evaluated system can: - Infer latent rules from very few examples. - Compose familiar operations in unfamiliar ways. - Maintain intermediate states. - Interpret task-local symbols. - Produce exact structured outputs. - Adapt at inference time. - Search candidate solutions efficiently enough to meet the evaluation budget. It does not directly establish competence in: - Natural-language communication. - Factual reasoning. - Scientific research. - Social understanding. - Tool use. - Autonomous planning in an external environment. - Software engineering. - Physical interaction. - Safety or alignment. - General learning across arbitrary real-world domains. A benchmark result should therefore be interpreted as one component of a broader evaluation rather than a binary test for the existence of AGI. ## Limitations and Criticism ### Narrow Construct Validity ARC-AGI-2 operationalizes intelligence as efficient adaptation to novel grid-transformation tasks. This follows the theoretical framework proposed by the benchmark's creator, but it is not the only accepted definition of intelligence. Other theories emphasize: - Language and communication. - World modelling. - Long-term learning. - Social cognition. - Embodied action. - Metacognition. - Creativity. - Transfer across natural domains. - Goal selection. - Autonomous agency. A system can excel on ARC-AGI-2 while lacking many of these capabilities. Conversely, a useful general-purpose system may perform poorly because its interface or training does not support exact grid construction. ## Human Priors Are Substantial The benchmark is often described as requiring minimal prior knowledge, but human solvers bring extensive developmental and perceptual priors. Humans already understand: - Objects and boundaries. - Space and direction. - Counting. - Symmetry. - Occlusion. - Containers. - Repetition. - Causal interactions. - Visual grouping. - Goal-directed transformation. An AI system may need to acquire comparable representations through training. The distinction between acceptable general priors and benchmark-specific preparation is not always precise. ## Grid Modality ARC-AGI-2 removes linguistic and factual confounds by using discrete grids. This improves experimental control but restricts ecological validity. Real reasoning problems often involve: - Noisy continuous perception. - Ambiguous language. - Missing information. - External tools. - Changing goals. - Long documents. - Social context. - Irreversible actions. - Uncertain or probabilistic outcomes. Success on clean 30 × 30 grids does not guarantee transfer to these conditions. ## Public Evaluation Overfitting The complete public evaluation set is available online. Researchers can inspect every task, create task-specific heuristics, tune prompts, or repeatedly modify a system in response to the same score. Even without directly adding the answers to a training set, repeated development against the public evaluation tasks can produce selection bias. For a credible generalization claim, the final system should be evaluated on a hidden set that was not used for: - Prompt design. - Architecture selection. - Hyperparameter tuning. - Error analysis. - Synthetic-data selection. - Ensemble construction. - Early stopping. ## Hidden-Set Reproducibility Semi-private and private tasks reduce contamination but limit independent verification. External researchers cannot fully inspect: - Task quality. - Reference outputs. - Difficulty distribution. - Possible redundancies. - Evaluator implementation. - Individual system failures. The ARC Prize Foundation can provide controlled verification, but complete scientific reproducibility is necessarily weaker than on the public set. ## Calibration Is Human-Relative The three evaluation sets were calibrated using human success rates. A similar human-facing difficulty distribution does not guarantee equal difficulty for every AI architecture. A model can exploit a feature that occurs more frequently in one set, struggle with a representation imbalance invisible to humans, or overfit indirectly to the public subset. The official estimate of less than one percentage point of expected difference therefore depends on the assumption that the evaluated system samples from the same effective task distribution and has not adapted to one partition. ## Exact-Match Brittleness Exact matching provides objective grading but gives no partial credit. A system that produces an otherwise correct 20 × 20 grid with one wrong cell receives the same binary result as a system that outputs an unrelated grid. The score does not reveal whether the error came from: - Incorrect abstraction. - Incorrect output dimensions. - One execution mistake. - A color-label mismatch. - A malformed serialization. - Failure on only one of several test inputs. Pixel correctness and per-task diagnostics can supplement the official score, but they are not always reported. ## Two Candidate Outputs The two-output protocol allows a solver to hedge between competing hypotheses. This is practical and applies equally to human and AI participants, but it changes the construct being measured. A system may obtain credit without