Pith. sign in

REVIEW 3 major objections 4 minor 56 references

JigShape: a jigsaw-puzzle benchmark with interlocking tab-and-blank pieces shows that vision-language models solve 4×4 puzzles only under supervision and collapse to near-random accuracy on 8×8 and larger grids.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A new VLM benchmark with interlocking jigsaw pieces shows frontier and fine-tuned vision-language models solve 4x4 puzzles but collapse to near random on 8x8 and larger grids.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A well-built benchmark whose headline 'scaling cliff' is confounded by an unexamined JSON scoring rule; the benchmark deserves peer review, the claim needs a fix. the 3 major comments →

arxiv 2607.27670 v2 pith:335BGFGF submitted 2026-07-30 cs.CV cs.AI

JigShape: Evaluating Visual-Geometric Reasoning in VLMs through Jigsaw Puzzles

Shawn Li , Wei Yang , Jike Zhong , Jiate Li , Jiawei Yang , You Qin , Ryan Rossi , Franck Dernoncourt
show 6 more authors
Roger Zimmermann Yue Wang Zhengzhong Tu Vicente Ordonez Mohit Bansal Yue Zhao
This is my paper
classification cs.CV cs.AI
keywords jigsaw puzzlesvision-language modelsgeometric reasoningspatial reasoning benchmarkconstraint satisfactionzero-shot evaluationscaling cliff
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

JigShape is a jigsaw puzzle benchmark built from interlocking tab-and-blank pieces, designed so every puzzle has exactly one correct arrangement. The paper uses it to test whether vision-language models can reason jointly about image content and geometric constraints. The headline finding is a “scaling cliff”: with supervised fine-tuning, models solve 4×4 grids almost perfectly, but on 8×8 and larger grids every model falls to near-random accuracy. The paper argues this means current architectures cannot keep satisfying geometric constraints as the number of pieces grows, and that scalable geometric reasoning remains an open challenge.

Core claim

The paper’s central claim is that current vision-language models do not perform scalable geometric constraint satisfaction. On the new benchmark, only one zero-shot frontier model beats the random baseline at 4×4, and its advantage disappears on 8×8; supervised fine-tuned models reach above 97% piece accuracy at 4×4 but drop below 5% at 12×12. Because the benchmark removes the ambiguity of rectangular cuts, the authors interpret this collapse as a reasoning failure rather than a labeling artifact. The shape-versus-no-shape ablation supports this interpretation: geometric constraints contribute most of the signal, and fine-tuned models rely on them almost exclusively.

What carries the argument

The tab-and-blank edge system: every internal piece boundary is tagged as a convex tab, a concave blank, or a flat boundary edge, and a valid placement requires complementary edges on adjoining pieces. This gives strong local compatibility constraints and, combined with visual content, guarantees unique ground truth; the same pieces without shape constraints serve as an ablation that isolates how much models use geometry.

Load-bearing premise

The central claim rests on treating every unparseable or truncated model output as a zero rather than as a formatting failure; if larger puzzles mostly trigger malformed responses, the observed “scaling cliff” may not be about geometry at all.

What would settle it

Count how often model outputs parse into valid JSON on 8×8 and 12×12 puzzles, and compare accuracy only among parseable outputs. If accuracies remain at chance while parse rates stay high, the cliff is a genuine geometric-reasoning failure; if parse rates drop sharply, part or all of the cliff is a generation or formatting artifact.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the scaling cliff holds, current VLM architectures cannot be relied upon for tasks that require assembling many locally constrained components, such as layout recovery or object assembly from fragments.
  • The benchmark supplies a controlled testbed for future models: a model that solves 8×8 or larger grids would demonstrate a genuine step in compositional geometric reasoning.
  • Supervised fine-tuning alone does not overcome the cliff; reweighting training data toward harder grids makes performance worse, suggesting the bottleneck is architectural rather than data distribution.
  • Because the no-shape ablation drops fine-tuned accuracy to near random, models trained on this task may overfit to geometric edge cues rather than integrating visual content.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper’s protocol scores unparseable JSON output as zero without reporting parse rates; if parse rates fall sharply on larger grids, part of the cliff could be an output-formatting artifact. Reporting parse rates would make the reasoning claim testable.
  • The tab-and-blank construction could transfer to other constrained-assembly tasks—tangram, tile layouts, or 3D block assembly—where local compatibility plus visual content defines a unique solution, offering a template for harder spatial-reasoning benchmarks.
  • A model that learns to use flat edges for corners and borders (as the fine-tuned models do) may solve 4×4 puzzles without general geometric reasoning, so the 4×4 result likely overstates what those models understand; interior-heavy larger grids are the real test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. JigShape introduces a benchmark for evaluating visual-geometric reasoning in VLMs using jigsaw puzzles with tab-and-blank interlocking pieces, claiming unambiguous ground truth via geometric constraints plus visual content. The benchmark spans four grid densities (4x4 to 16x16) with roughly 95K instances, plus a paired shape/no-shape ablation at 4x4. The authors evaluate five frontier VLMs (six configurations) zero-shot and two fine-tuned models. The headline findings are: only GPT-5.5 exceeds random baseline on 4x4; fine-tuned models exceed 97% PA on 4x4; and all models collapse on larger grids, interpreted as a 'scaling cliff' in geometric constraint satisfaction.

