Pith. sign in

REVIEW 3 major objections 6 minor 63 references

Beyond Task-Specific Reasoning: A Unified Conditional Generative Framework for Abstract Visual Reasoning

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Four classic abstract visual reasoning tasks reduce to one conditional probability, and a single trained model answers all of them.

desk verdict The paper's real contribution is the empirical zero-shot transfer of a single RPM-trained model to other AVR task formats; the theoretical reduction is elementary and Proposition 3.5 has a genuine proof gap. read the letter →

arxiv 2507.11761 v1 pith:IGLXYZKD submitted 2025-07-15 cs.CV cs.AI

Fan Shi , Bin Li , Xiangyang Xue This is my paper
classification cs.CVcs.AI
keywords abstractvisualreasoningconditionalgenerativemodelRaven'sProgressiveMatriceszero-shotanalogyproblemodd-one-outSyntheticTaskdiscreterepresentationlearning
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 claims that four classic abstract visual reasoning tasks—Raven's Progressive Matrices, Visual Analogy Problems, Odd-One-Out, and Synthetic Visual Reasoning Tasks—can all be reduced to a single operation: estimating the conditional probability $p(x\mid I^p_{\neg i})$ that a candidate image fits a problem panel given the rest of the panel. If that reduction is correct, the same trained conditional generative model can generate answers and select among candidates, and it can transfer to task formats it never saw during training. The paper proves this equivalence for the four task types and instantiates it in a transformer-based model that predicts target image patches from context patches. Trained once on RAVEN and PGM, the model reaches 64.6% and 38.1% accuracy on those in-distribution tests and shows partial zero-shot transfer on new task formats, while out-of-distribution zero-shot performance falls to barely above random. That contrast marks the boundary of the claimed generalization.

What carries the argument

The carrying object is the predictability estimator, a conditional generative network trained to approximate $p(x\mid I^p_{\neg i})$. The argument works because the estimator factorizes the target image into $M$ discrete patch codes and models $p(Z_t\mid Z_C) = \prod_{m=1}^{M} p(Z_{t,m}\mid Z_{t,<m}, Z_C)$, so the same network produces a likelihood for candidate images and a sampler for generated answers. The instantiation UCGS-T has five modules: a VQ-VAE image encoder/decoder mapping images to and from a fixed codebook; a patch encoder with learnable class tokens that extract $K$ visual concepts per image; a concept encoder that groups concepts across context images and predicts target concepts with a shared transformer decoder; and an autoregressive patch decoder that turns predicted concepts into patch codes. The judgment functions—argmax for RPM/VAP, argmin for O3, argmax over panels for SVRT—are task-specific wrappers around this one shared score.

What would settle it

A concrete check: train UCGS-T on RAVEN and PGM as described, then evaluate it on a VAP variant whose source row uses an XOR rule and whose target row uses a progression rule over the same visual concepts. If selection accuracy on this mixed-rule VAP stays near the 25% random baseline, the learned conditional probability is not a general rule-compliance oracle and the reduction of Proposition 3.3 does not transfer beyond the rule structures seen in training.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes a reduction claim: for any abstract visual reasoning problem whose panel is a set of images with one distinguished slot, the correct answer is the image that maximizes the conditional predictability $p(x\mid I^p_{\neg i})$, the odd-one-out is the image that minimizes it, and an SVRT query belongs to the panel for which it is most predictable (Propositions 3.3–3.5). Because these judgments are expressed purely through one score, answer selection and answer generation are the same operation; the model can sample from $p(x\mid I^p_{\neg i})$ to create completions and evaluate candidates by the same likelihood. The paper instantiates this as UCGS-T, which first compresses each image into discrete patch codes, extracts $K$ visual concepts per image via slot attention, combines context concepts to predict target concepts, then autoregressively decodes target patches. The empirical payload is that a single multi-task training run on RAVEN and PGM produces 64.6% on RAVEN, 38.1% on PGM, and transfer to three task formats constructed from RAVEN concepts, with accuracy falling to near-random on the out-of-distribution VAP and SVRT splits.