selecting one final interpretation if it can place two plausible answers in its submission. Comparisons with benchmarks allowing one answer, many sampled answers, or verifier-guided selection require caution. ## Small Evaluation Sets Each calibrated set contains 120 tasks. When accuracy is reported at the task level, one additional solved task changes the score by approximately **0.83 percentage points**. Differences of one or two points may therefore reflect only a small number of examples. They can also be influenced by: - Sampling variation. - Reasoning-budget configuration. - Output-selection strategy. - Random seeds. - Tool failures. - Parser errors. - Changes in model endpoints. Confidence intervals and per-task results provide more information than a point estimate alone. ## Human Baseline Interpretation The reported 66% average does not describe an idealized expert or the maximum human score. Participants received different task subsets and worked under time limits. The following statements are simultaneously true: - The average participant solved approximately 66% of attempted tasks. - Every retained task was solved by at least two people. - No single reported participant was required to solve the entire evaluation collection. - An AI system can exceed 66% without solving every task humans can solve collectively. Statements that a model is “superhuman” on ARC-AGI-2 should specify whether they refer to average individual accuracy, best individual performance, aggregate human coverage, time, cost, or another criterion. ## Monetary Cost Is an Imperfect Efficiency Metric API price per task is easy to calculate but is not a stable physical measure. Prices can change independently of model efficiency, and providers may subsidize or bundle inference. The metric generally excludes: - Model-training compute. - Research and engineering cost. - Hardware manufacturing. - Energy and cooling. - Synthetic-data generation. - Failed development experiments. - Human prompt design. - Proprietary verifier infrastructure. Cost remains useful for comparing deployed configurations at a particular time, but it should not be treated as a complete measure of cognitive efficiency. ## Protocol Dependence ARC-AGI-2 performance can depend heavily on: - Reasoning-token budget. - Number of generated hypotheses. - Use of code execution. - Test-time fine-tuning. - Synthetic augmentations. - Search depth. - Ensemble size. - Candidate-ranking model. - Access to the public training and evaluation tasks. - Wall-clock limit. - Hardware. - Output parser. - Whether the system is allowed to call a closed model. The difference between GPT-5.6 Sol's low-setting score of 42.5% and maximum-setting score of 92.5% illustrates that the evaluated inference configuration can be as important as the model family name[\[5\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-gpt56_sol-5). ## Rapid Saturation ARC-AGI-2 was introduced when leading baselines scored approximately 3%. A verified result above 90% was reported little more than a year later. This rapid progress has two possible interpretations: - Reasoning and test-time computation improved substantially. - A finite static benchmark can be optimized faster than anticipated once it becomes a major research target. The two explanations are not mutually exclusive. Continued progress should be tested on new hidden tasks, generated task families, alternative modalities, and interactive environments. ## Benchmark-Specific Knowledge The ARC Prize 2025 report argued that frontier systems increasingly benefit from broad knowledge of transformation templates, program patterns, and ARC-like tasks[\[10\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arc2025_report-10). A model may not memorize a specific evaluation grid yet still acquire an extensive library of benchmark-relevant operations from: - Public ARC tasks. - Community explanations. - Synthetic ARC generators. - Solver source code. - Published trajectories. - Related benchmark tasks. - Training data created specifically for ARC. This raises a boundary question: at what point does accumulated benchmark-specific knowledge replace the intended acquisition of a genuinely new skill? ## Subjective Task Authorship ARC tasks are manually authored or curated. Decisions about whether a task is: - Unique. - Elegant. - Ambiguous. - Brute-force resistant. - Dependent only on acceptable priors. - Redundant with another task. - Solvable through the intended abstraction. require expert judgment. Human validation reduces outright errors but cannot prove that every task has one uniquely natural interpretation. Two solvers may construct different rules that fit all demonstrations yet diverge on a test input. ## Static Input-Output Setting The solver cannot request another example, probe the rule, or interact with the environment. It receives a fixed amount of evidence and must commit to an answer. Real learning often involves active experimentation. ARC-AGI-3 was introduced partly to evaluate this broader adaptive process through exploration, planning, memory, and goal discovery[\[12\]](https://systems-analysis.info/eng/ARC-AGI-2#cite_note-arcagi3_paper-12). ## External