Significance. If the central claims hold, JigShape would be a valuable diagnostic resource: the tab-and-blank design is a thoughtful departure from rectangular-cut benchmarks, the paired shape/no-shape ablation is a principled way to isolate geometric-cue usage, and the position-based error analysis (corners vs. edges vs. interiors) gives actionable signal. The construction pipeline is described in detail and the dataset is being released, which are strengths. However, the paper's most consequential conclusion—that models exhibit a fundamental scaling cliff in geometric reasoning—is directly confounded by the evaluation protocol's treatment of invalid JSON outputs, and the current manuscript does not supply the data needed to separate a reasoning failure from an output-format artifact. The benchmark itself is useful, but the main interpretive claim is not yet supported.

major comments (3)
  1. [Appendix C; Table 4; Section 5.1] The 'scaling cliff' claim is load-bearing but confounded by the JSON parsing protocol. Appendix C states 'responses with invalid JSON receive zero accuracy on all metrics,' and Table 4 reports all metrics under this rule, yet parse success rates, output-token usage, and refusal/truncation rates are never reported. Larger grids require outputting N^2 piece–position pairs (64, 144, 256), so the probability of truncated or malformed JSON grows with grid size; a model with partially correct placements but format failures would be scored as zero, mechanically producing the observed cliff. The 8x8 GPT-5.5 PA of 4.37% is above the 1.56% random baseline, so 'near-random' is not established without a parse-rate breakdown. The authors must report per-model, per-grid parse success and, ideally, re-analyze with a tolerant parser or partial-credit scoring for valid-but-incompletely-parsed responses.
  2. [Appendix E.2; Section 3.1] The uniqueness verification is insufficient and potentially circular. Appendix E.2 describes a backtracking solver that uses exact edge signatures and 'achieves 100% accuracy,' but finding the constructed solution does not prove that no other arrangement satisfies the edge constraints. Since edge compatibility is binary (tab matches blank, flat matches flat), many tilings can satisfy local compatibility; visual content is not used by the solver. To substantiate 'unambiguous ground truth,' the authors need an exhaustive enumeration of edge-compatible arrangements or a formal argument (with the actual edge-type distributions) showing uniqueness. As written, the claim that every instance has a unique solution is not established by the reported experiment.
  3. [Abstract; Section 5.1; Conclusion] The narrative 'all models collapse on larger grids' and 'near-random performance on 8×8' is not consistent with the reported numbers. Table 4 shows the SFT models at 27.34% and 33.58% PA on 8x8 versus a 1.56% random baseline—more than an order of magnitude above chance. GPT-5.5 at 4.37% is also not literally near the 1.56% baseline. The paper should either quantify the cliff precisely or soften the 'near-random' characterization; otherwise the abstract overstates the result.
minor comments (4)
  1. [Appendix E.3; Table 10] The text says GPT-5.5 achieves 25% PA in the extreme60 condition, but Table 10 lists GPT-5.5 at 2.50% and Qwen SFT at 25.0%. The numbers and the prose appear to be swapped; please correct.
  2. [Appendix C] The evaluation protocol should specify how duplicate piece positions, missing entries, and extra fields in otherwise valid JSON are handled. This matters for fair scoring because large outputs are prone to partial omissions even when the JSON parses.
  3. [Section 3.4] The train split is 91,968 instances but the SFT protocol in Section 5 uses only 6.5K samples. Clarify whether the remaining train instances are intended for future use and why the SFT subset is the relevant training condition.
  4. [Section 5.3 / Table 7] The position-based error analysis is informative, but the table would benefit from confidence intervals or significance tests; with 250 eval samples the corner vs. interior gaps are large, but error bars would aid comparison.