Load-bearing premise

The load-bearing premise is that the conditional probability that an image fits a panel, learned from RAVEN and PGM training, remains a faithful measure of rule-compliance when the panel layout, task format, or rule vocabulary is new.

Editorial extensions

If this is right

  • One round of multi-task training on RAVEN and PGM yields a single model that answers RPM, analogy, odd-one-out, and categorization problems, so solving a new AVR task format no longer requires retraining or architecture changes when concepts and rules are in distribution.
  • Since answer generation and answer selection share one likelihood, any improvement to the generative model—better image tokenizer, decoder, or concept encoder—should improve both completion quality and selection accuracy.
  • The near-random out-of-distribution results (28.8% on VAP vs 25% random, 52.8% on SVRT vs 50% random) imply that the learned predictability is not a universal rule oracle; the framework's zero-shot reach is bounded by the visual concepts and abstract rules present in training.
  • The framework transfers across panel structures within the training rule vocabulary, as shown by UCGS-T scoring 35.8% on six-image VAP-ID and 33.6% on five-image O3-ID after training on nine-image RPM panels.

Reading between the lines

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

  • A testable extension the paper leaves implicit: apply the same conditional likelihood to Bongard-style concept learning, where the context is a set of positive and negative exemplars and the score decides whether a query image fits the positive class.
  • The sharp out-of-distribution drop suggests the model learns a dataset-specific rule vocabulary rather than abstract rule compliance; a calibration study comparing $p(x\mid \text{context})$ against human rule-compliance judgments could settle which.
  • Because the framework requires only a panel of images with one target, it can be extended to multi-cell completion by factorizing over several missing positions, which would connect it to image-inpainting formulations.
  • The discrete codebook is likely the bottleneck for real-world transfer; swapping in a higher-capacity image tokenizer while freezing the concept and panel encoders would test whether the reasoning layers, not the image compression, limit out-of-distribution generalization.
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 / 6 minor

Summary. Shi et al. propose UCGS, a framework that reduces four abstract visual reasoning benchmarks (RPM, VAP, O3, SVRT) to estimating the conditional predictability p(x|context) of a target image given the remaining panel images. The paper presents Propositions 3.3–3.5 as proofs of this reduction, then instantiates the framework in a VQ-VAE plus Transformer model (UCGS-T) trained once on RAVEN and PGM and applied without retraining to all evaluation tasks. Experiments report 64.6% on RAVEN, 38.1% on PGM, mixed results on three constructed in-distribution zero-shot tasks (O3-ID, VAP-ID, SVRT-ID), and out-of-distribution zero-shot accuracies that are close to chance on G1-set, VAP, and SVRT.

Significance. If the reduction were rigorously established, the paper would offer a genuinely unified treatment of four distinct AVR benchmarks with a single conditional generative model, and the empirical comparison (64.6% RAVEN vs. 54.5% RAISE) shows the instantiation is competitive among generative RPM solvers. The construction of ID-ZS tasks from RAVEN is a useful methodology for probing task-format generalization. However, the proof of Proposition 3.5 conflates dataset sampling probabilities with model-assigned panel likelihoods, and the OOD zero-shot results are modest, with SVRT at 52.8% versus 50% random on 253 test problems. These issues are fixable, but they currently limit the strength of the paper's central claims.

