Pith. sign in

REVIEW 3 major objections 4 minor 43 references

TraceViT: Grounded Trace Supervision for Visual Abstract Reasoning

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read The paper's central claim is that training a looped visual reasoner to follow grounded intermediate grid-state milestones—not just the final answer—improves abstract-reasoning accuracy, reaching 67.8% pass@2 on ARC-AGI-1 and 24.3% on ARC-AG

desk verdict Worth refereeing, but the trace-chain semantics are untested and the ARC-2 comparison is confounded. read the letter →

arxiv 2607.29586 v1 pith:NPFLKBUY submitted 2026-07-31 cs.CV cs.AI

classification cs.CVcs.AI
keywords AbstractionandReasoningCorpusvisualabstractloopedreasonertracesupervisiontransformationchainssoftalignmentobject-centricrepresentationsfew-shotgrid
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a looped visual reasoner—a compact network that refines a puzzle grid over several iterations—learns better when its intermediate iterations are supervised to match meaningful intermediate grid states, not just the final output. The authors obtain those intermediate states (the 'trace') by rewriting programmatic solutions of ARC puzzles into step-by-step transformation chains and verifying them, then feed each iteration both a compact reference built from the few-shot demonstrations and an object-centric summary of the current grid. Soft alignment enforces only the order of the milestones, letting the model allocate iterations freely. The empirical core is an interaction: trace supervision alone slightly hurts, but combined with this grounding it lifts ARC-AGI-1 pass@2 from 62.3% to 65.5% (Medium) and to 67.8% (Large), while on the harder ARC-AGI-2 it reaches 24.3%. A sympathetic reader would care because the approach improves a compact 11M–18M model without extra test-time compute, and because the memory-versus-workspace interaction is a reusable lesson for any iterative visual reasoner.

What carries the argument

The load-bearing mechanism is the semantically monotonic transformation chain: a sequence of intermediate grids obtained by taking programmatic implementations of ARC tasks and rewriting them so that every action is a single, nameable step toward the answer, then verifying the rewrites with automated gates and human filmstrip review. Around this, the architecture has three parts. The task-reference encoder turns the few-shot demonstrations into a compact set of tokens re-injected at every iteration. The object workspace runs slot attention over the current grid patches, carrying an object-centric summary across iterations. Soft trace alignment then compares the loop's decoded predictions wit

What would settle it

Take a set of ARC tasks and obtain independent intermediate-state annotations from human solvers; train the identical TraceViT architecture twice, once with the paper's program-rewritten chains and once with these human-validated chains. If pass@2 with human-verified chains does not exceed the final-only baseline (or does not track the claimed gains), the benefit is an artifact of the rewriting process rather than of trace supervision per se. More directly, if the model's intermediate predictions systematically follow program-rewritten milestones but diverge from human-annotated milestones in

Watch

Extended reading notes

Core claim

The paper's central discovery is a training principle with a twist: in a looped visual reasoner, where a shared core is applied repeatedly to refine grid predictions, supervising every iteration against a chain of intermediate grids—each produced by one meaningful action from input to output—is beneficial only when the loop is grounded. Grounding means that a task reference built from the demonstration pairs and an object workspace extracted from the current grid state are re-supplied at every iteration, so that pushing hidden states toward partial outputs does not erase the input information. With grounding, trace supervision lifts ARC-AGI-1 pass@2 from 62.3% to 65.5% (Medium) and to 67.8%

Load-bearing premise

The whole result rests on the assumption that the automatically generated intermediate-grid chains are semantically valid—each step a single meaningful action toward the correct output—since no automated check can fully verify step semantics and humans are the final judge.