Circularity Check

0 steps flagged

No significant circularity: the evaluation is an external benchmark with independent baselines and no fitted parameters feeding the conclusions.

full rationale

The paper's derivation chain is: generate puzzle instances with random tab/blank polarity from public images (Sec. 3); define metrics against the known ground-truth permutation (Sec. 4.1 and Appendix D); measure five frontier VLMs and two SFT models (Sec. 5); and interpret the cross-grid drop as a scaling cliff. No parameter is fitted from the model outputs into the benchmark, and the SFT training mix is not tuned to the eval results. The random baselines are derived analytically from the same edge statistics as the generator (Appendix D), which is the correct null-model construction, not circularity. The uniqueness check (Appendix E.2) uses a classical backtracking solver independent of the VLMs; the statement that 100% solver accuracy 'confirms' uniqueness is logically under-supported, but that is a validation gap, not a circular inference. The self-citations (Li et al. 2026a-c, Shawn et al. 2025) appear only in related work and do not support the load-bearing claims. The strongest validity threat is Appendix C's rule that 'responses with invalid JSON receive zero accuracy on all metrics' combined with no parse-rate or token-usage reporting: on 8x8 and larger grids the required JSON has O(N^2) entries, so format failures could mechanically create the observed collapse. That would undermine the scaling-cliff conclusion, but it is an experimental-validity concern rather than a circularity, because the reported numbers are still external measurements and not defined by the paper's assumptions. The no-shape ablation's use of the original arrangement as ground truth could also penalize visually valid alternatives, but again this is a validity issue. Accordingly, no step reduces to its own inputs, and the circularity score is 0.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

No new physical or conceptual entities are introduced. The tab, blank, and flat edge types are design elements borrowed from real jigsaw puzzles, not invented theoretical constructs. The only hand-chosen quantities are the tab ratios and training mix, which are explicit design decisions rather than hidden fitted parameters.

free parameters (2)
  • Tab ratio = 0.15 (4x4), 0.20 (8x8), 0.24 (12x12), 0.28 (16x16)
    Chosen by hand so tabs remain visually perceptible as piece size shrinks (Sec. 3.3); affects puzzle difficulty and could influence model performance, but is not fitted to model scores.
  • Mixed-grid training composition = 4x4:3K, 8x8:2K, 12x12:1K, 16x16:0.5K (6.5K total)
    Training distribution chosen by the authors (Sec. 5); fine-tuning results are conditional on this schedule, but it is an experimental design choice rather than a fitted constant.
axioms (3)
  • domain assumption Each JigShape instance has a unique solution under tab-blank constraints.
    Stated in Sec. 3.1 and claimed verified in Appendix E.2 by a constraint solver; the appendix reports the solver matches ground truth on all test instances but does not show an exhaustive uniqueness proof.
  • domain assumption Model outputs must be parseable JSON; unparseable responses are scored zero.
    Appendix C sets this protocol; it is load-bearing for all accuracy numbers, especially on large grids, but is never validated against parse-rate statistics.
  • standard math Random baselines for AA and SC assume uniform random placement sampling without replacement.
    Baseline derivations in Appendix D rely on standard combinatorics with a stated without-replacement correction, and Monte Carlo simulation is cited as confirmation.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of JigShape: Evaluating Visual-Geometric Reasoning in VLMs through Jigsaw Puzzles." pith.science (2026). https://pith.science/paper/335BGFGF

@misc{pith2026260727670,
  author       = {Pith},
  title        = {Pith review of: JigShape: Evaluating Visual-Geometric Reasoning in VLMs through Jigsaw Puzzles},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/335BGFGF}},
  note         = {Machine review of arXiv:2607.27670}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Jigsaw puzzle solving requires jointly reasoning about visual content and geometric constraints, yet existing benchmarks use rectangular cuts that create ambiguous ground truth in texture-repeated regions. We introduce \textit{\ours{}}, a benchmark with tab-and-blank interlocking pieces where geometric constraints provide strong local compatibility requirements that, combined with visual content, yield unambiguous ground truth. Across 95K instances at four grid densities (4$\times$4 to 16$\times$16), we find that \textbf{zero-shot VLMs largely lack geometric reasoning}: only one of five frontier models (GPT-5.5) exceeds random baseline on 4$\times$4 puzzles, while all others perform at chance level. While supervised fine-tuning achieves $>$97\% on 4$\times$4, \textbf{all models collapse on larger grids}: GPT-5.5 drops from 70\% to near-random on 8$\times$8, and even fine-tuned models fall below 5\% on 12$\times$12. This ``scaling cliff'' suggests current architectures cannot maintain consistent constraint satisfaction as the number of pieces increases. \ours{} establishes scalable geometric reasoning as an open challenge for vision-language models.