major comments (3)
  1. [§3.1, Proposition 3.5] The proof of the SVRT reduction is invalid as written. Definition 3.1 defines p(I^p) as the joint probability or 'correctness' of a panel under the generative model, but the proof then asserts p(I^l)=p(I^r)=1/|D| because the panels are 'sampled from the dataset D uniformly without overlap.' The marginal likelihood assigned by a trained generative model is not the empirical sampling frequency; nothing in the training objective (Eq. 9) enforces equal likelihoods for all panels in D. Consequently, argmax_{I^p} p(x|I^p) is not equivalent to argmax_{I^p} p(I^p_+x) unless the panel priors are equal or the model is calibrated to the uniform sampling distribution. The reduction can be repaired by stating an explicit uniform prior over the two panels in the SVRT decision rule, or by treating the rule as a heuristic, but as written the proof does not go through.
  2. [§3.1, Proposition 3.4] Equation (2) writes p(I^p_¬i)=p(I^p)/p(I^p_i|I^p_¬i) and then argues that removing the odd image leaves a 'large' panel probability. This requires p(I^p_i|I^p_¬i)>0; if the model assigns zero conditional probability to the odd image, the expression is an indeterminate 0/0 form. The proof also relies on an informal threshold ('large' vs. '→0') without specifying how to compare panels of different sizes. Please state regularity conditions (strict positivity of conditional probabilities and a separation between rule-compliant and rule-violating panels) or recast Proposition 3.4 as a heuristic decision rule rather than a theorem.
  3. [§4.1, Table 1] The zero-shot claims rest on single-run accuracies with no error bars or significance tests. The OOD gains are modest: G1-set 30.4% vs. 22.5% random, VAP 28.8% vs. 25% random, and SVRT 52.8% vs. 50% random. With only 253 SVRT test problems, the SVRT result is within binomial sampling noise (standard error approximately 3.1 percentage points). Please report multiple seeds and confidence intervals, or at least a binomial test for SVRT, to support the claim that UCGS exhibits zero-shot reasoning on unseen tasks. The near-chance SVRT result in particular weakens that claim as currently stated.
minor comments (6)
  1. [Table 1, random-guess row] The random-guess baseline for O3-ID is listed as 11.1%, but with five images and one odd image the chance level is 20%. Please correct the table and any related discussion.
  2. [Appendix B.1, Patch Decoder] The text says 'The panel encoder is a 12-layer Transformer decoder' in the Patch Decoder paragraph; this appears to be a copy-paste error and should refer to the patch decoder.
  3. [Figure 2 and §3.2.3] The terms 'concept encoder' and 'panel encoder' are used interchangeably across Figure 2, the section heading, and the body text; please standardize the terminology.
  4. [§4.1.4] The GPT-4o comparison uses subsets of different sizes (700, 1K, 253) without confidence intervals; a brief note on sampling variability would improve the comparison.
  5. [§3.1, Propositions 3.3–3.5] The propositions are stated as mathematical proofs, but they rely on the true data distribution ranking rule-compliant panels above rule-violating ones. Since the model is trained on a finite sample, this is an empirical assumption; consider stating it explicitly as a condition rather than leaving it implicit.
  6. [General] The paper does not state whether code or trained models will be released. For reproducibility, please provide code or clearly state the release plan.

Circularity Check

2 steps flagged · score 6.0 of 10

The reduction proofs are consequences of the paper's own definition of correctness as joint probability, making the theoretical unification true by construction; Proposition 3.5 adds an unsupported equality between model probability and sampling frequency.

  1. self definitional [Section 3.1, Definition 3.1 and Proposition 3.3 (proof pattern reused in 3.4 and 3.5)]
    "The correctness of I p is the joint probability p(I p). I p is rule-compliant if p(I p) is large, and rule-violating if p(I p) → 0. ... The correct answer x⋆ make up the modified panel with the largest correctness, given by argmax x∈I s p(x|I p ¬N )."

    The claimed reduction is the definition restated: 'correctness' is stipulated to be the joint probability, so the statement that the right completion maximizes p(x|context) (and the odd image minimizes it) adds no independent content. Eq. (1) is just the chain rule; the identification of the dataset's correct answer with 'largest correctness' is assumed, not derived. Consequently, the 'proof' that RPM/VAP/O3 reduce to predictability estimation is true by construction.

  2. other [Section 3.1, Proposition 3.5]
    "Consider that I l and I r are sampled from the dataset D uniformly without overlap, in Equation 3 we have p(I l) =p(I r) = 1/|D|. Therefore, x belongs to the panel I ⋆ = argmaxI p∈{I l,I r} p(x|I p)."