Editorial extensions

If this is right

  • If the central claim holds, compact looped visual reasoners can be improved by process-level supervision without extra parameters or test-time cost, narrowing the gap to much larger LLM-based ARC solvers.
  • The trace–grounding interaction predicts that any auxiliary loss steering intermediate states toward task-specific targets in a recurrent network must be paired with externalized memory; otherwise it will impair performance.
  • The soft-alignment mechanism with skip penalties gives a recipe for supervising variable-length process targets when milestone count and iteration budget differ, applicable to other iterative refinement models.
  • The error decomposition shows that reranking already-generated candidate grids can recover at most about 8.5 points on ARC-AGI-1; the remaining tasks require the per-view solver itself to improve.
  • Because untraced instances still receive final-state supervision, the approach degrades gracefully for tasks without a programmatic decomposition, but those tasks get no trace benefit.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • This suggests a general principle: auxiliary losses that push intermediate hidden states toward interpretable targets act as information bottlenecks unless the network has a separate channel that preserves the input; the same lesson could transfer to other domains that refine outputs over multiple steps.
  • The method's ceiling is set by the quality of the chain-construction pipeline; a testable extension is to replace LLM-rewrite-plus-human-review with automatic semantic verification (e.g., checking each step is a known atomic operation), which would let the approach scale beyond ARC.
  • Because the test-time procedure is identical across compared systems, the TraceViT gains come from offline training signal rather than extra inference compute; an implication is that stronger per-view solvers, not more views, are the path to further gains—consistent with the paper's own coverage-error analysis.
  • The soft-trace-alignment objective, combining ideas from CTC and Soft-DTW, could be reused to supervise large language models or diffusion models where intermediate states are semantically meaningful but not time-aligned.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes TraceViT, a looped vision transformer for ARC trained with intermediate grid-state supervision. Transformation chains are obtained by rewriting and verifying programmatic task implementations, and each loop iteration is grounded by a task reference and an object workspace. A soft trace alignment loss supervises intermediate predictions in an order-only manner. The paper reports 67.8% pass@2 on ARC-AGI-1 and 24.3% on ARC-AGI-2, and a 2x2 ablation showing that trace supervision improves accuracy only when combined with grounding.

Significance. If the empirical claims hold, the paper makes a useful contribution: a scalable pipeline for generating intermediate-state annotations for ARC, and evidence that grounding enables trace supervision to improve a compact visual reasoner. The paper is generally well-written and candid about limitations; the controlled 2x2 design, the detailed corpus-construction audit, and the explicit error decomposition are strengths, as is the planned release of code and data. However, the central interaction rests on a single training seed and on a chain-validity premise that is not tested by any semantic-content control, so the empirical support is not yet fully convincing.

major comments (3)
  1. [Table 2; Appendix D (Table 4)] The central interaction is reported for a single training seed (seed 42) with no error bars. The key differences are small: trace+grounding vs grounding-only is 65.5 vs 63.9 (1.6 points), and trace-alone vs baseline is 61.6 vs 62.3. With binary pass@2 over 400 tasks, run-to-run variation across seeds is expected to be of the same order as these differences. Please report means and standard deviations over at least three seeds (or otherwise establish robustness), and state the number of independent test-time training runs used for Table 2. Without this, the claimed 'clear interaction' is not statistically supported.
  2. [§Constructing Transformation Chains; Appendix A.2–A.4] The load-bearing premise is that the GPT-5.5-rewritten chains are semantically monotonic and that each milestone is a correct intermediate state. The paper concedes (Appendix A.4) that no automated gate can assess step semantics and that human review is the final adjudicator. However, no experiment manipulates the semantic content of the chains. A control with permuted milestones, reversed chains, or randomly sampled intermediate states (keeping the same L_align loss and grounding) would determine whether the benefit in Table 2 comes from the semantic ordering/content or from an additional regularizing auxiliary loss. Since the corpus is not yet released, the chain validity cannot be independently checked. This control is essential to the paper's claim that 'grounded trace supervision' is effective.
  3. [§Comparison with Prior ARC Solvers (Table 1)] The ARC-AGI-2 comparison with LoopViT (24.3 vs 14.2) is not matched for training data or protocol. The compact-solver results are quoted from other papers; if LoopViT did not use the 877,962-record ARC-AGI-2 corpus or the ARC-AGI-1 warm start, the gain could be due to data scale rather than trace supervision. Please clarify the training data for the baselines and, ideally, add an ARC-AGI-2 ablation with the same architecture and data, with and without trace supervision, to show the benefit transfers beyond ARC-AGI-1.
minor comments (4)
  1. [Appendix D, Table 4] 'Independent runs 2' is ambiguous; specify whether it refers to two test-time training runs or two full training runs, and how the results are aggregated.
  2. [Eq. (10) and §Method] The final-state loss weighting uses w(i)=1+α[y(i)≠x(i)], treating the input as the predecessor of the output even for traced examples where the true predecessor is T_{K-1}. This is inconsistent with Eq. (9); please clarify or correct the choice.
  3. [Figure 2] The pipeline figure is dense; consider enlarging the alignment panel or pointing to the relevant appendix equations for the soft-min DP.
  4. [§Experiments] Please report the distribution of chain lengths K relative to the loop budget N=6, so readers can see how often the K>N regime of Appendix B actually occurs.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TraceViT's supervision targets come from externally generated program traces, and its predictions are evaluated on held-out ARC tasks.