Figures

Figures reproduced from arXiv: 2607.27670 by Franck Dernoncourt, Jiate Li, Jiawei Yang, Jike Zhong, Mohit Bansal, Roger Zimmermann, Ryan Rossi, Shawn Li, Vicente Ordonez, Wei Yang, You Qin, Yue Wang, Yue Zhao, Zhengzhong Tu.

Figure 1
Figure 1. Figure 1: Overview of JigShape. Left: Rectangular cuts create ambiguous ground truth in repeated￾texture regions; our tab-and-blank constraints enforce local compatibility, yielding unambiguous solutions when combined with visual content. Right: Benchmark taxonomy with four grid densities (4×4 to 16×16) and shape vs. no-shape ablation for controlled evaluation. and ask models to predict the original arrangement. Jig… view at source ↗
Figure 2
Figure 2. Figure 2: Example 12×12 puzzle instance from JigShape. Left: shuffled layout with 144 labeled pieces showing tab-and-blank edge shapes. Right: original source image representing ground truth. Geometric constraints (tabs must match blanks) enforce local compatibility, while visual content disambiguates pieces with identical edge signatures. See Appendix F for examples at other grid sizes. 3.4 SOURCE IMAGES AND STATIS… view at source ↗
Figure 3
Figure 3. Figure 3: Visual similarity hint. Left: original scrambled pieces. Middle: pieces colored by cluster [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Example instances from JigShape. Left: shuffled puzzle layout with labeled pieces. Right: [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 36 canonical work pages

  1. [1]

    Efros , title =

    Carl Doersch and Abhinav Gupta and Alexei A. Efros , title =. Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages =

    work page

  2. [2]

    Yuille , title =

    Chen Wei and Lingxi Xie and Xutong Ren and Yingda Xia and Chi Su and Jiaying Liu and Qi Tian and Alan L. Yuille , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    work page

  3. [3]

    Essa , title =

    Unaiza Ahsan and Rishi Madhok and Irfan A. Essa , title =. 2019 IEEE Winter Conference on Applications of Computer Vision (WACV) , pages =

    work page 2019

  4. [4]

    Yingyi Chen and Xi Shen and Yahui Liu and Qinghua Tao and Johan A. K. Suykens , title =. Pattern Recognition Letters , volume =

    work page

  5. [5]

    CoRR , volume =

    Penghao Wu and Yushan Zhang and Haiwen Diao and Bo Li and Lewei Lu and Ziwei Liu , title =. CoRR , volume =. 2025 , eprinttype =

    work page 2025

  6. [6]

    OpenAI Technical Report , year=

  7. [7]

    Anthropic Technical Report , year=

    The Claude 3 Model Family: Opus, Sonnet, Haiku , author=. Anthropic Technical Report , year=

    work page

  8. [8]

    arXiv preprint arXiv:2312.11805 , year=

    Gemini: A Family of Highly Capable Multimodal Models , author=. arXiv preprint arXiv:2312.11805 , year=

    Pith/arXiv arXiv

  9. [9]

    Transactions of the Association for Computational Linguistics , year=

    Visual Spatial Reasoning , author=. Transactions of the Association for Computational Linguistics , year=

    work page

  10. [10]

    Chen, Boyuan and Xu, Zhuo and Kirmani, Sean and Ichter, Brian and Sadigh, Dorsa and Guibas, Leonidas and Xia, Fei , journal=

    work page

  11. [11]

    arXiv preprint arXiv:2310.19785 , year=

    What's ``up'' with vision-language models? Investigating their struggle with spatial reasoning , author=. arXiv preprint arXiv:2310.19785 , year=

    Pith/arXiv arXiv

  12. [12]