    The proof switches the meaning of p: Definition 3.1 defines p(I^p) as the model's joint 'correctness' probability, but here p(I^l)=1/|D| is the empirical sampling frequency. No argument or experiment shows the learned panel likelihood is calibrated to the uniform dataset prior. The equivalence argmax p(x|I^p) = argmax p(I^p_{+x}) required for SVRT depends entirely on this unproved equality, so the reduction is not established by the framework.

full rationale

The paper's central theoretical contribution—that RPM, VAP, O3, and SVRT reduce to estimating p(x|context)—is largely self-definitional: Definition 3.1 equates panel correctness with joint probability, so Propositions 3.3 and 3.4 follow immediately by construction. Proposition 3.5 has an additional unsupported identification of the model's marginal probability with the empirical sampling frequency, which is a proof gap rather than a separate circularity. These issues affect the unification proof, not the empirical instantiation: UCGS-T is trained with a standard likelihood objective on RAVEN/PGM and evaluated on external benchmarks (G1-set, VAP, SVRT), so the main experimental results are independent of the tautological proofs. The authors' self-citations (Shi et al. 2021, 2023, 2024) are background or architectural inspiration and are not load-bearing fitted constants. Overall, the reduction claims are partially circular by definition, but there is no fitting-renamed-as-prediction and no self-citation chain.

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

The central claim rests on four main axioms: the likelihood-correctness correspondence, the uniform-prior assumption for SVRT, the patch-space approximation, and the O3 removal behavior. None of these is verified on out-of-distribution data beyond the near-chance OOD results.

assumptions (4)
  • domain assumption Rule-compliant panels have high probability under the learned distribution, and rule-violating panels have low probability.
    Underlies Propositions 3.3-3.5 and the whole substitution of panel correctness with likelihood; the model is trained only to maximize likelihood of RAVEN/PGM panels, so it is not guaranteed to track correctness on unseen tasks.
  • domain assumption SVRT problem panels I_l and I_r are drawn uniformly from dataset D, giving p(I_l)=p(I_r)=1/|D|.
    Used in the proof of Proposition 3.5, Eq. 3, to cancel the panel priors; no justification is given that the panel distribution is uniform.
  • domain assumption Predictability in discrete patch space p(Z_t|Z_C) approximates image-level predictability p(I_t|I_C).
    Section 3.2.1 reduces image generation to patch generation; VQ-VAE reconstruction loss limits how faithfully the codebook preserves rule-relevant details.
  • ad hoc to paper In O3 problems, removing the odd image leaves a rule-compliant panel whose probability is large, while the full panel's probability is finite.
    Proposition 3.4 writes p(I_¬i) = p(I)/p(I_i|I_¬i) and treats the ratio as proportional to 1/p(I_i|I_¬i), which assumes p(I) is not negligible.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Task-Specific Reasoning: A Unified Conditional Generative Framework for Abstract Visual Reasoning." pith.science (2026). https://pith.science/paper/IGLXYZKD

@misc{pith2026250711761,
  author       = {Pith},
  title        = {Pith review of: Beyond Task-Specific Reasoning: A Unified Conditional Generative Framework for Abstract Visual Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IGLXYZKD}},
  note         = {Machine review of arXiv:2507.11761}
}
read the original abstract

Abstract visual reasoning (AVR) enables humans to quickly discover and generalize abstract rules to new scenarios. Designing intelligent systems with human-like AVR abilities has been a long-standing topic in the artificial intelligence community. Deep AVR solvers have recently achieved remarkable success in various AVR tasks. However, they usually use task-specific designs or parameters in different tasks. In such a paradigm, solving new tasks often means retraining the model, and sometimes retuning the model architectures, which increases the cost of solving AVR problems. In contrast to task-specific approaches, this paper proposes a novel Unified Conditional Generative Solver (UCGS), aiming to address multiple AVR tasks in a unified framework. First, we prove that some well-known AVR tasks can be reformulated as the problem of estimating the predictability of target images in problem panels. Then, we illustrate that, under the proposed framework, training one conditional generative model can solve various AVR tasks. The experiments show that with a single round of multi-task training, UCGS demonstrates abstract reasoning ability across various AVR tasks. Especially, UCGS exhibits the ability of zero-shot reasoning, enabling it to perform abstract reasoning on problems from unseen AVR tasks in the testing phase.

Figures

Figures reproduced from arXiv: 2507.11761 by the authors.