full rationale

The paper's central claim—that grounded trace supervision improves a looped visual reasoner on ARC-AGI-1 and ARC-AGI-2—does not reduce to its inputs. The transformation chains are constructed from independent programmatic task implementations (RE-ARC and ARC-GEN) by GPT-5.5 rewriting, then checked by automated gates and human filmstrip review (Appendix A). These chains are fixed supervision targets for training; they are not defined in terms of the model's outputs, nor are any fitted model parameters embedded in the milestone grids. The final evaluation is on the official held-out ARC-AGI-1 and ARC-AGI-2 evaluation tasks, and Appendix D explicitly states that chain supervision is used only offline, with the same final-state-only test-time objective for all compared systems. The key 2x2 ablation in Table 2 holds the input–output training set and data pipeline fixed and varies only whether the L_align milestone term and the grounding structures are enabled; the trace signal is therefore isolated rather than constructed to guarantee the reported interaction. The cited backbone LoopViT is an architectural starting point and comparison baseline, not a self-citation invoked to justify the trace-supervision result. The admitted reliance on human review of chain semantics (Appendix A.4) is a correctness/validity risk, not a circularity risk: even if some chains were semantically imperfect, the target grids are still externally derived program states rather than restatements of the model's predictions. No equation in the paper equates a predicted quantity with a fitted parameter or defines a supervision target in terms of the model's own output. The derivation chain is therefore self-contained with respect to the stated claim, and the appropriate circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The central claim rests on the reliability of LLM-generated chains, on architectural assumptions about grounding, and on hyperparameters chosen without sensitivity analysis.

free parameters (6)
  • change weight alpha = 3.0
    Eq. (9)-(10), Table 4; chosen by hand to upweight changed cells in milestone and final-state costs; no sensitivity analysis reported.
  • alignment weight beta = 0.2 (Medium) / 0.3 (Large)
    Eq. (8), Table 4; weights the trace-alignment loss; chosen by hand.
  • softmin temperature gamma = 0.5
    Eq. (5)-(6), Table 4; controls sharpness of path marginalization; chosen by hand.
  • skip penalty lambda_skip = 0.3
    Eq. (11)-(13), Appendix B, Table 4; cost per skipped milestone when K>N; chosen by hand.
  • loop iterations N = 6
    Appendix C.1; fixed unroll depth; central to alignment capacity and all reported results.
  • final-state loss weight lambda_out = 2.0
    Table 4; weights the final-output loss in the objective; chosen by hand.
assumptions (4)
  • standard math Soft-min dynamic programming (Eqs. 6, 13) is differentiable and yields valid posterior occupancies dF/dC.
    Inherited from CTC/Soft-DTW; the paper cites Graves et al. 2006 and Cuturi & Blondel 2017.
  • domain assumption GPT-5.5 rewrites of programmatic implementations preserve task semantics and training distribution after automated gates and human review.
    Appendix A.2-A.4; the paper explicitly relies on human review as final adjudicator, which is not independently machine-verified.
  • domain assumption The task reference and object workspace externalize memory sufficiently so that intermediate supervision does not destroy input information.
    Method section 'Grounding the Loop'; the 2x2 ablation is the only empirical support for this architectural assumption.
  • domain assumption Chain milestones are orderable and the K<=N / K>N alignment regimes cover the actual distribution of chain lengths.
    Eq. (6) and Appendix B assume monotonic paths and skip penalties match real chains; no corpus statistics for K are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TraceViT: Grounded Trace Supervision for Visual Abstract Reasoning." pith.science (2026). https://pith.science/paper/NPFLKBUY