    Fu, Xingyu and Hu, Yushi and Li, Bangzheng and Feng, Yu and Wang, Haoyu and Lin, Xudong and Roth, Dan and Smith, Noah A and Ma, Wei-Chiu and Krishna, Ranjay , journal=

    work page

  13. [13]

    2025 , eprint=

    Jigsaw-Puzzles: From Seeing to Understanding to Reasoning in Vision-Language Models , author=. 2025 , eprint=

    work page 2025

  14. [14]

    arXiv preprint arXiv:2503.23064 , year=

    Ren, Yufan and Tertikas, Konstantinos and Maiti, Shalini and Han, Junlin and Zhang, Tong and S. arXiv preprint arXiv:2503.23064 , year=

    Pith/arXiv arXiv

  15. [15]

    Tang, Kexian and Gao, Junyao and Zeng, Yanhong and Duan, Haodong and Sun, Yanan and Xing, Zhening and Liu, Wenran and Lyu, Kaifeng and Chen, Kai , journal=

    work page

  16. [16]

    Xu, Zhuoning and Liu, Xinyan , journal=

    work page

  17. [17]

    Elkin, Gur and Shahar, Ofir Itzhak and Ben-Shahar, Ohad , journal=

    work page

  18. [18]

    Scarpellini, Gianluca and Fiorini, Stefano and Giuliari, Francesco and Morerio, Pietro and Del Bue, Alessio , booktitle=

    work page

  19. [19]

    Agustsson, Eirikur and Timofte, Radu , booktitle=

    work page

  20. [20]

    Gu, Shuhang and Lugmayr, Andreas and Danelljan, Martin and Fritsche, Manuel and Lamour, Julien and Timofte, Radu , booktitle=

    work page

  21. [21]

    2024 , note=

    Unsplash , author=. 2024 , note=

    work page 2024

  22. [22]

    Prentice-Hall , year=

    Applied Analysis , author=. Prentice-Hall , year=

    work page

  23. [23]

    Bai, Shuai and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and Song, Sibo and Dang, Kai and Wang, Peng and Wang, Shijie and Tang, Jun and others , journal=

  24. [24]

    arXiv preprint arXiv:2408.00118 , year=

    Gemma 2: Improving Open Language Models at a Practical Size , author=. arXiv preprint arXiv:2408.00118 , year=

    Pith/arXiv arXiv

  25. [25]

    Yang, An and others , journal=

  26. [26]

    Liu, Haotian and Li, Chunyuan and Wu, Qingyang and Lee, Yong Jae , journal=

  27. [27]

    Dai, Wenliang and Li, Junnan and Li, Dongxu and Tiong, Anthony Meng Huat and Zhao, Junqi and Wang, Weisheng and Li, Boyang and Fung, Pascale and Hoi, Steven , journal=

    work page

  28. [28]

    Li, Junnan and Li, Dongxu and Savarese, Silvio and Hoi, Steven , journal=

    work page

  29. [29]

    Lu, Pan and Bansal, Hritik and Xia, Tony and Liu, Jiacheng and Li, Chunyuan and Hajishirzi, Hannaneh and Cheng, Hao and Chang, Kai-Wei and Galley, Michel and Gao, Jianfeng , journal=

  30. [30]

    Yue, Xiang and Ni, Yuansheng and Zhang, Kai and Zheng, Tianyu and Liu, Ruoqi and Zhang, Ge and Stevens, Samuel and Jiang, Dongfu and Ren, Weiming and Sun, Yuxuan and others , journal=

  31. [31]

    CVPR , year=

    A Fully Automated Greedy Square Jigsaw Puzzle Solver , author=. CVPR , year=

    work page

  32. [32]

    CVPR , year=

    A Genetic Algorithm-Based Solver for Very Large Jigsaw Puzzles , author=. CVPR , year=

    work page

  33. [33]

    ECCV , year=

    Solving Square Jigsaw Puzzles with Loop Constraints , author=. ECCV , year=

    work page

  34. [34]

    Johnson, Justin and Hariharan, Bharath and van der Maaten, Laurens and Fei-Fei, Li and Zitnick, C Lawrence and Girshick, Ross , journal=

    work page

  35. [35]

    Hudson, Drew A and Manning, Christopher D , journal=

  36. [36]

    ICLR , year=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. ICLR , year=

  37. [37]

    ICML , year=

    Learning Transferable Visual Models From Natural Language Supervision , author=. ICML , year=