Figure 1
Figure 1. An illustration of abstract visual reasoning (AVR) tasks. This paper involves four AVR tasks, despite their differences in the predefined visual concepts (e.g., object shape and color) and abstract rules (e.g., progressive change and logical rule), all assess the ability to infer abstract rules from visual stimuli. (a) Raven’s Progressive Matrices (RPMs) are visual puzzles where participants are given a problem pane… view at source ↗
Figure 2
Figure 2. An Overview of UCGS-T. The image encoder extracts high-level features from each context panel image, which are mapped into discrete context patch representations via vector quantization. The patch encoder captures image-level visual concepts from context patches that encode local information. The panel encoder integrates visual concepts of the context images, understanding abstract rules on the panel, and predicting… view at source ↗
Figure 3
Figure 3. Comparison of generation results on RAVEN (left) and PGM (right). In the visualization result of each model, the bottom-right position is the prediction and the remaining context images are ground truths. The figure illustrates qualitative differences between models, with errors and artifacts appearing in some results predicted by baselines. Note that there is noise in the problem panel, therefore the generated resu… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Construction of the ID tasks. O3-ID. As shown in Figure 4a, we replace the bottom-right image of each 3 × 3 image panel with a randomly selected distractor from the candidate panel to construct odd-one-out tests from RAVEN. This process transforms the bottom-right imag…
Figure 5
Figure 5. Figure 5: Task-specific prompts of GPT-4o. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: Comparison of generation results on RAVEN. [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: Comparison of generation results on PGM. [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 43 canonical work pages

  1. [1]

    Measuring abstract reasoning in neural networks

    Barrett, D., Hill, F., Santoro, A., Morcos, A., and Lillicrap, T. Measuring abstract reasoning in neural networks. In International conference on machine learning, pp.\ 511--520. PMLR, 2018

    work page 2018

  2. [2]

    K., Yantis, S., and Geman, D

    Fleuret, F., Li, T., Dubout, C., Wampler, E. K., Yantis, S., and Geman, D. Comparing machines and humans on a visual categorization test. Proceedings of the National Academy of Sciences, 108 0 (43): 0 17621--17625, 2011

    work page 2011

  3. [3]

    Learning to make analogies by contrasting abstract relational structure

    Hill, F., Santoro, A., Barrett, D., Morcos, A., and Lillicrap, T. Learning to make analogies by contrasting abstract relational structure. In International Conference on Learning Representations, 2019

    work page 2019

  4. [4]

    Kim, H., Mnih, A., Schwarz, J., Garnelo, M., Eslami, A., Rosenbaum, D., Vinyals, O., and Teh, Y. W. Attentive neural processes. arXiv preprint arXiv:1901.05761, 2019

    arXiv 1901

  5. [5]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    arXiv 2014

  6. [6]

    and \.Z ychowski, A

    Ma \'n dziuk, J. and \.Z ychowski, A. Deepiq: A human-inspired ai system for solving iq test problems. In 2019 International Joint Conference on Neural Networks (IJCNN), pp.\ 1--8. IEEE, 2019

    work page 2019

  7. [7]

    Learning to reason over visual objects

    Mondal, S. S., Webb, T., and Cohen, J. D. Learning to reason over visual objects. arXiv preprint arXiv:2303.02260, 2023

    work page Pith review arXiv 2023

  8. [8]

    Pytorch: An imperative style, high-performance deep learning library

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019

    2019

Show all 63 references
  1. [9]

    Neural discrete representation learning

    Van Den Oord, A., Vinyals, O., et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017

    2017

  2. [10]

    N., Kaiser, ., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, ., and Polosukhin, I. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  3. [11]

    Raven: A dataset for relational and analogical visual reasoning

    Zhang, C., Gao, F., Jia, B., Zhu, Y., and Zhu, S.-C. Raven: A dataset for relational and analogical visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 5317--5327, 2019

    2019

  4. [12]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  5. [13]

    L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  6. [14]

    Bai, Y., Geng, X., Mangalam, K., Bar, A., Yuille, A., Darrell, T., Malik, J., and Efros, A. A. Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785, 2023

    2023 arXiv

  7. [15]

    Visual prompting via image inpainting

    Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., and Efros, A. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35: 0 25005--25017, 2022