@misc{pith2026260729586,
  author       = {Pith},
  title        = {Pith review of: TraceViT: Grounded Trace Supervision for Visual Abstract Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NPFLKBUY}},
  note         = {Machine review of arXiv:2607.29586}
}
read the original abstract

The Abstraction and Reasoning Corpus (ARC) tests whether a model can infer an unseen transformation from a few input-output examples and apply it to a new grid. Looped visual reasoners refine predictions over multiple iterations, but conventional training constrains only the final output, leaving intermediate refinements unconstrained. We propose that these refinements should instead follow the transformation step by step. We introduce TraceViT, a looped visual reasoner trained with semantically monotonic transformation chains. We obtain these chains by rewriting and verifying programmatic task implementations, decomposing each solution into intermediate grid states. Each iteration is grounded by a task reference derived from the few-shot demonstrations and an object workspace representing the current grid state. Because these chains may differ in length from the loop, soft trace alignment enforces only their ordering, letting the model allocate iterations freely. TraceViT achieves 67.8% pass@2 on ARC-AGI-1 and 24.3% on ARC-AGI-2. Controlled ablations on ARC-AGI-1 show that trace supervision becomes beneficial only when paired with grounding. Code and data will be available at https://github.com/LiuBinnan/TraceViT.

Figures

Figures reproduced from arXiv: 2607.29586 by the authors.

Figure 1
Figure 1. Grounded trace supervision. Left: an ARC task provides demonstrations and a test input. Right: TraceViT it￾erates a shared visual core, decoding a grid at every step. Dur￾ing training, intermediate predictions are aligned with trans￾formation chain milestones. See [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed method. Top: the few-shot demonstrations are encoded into a compact task-reference prefix G (left) via cross-attention from learned queries Q to the joint demonstration context X; a looped visual reasoner (center) iterates a shared core F for N steps. Iteration t receives the static reference G and the dynamic slot bank St−1; the object workspace (right) extracts St from the updated state fo… view at source ↗
Figure 3
Figure 3. Per-step predictions and slot assignments on three tasks. Prediction rows show the input and six loop outputs. Assignment rows show the argmax over eight slots at each step. Process Visualization Evolution of intermediate predictions [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Failure cases for trajectory-wide recoloring. Official pairs establish the task semantics. Records labeled identity or production preserve those semantics, whereas each improper recolor applies a consistent palette mapping across the full trajectory but changes the rul…
Figure 5
Figure 5. Figure 5: complements the three examples in the main paper with three additional tasks, showing how predictions and workspace assignments evolve across loop iterations. input step 1 step 2 step 3 step 4 step 5 output [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 7 linked inside Pith

  1. [1]

    Aky \" u rek, E.; Damani, M.; Zweiger, A.; Qiu, L.; Guo, H.; Pari, J.; Kim, Y.; and Andreas, J. 2025. The Surprising Effectiveness of Test-Time Training for Few-Shot Learning. In Proceedings of the Forty-Second International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research. PMLR

  2. [2]

    Alford, S.; Gandhi, A.; Rangamani, A.; Banburski, A.; Wang, T.; Dandekar, S.; Chin, J.; Poggio, T.; and Chin, P. 2021. Neural-guided, Bidirectional Program Search for Abstraction and Reasoning. In Complex Networks & Their Applications X , 657--668. Springer

  3. [3]

    ARC Prize Foundation . 2025. ARC-AGI Benchmarking: Leaderboard for the ARC-AGI Benchmark. https://arcprize.org/leaderboard. Accessed: 2026-07-25

  4. [4]

    Chollet, F. 2019. On the Measure of Intelligence. arXiv:1911.01547

  5. [5]

    Chollet, F.; Knoop, M.; Kamradt, G.; Landers, B.; and Pinkard, H. 2025. ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems. arXiv:2505.11831

  6. [6]

    Cole, J.; and Osman, M. 2025. Don't throw the baby out with the bathwater: How and why deep learning for ARC. arXiv:2506.14276

  7. [7]

    Cuturi, M.; and Blondel, M. 2017. Soft-DTW : A Differentiable Loss Function for Time-Series. In Proceedings of the Thirty-Fourth International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, 894--903. PMLR

  8. [8]