  38. [38]

    Thrush, Tristan and Jiang, Ryan and Bartolo, Max and Singh, Amanpreet and Williams, Adina and Kiela, Douwe and Ross, Candace , journal=

    work page

  39. [39]

    VALSE : A Task-Independent Benchmark for Vision and Language Models Centered on Linguistic Phenomena

    Parcalabescu, Letitia and Cafagna, Michele and Muradjan, Lilitta and Frank, Anette and Calixto, Iacer and Gatt, Albert. VALSE : A Task-Independent Benchmark for Vision and Language Models Centered on Linguistic Phenomena. ACL. 2022

    work page 2022

  40. [40]

    European Conference on Computer Vision (ECCV) , pages=

    Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles , author=. European Conference on Computer Vision (ECCV) , pages=

    work page

  41. [41]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Masked Jigsaw Puzzle: A Versatile Position Embedding for Vision Transformers , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    work page

  42. [42]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Liu, Jinyang and Teshome, Wondmgezahu and Ghimire, Sandesh and Sznaier, Mario and Camps, Octavia , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    work page 2024

  43. [43]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Domain Generalization by Solving Jigsaw Puzzles , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    work page

  44. [44]

    Transactions on Machine Learning Research , year=

    Jigsaw-R1: A Study of Rule-based Visual Reinforcement Learning with Jigsaw Puzzles , author=. Transactions on Machine Learning Research , year=

    work page

  45. [45]

    and Taati, Babak and Grzeszczuk, Radek , journal=

    Jeddi, Ahmadreza and Karaimer, Hakki Can and Nguyen, Hue and Wang, Zhongling and Zhao, Ke and Rajabi, Javad and Zhang, Ran and Goyal, Raghav and Derpanis, Konstantinos G. and Taati, Babak and Grzeszczuk, Radek , journal=. PuzzleCraft: Exploration-Aware Curriculum Learning for Puzzle-Based

    work page

  46. [46]

    Spatial-

    Liu, Yuhong and Zhang, Beichen and Zang, Yuhang and Cao, Yuhang and Xing, Long and Dong, Xiaoyi and Duan, Haodong and Lin, Dahua and Wang, Jiaqi , journal=. Spatial-

    work page

  47. [47]

    AAAI , author=

    Siamese-Discriminant Deep Reinforcement Learning for Solving Jigsaw Puzzles with Large Eroded Gaps , volume=. AAAI , author=. 2023 , month=

    work page 2023

  48. [48]

    AAAI , author=

    Panoptic Scene Graph Generation with Semantics-Prototype Learning , volume=. AAAI , author=. 2024 , month=. doi:10.1609/aaai.v38i4.28098 , number=

    doi:10.1609/aaai.v38i4.28098 2024

  49. [49]

    2023 , isbn =

    Li, Li and Wang, Chenwei and Qin, You and Ji, Wei and Liang, Renjie , title =. 2023 , isbn =. doi:10.1145/3581783.3611847 , booktitle =

    arXiv 2023

  50. [50]

    CVPR , month =

    Li, Shawn and Gong, Huixian and Dong, Hao and Yang, Tiankai and Tu, Zhengzhong and Zhao, Yue , title =. CVPR , month =. 2025 , pages =

    2025

  51. [51]

    ICCV , month =

    Secure On-Device Video OOD Detection Without Backpropagation , author=. ICCV , month =

  52. [52]

    Treble Counterfactual VLM s: A Causal Approach to Hallucination

    Shawn, Li and Qu, Jiashu and Song, Linxin and Zhou, Yuxiao and Qin, Yuehan and Yang, Tiankai and Zhao, Yue. Treble Counterfactual VLM s: A Causal Approach to Hallucination. EMNLP. 2025

    2025

  53. [53]

    Defenses Against Prompt Attacks Learn Surface Heuristics , author=. ACL. 2026

    2026

  54. [54]

    2026 , eprint=

    The Autonomy Tax: Defense Training Breaks LLM Agents , author=. 2026 , eprint=

    2026

  55. [55]

    2026 , eprint=

    Geometry over Density: Few-Shot Cross-Domain OOD Detection , author=. 2026 , eprint=

    work page 2026

  56. [56]

    2026 , eprint=

    FORTIS: Benchmarking Over-Privilege in Agent Skills , author=. 2026 , eprint=

    work page 2026

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.