links - Official ARC-AGI-2 overview - Official ARC-AGI-2 repository - ARC-AGI-1 and ARC-AGI-2 technical guide - Interactive ARC task interface - ARC-AGI-2 technical paper - Official verified-results page - ARC Prize 2026 ARC-AGI-2 competition ## Literature - Chollet, F. (2019). *On the Measure of Intelligence*. arXiv:1911.01547. - Chollet, F.; Knoop, M.; Kamradt, G.; Landers, B.; Pinkard, H. (2025). *ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems*. arXiv:2505.11831. - Johnson, A. et al. (2021). *Fast and Flexible: Human Program Induction in Abstract Reasoning Tasks*. arXiv:2103.05823. - Moskvichev, A.; Odouard, V.; Mitchell, M. (2024). *The ConceptARC Benchmark: Evaluating Understanding and Generalization in the ARC Domain*. Transactions on Machine Learning Research. arXiv:2305.07141. - Wang, R. et al. (2024). *H-ARC: A Robust Estimate of Human Performance on the Abstraction and Reasoning Corpus Benchmark*. arXiv:2409.01374. - Chollet, F. et al. (2024). *ARC Prize 2024: Technical Report*. arXiv:2412.04604. - Chollet, F.; Knoop, M.; Kamradt, G.; Landers, B. (2026). *ARC Prize 2025: Technical Report*. arXiv:2601.10904. - ARC Prize Foundation (2026). *ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence*. arXiv:2603.24621. ## References 1. ↑ [1.00](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_paper_1-0) [1.01](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_paper_1-1) [1.02](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_paper_1-2) [1.03](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_paper_1-3) [1.04](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_paper_1-4) [1.05](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_paper_1-5) [1.06](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_paper_1-6) [1.07](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_paper_1-7) [1.08](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_paper_1-8) [1.09](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_paper_1-9) [1.10](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_paper_1-10) Chollet, F.; Knoop, M.; Kamradt, G.; Landers, B.; Pinkard, H. "ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems". *arXiv:2505.11831*, 2025, revised 2026. [1] 2. ↑ [2.0](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_official_2-0) [2.1](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_official_2-1) [2.2](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_official_2-2) ARC Prize Foundation. "ARC-AGI-2". [2] 3. ↑ [3.0](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_repo_3-0) [3.1](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_repo_3-1) [3.2](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi2_repo_3-2) ARC Prize Foundation. "Abstraction and Reasoning Corpus for Artificial General Intelligence v2". GitHub repository. [3] 4. ↑ [4.0](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arc2025_results_4-0) [4.1](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arc2025_results_4-1) ARC Prize Foundation. "ARC Prize 2025 Results and Analysis". 5 December 2025. [4] 5. ↑ [5.0](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-gpt56_sol_5-0) [5.1](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-gpt56_sol_5-1) [5.2](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-gpt56_sol_5-2) ARC Prize Foundation. "GPT-5.6 Sol — ARC-AGI Results". 9 July 2026. [5] 6. ↑ [6.0](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arc2026_competition_6-0) [6.1](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arc2026_competition_6-1) [6.2](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arc2026_competition_6-2) ARC Prize Foundation. "ARC Prize 2026 — ARC-AGI-2 Competition". [6] 7. ↑ [7.0](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-measure_intelligence_7-0) [7.1](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-measure_intelligence_7-1) Chollet, F. "On the Measure of Intelligence". *arXiv:1911.01547*, 2019. [7] 8. [↑](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arc2024_report_8-0) ARC Prize Foundation. "ARC Prize 2024: Technical Report". *arXiv:2412.04604*, 2024. [8] 9. ↑ [9.0](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arc_guide_9-0) [9.1](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arc_guide_9-1) ARC Prize Foundation. "ARC-AGI-1 & ARC-AGI-2 Guide". [9] 10. ↑ [10.0](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arc2025_report_10-0) [10.1](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arc2025_report_10-1) [10.2](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arc2025_report_10-2) Chollet, F.; Knoop, M.; Kamradt, G.; Landers, B. "ARC Prize 2025: Technical Report". *arXiv:2601.10904*, 2026. [10] 11. [↑](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-trm_paper_11-0) Jolicoeur-Martineau, A. "Less is More: Recursive Reasoning with Tiny Networks". *arXiv:2510.04871*, 2025. [11] 12. ↑ [12.0](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi3_paper_12-0) [12.1](https://systems-analysis.info/eng/ARC-AGI-2#cite_ref-arcagi3_paper_12-1) ARC Prize Foundation. "ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence". *arXiv:2603.24621*, 2026. [12]