    Ferr \'e , S. 2024. Tackling the Abstraction and Reasoning Corpus ( ARC ) with Object-centric Models and the MDL Principle. In Advances in Intelligent Data Analysis XXII , volume 14641 of Lecture Notes in Computer Science, 3--15. Springer

Show all 43 references
  1. [9]

    Franzen, D.; Disselhoff, J.; and Hartmann, D. 2025 a . The ARChitects - Technical Report: ARC Prize 2025 Solution Summary. https://lambdalabsml.github.io/ARC2025_Solution_by_the_ARChitects/. Accessed: 2026-06-17

  2. [10]

    Franzen, D.; Disselhoff, J.; and Hartmann, D. 2025 b . Product of Experts with LLMs: Boosting Performance on ARC Is a Matter of Perspective. In Proceedings of the Forty-Second International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research. PMLR

  3. [11]

    Gao, G.; Schölkopf, B.; and Geiger, A. 2026. Slots, Transitions, Loops: Learning Composable World Models for ARC. arXiv:2606.12316

  4. [12]

    Graves, A.; Fern \'a ndez, S.; Gomez, F.; and Schmidhuber, J. 2006. Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks. In Proceedings of the Twenty-Third International Conference on Machine Learning, volume 148 of ACM Int...

  5. [13]

    Greenblatt, R. 2024. Getting 50\ https://redwoodresearch.substack.com/p/getting-50-sota-on-arc-agi-with-gpt. Accessed: 2026-06-17

  6. [14]

    Hodel, M. 2024 a . Addressing the Abstraction and Reasoning Corpus via Procedural Example Generation. arXiv:2404.07353

  7. [15]

    Hodel, M. 2024 b . Domain Specific Language for the Abstraction and Reasoning Corpus. https://github.com/michaelhodel/arc-dsl. Accessed: 2026-06-17

  8. [16]

    D.; Wang, R.; Zhu, Y

    Hu, K.; Cy, A.; Qiu, L.; Ding, X. D.; Wang, R.; Zhu, Y. E.; Andreas, J.; and He, K. 2025. ARC Is a Vision Problem! arXiv:2511.14761

  9. [17]

    Jia, B.; Liu, Y.; and Huang, S. 2023. Improving Object-centric Learning with Query Optimization. In Proceedings of the Eleventh International Conference on Learning Representations

  10. [18]

    Jolicoeur-Martineau, A. 2025. Less is More: Recursive Reasoning with Tiny Networks. arXiv:2510.04871

  11. [19]

    Kim, S.; Choi, H.; Lee, S.; and Kim, S. 2026. ARCTraj: A Dataset and Benchmark of Human Reasoning Trajectories for Abstract Problem Solving. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2687--2698. New York, NY, USA: ACM

  12. [20]

    P.; and Ba, J

    Kingma, D. P.; and Ba, J. 2015. Adam: A Method for Stochastic Optimization. In Proceedings of the Third International Conference on Learning Representations

  13. [21]

    K.; Lake, B

    LeGris, S.; Vong, W. K.; Lake, B. M.; and Gureckis, T. M. 2025. A Comprehensive Behavioral Dataset for the Abstraction and Reasoning Corpus. Scientific Data, 12(1): 1380

  14. [22]

    Lehmann, J.; Khushbakht, S.; Salehfard, N.; Nishat, N. A. Z.; Bhandiwad, D.; Aioanei, A.; and Vahdati, S. 2026. ARC-TGI: Human-Validated Task Generators with Reasoning Chain Templates for ARC-AGI. arXiv:2603.05099

  15. [23]

    Lei, C.; Lipovetzky, N.; and Ehinger, K. A. 2024. Generalized planning for the abstraction and reasoning corpus. In Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence , volume 38, 20168--20175

  16. [24]

    Li, W.; Xu, Y.; Sanner, S.; and Khalil, E. B. 2025 a . Tackling the Abstraction and Reasoning Corpus with Vision Transformers: the Importance of 2D Representation, Positions, and Objects. Transactions on Machine Learning Research, 2025

  17. [25]

    Li, W.-D.; Hu, K.; Larsen, C.; Wu, Y.; Alford, S.; Woo, C.; Dunn, S.; Tang, H.; Zheng, W.-L.; Pu, Y.; et al. 2025 b . Combining Induction and Transduction for Abstract Reasoning. In Proceedings of the Thirteenth International Conference on Learning Representations

  18. [26]

    Liu, Z.; and Shang, Y. 2026. Reasoning is a Modality. arXiv:2601.13562

  19. [27]

    Locatello, F.; Weissenborn, D.; Unterthiner, T.; Mahendran, A.; Heigold, G.; Uszkoreit, J.; Dosovitskiy, A.; and Kipf, T. 2020. Object-Centric Learning with Slot Attention. In Advances in Neural Information Processing Systems 33 (NeurIPS 2020), 11525--11538. New York, NY, USA:...

  20. [28]

    McGovern, R. K. 2025. Test-time Adaptation of Tiny Recursive Models. arXiv:2511.02886

  21. [29]

    Moffitt, M. D. 2025. ARC-GEN: A Mimetic Procedural Benchmark Generator for the Abstraction and Reasoning Corpus. arXiv:2511.00162

  22. [30]

    S.; Cohen, J

    Mondal, S. S.; Cohen, J. D.; and Webb, T. W. 2024. Slot Abstractors : Toward Scalable Abstract Visual Reasoning. In Proceedings of the Forty-First International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, 36088--36105. PMLR

  23. [31]

    Ouellette, S. 2024. Towards Efficient Neurally-Guided Program Induction for ARC-AGI. arXiv:2411.17708

  24. [32]

    Pourcel, J.; Colas, C.; and Oudeyer, P. 2025. Self-Improving Language Models for Evolutionary Program Synthesis: A Case Study on ARC-AGI . In Proceedings of the Forty-Second International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research. PMLR

  25. [33]

    Roye-Azar, A.; Vargas-Naranjo, S.; Ghai, D.; Balamurugan, N.; and Amir, R. 2026. Tiny Recursive Models on ARC-AGI-1: Inductive Biases, Identity Conditioning, and Test-Time Compute. arXiv:2512.11847

  26. [34]

    H.; Liu, Y.; and Yang, H

    Shu, W.-J.; Qiu, X.; Zhu, R.-J.; Chen, H. H.; Liu, Y.; and Yang, H. 2026. LoopViT: Scaling Visual ARC with Looped Transformers. arXiv:2602.02156

  27. [35]

    Singhal, K.; and Shroff, G. 2025. ConceptSearch : Towards Efficient Program Search Using LLMs for Abstraction and Reasoning Corpus ( ARC ). In Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence , volume 39, 20506--20513. AAAI Press

  28. [36]

    Sorokin, I.; and Puget, J.-F. 2025. NVARC solution to ARC-AGI-2 2025. https://github.com/1ytic/NVARC. Accessed: 2026-06-17

  29. [37]

    Strandgaard, S. 2026. ARC-Interactive-History-Dataset. https://github.com/neoneye/ARC-Interactive-History-Dataset. Accessed: 2026-06-17

  30. [38]

    Wang, G.; Li, J.; Sun, Y.; Chen, X.; Liu, C.; Wu, Y.; Lu, M.; Song, S.; and Yadkori, Y. A. 2025. Hierarchical Reasoning Model. arXiv:2506.21734

  31. [39]

    Wang, W.; and Reid, F. 2026. Tiny Recursive Reasoning with Mamba-2 Attention Hybrid. arXiv:2602.12078

  32. [40]

    Wind, J. S. 2020. DSL solution to the ARC challenge. https://github.com/top-quarks/ARC-solution. Accessed: 2026-06-17

  33. [41]

    B.; and Sanner, S

    Xu, Y.; Khalil, E. B.; and Sanner, S. 2023. Graphs, constraints, and search for the abstraction and reasoning corpus. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence , volume 37, 4115--4122. Menlo Park, Calif.: AAAI Press

  34. [42]

    Xu, Y.; Li, W.; Vaezipoor, P.; Sanner, S.; and Khalil, E. B. 2024. LLMs and the Abstraction and Reasoning Corpus: Successes, Failures, and the Importance of Object-based Representations. Transactions on Machine Learning Research, 2024

  35. [43]

    Zhang, B.; Zang, Y.; Dong, X.; Cao, Y.; Duan, H.; Lin, D.; and Wang, J. 2025. Think Visually, Reason Textually: Vision-Language Synergy in ARC. arXiv:2511.15703

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.