    2022

  8. [16]

    Scale-localized abstract reasoning

    Benny, Y., Pekar, N., and Wolf, L. Scale-localized abstract reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12557--12565, 2021

    2021

  9. [17]

    D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language models are few-shot learners. Advances in neural information processing systems, 33: 0 1877--1901, 2020

    1901

  10. [18]

    Cao, X., Lai, B., Ye, W., Ma, Y., Heintz, J., Chen, J., Cao, J., and Rehg, J. M. What is the visual cognition gap between humans and multimodal llms? arXiv preprint arXiv:2406.10424, 2024

    2024 arXiv

  11. [19]

    Cattell, R. B. Theory of fluid and crystallized intelligence: A critical experiment. Journal of educational psychology, 54 0 (1): 0 1, 1963

    1963

  12. [20]

    Infogan: Interpretable representation learning by information maximizing generative adversarial nets

    Chen, X., Duan, Y., Houthooft, R., Schulman, J., Sutskever, I., and Abbeel, P. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in neural information processing systems, 29, 2016

    2016

  13. [21]

    On the measure of intelligence

    Chollet, F. On the measure of intelligence. arXiv preprint arXiv:1911.01547, 2019

    1911 arXiv

  14. [22]

    C., Griffiths, T

    Dedhia, B., Chang, M., Snell, J. C., Griffiths, T. L., and Jha, N. K. Im-promptu: In-context composition from image prompts. arXiv preprint arXiv:2305.17262, 2023

    2023 arXiv

  15. [23]

    A., and J \"a kel, F

    Depeweg, S., Rothkopf, C. A., and J \"a kel, F. Solving bongard problems with a visual language and pragmatic constraints. Cognitive Science, 48 0 (5): 0 e13432, 2024

    2024

  16. [24]

    and Mordatch, I

    Du, Y. and Mordatch, I. Implicit generation and modeling with energy based models. Advances in Neural Information Processing Systems, 32: 0 3608--3618, 2019

    2019

  17. [25]

    Neural diffusion processes

    Dutordoir, V., Saul, A., Ghahramani, Z., and Simpson, F. Neural diffusion processes. In International Conference on Machine Learning, pp.\ 8990--9012. PMLR, 2023

    2023

  18. [26]

    J., Eslami, S., and Teh, Y

    Garnelo, M., Schwarz, J., Rosenbaum, D., Viola, F., Rezende, D. J., Eslami, S., and Teh, Y. W. Neural processes. arXiv preprint arXiv:1807.01622, 2018

    2018 arXiv

  19. [27]

    Generative adversarial nets

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial nets. Advances in neural information processing systems, 27, 2014

    2014

  20. [28]

    Gray, J. R. and Thompson, P. M. Neurobiology of intelligence: science and ethics. Nature Reviews Neuroscience, 5 0 (6): 0 471--482, 2004

    2004

  21. [29]

    Masked autoencoders are scalable vision learners

    He, K., Chen, X., Xie, S., Li, Y., Doll \'a r, P., and Girshick, R. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 16000--16009, 2022

    2022

  22. [30]

    A neuro-vector-symbolic architecture for solving raven’s progressive matrices

    Hersche, M., Zeqiri, M., Benini, L., Sebastian, A., and Rahimi, A. A neuro-vector-symbolic architecture for solving raven’s progressive matrices. Nature Machine Intelligence, 5 0 (4): 0 363--375, 2023

    2023

  23. [31]

    Learning to make analogies by contrasting abstract relational structure

    Hill, F., Santoro, A., Barrett, D., Morcos, A., and Lillicrap, T. Learning to make analogies by contrasting abstract relational structure. In International Conference on Learning Representations, 2019 a

    2019

  24. [32]

    G., Morcos, A

    Hill, F., Santoro, A., Barrett, D. G., Morcos, A. S., and Lillicrap, T. Learning to make analogies by contrasting abstract relational structure. arXiv preprint arXiv:1902.00120, 2019 b

    1902 arXiv

  25. [33]

    Stratified rule-aware network for abstract visual reasoning

    Hu, S., Ma, Y., Liu, X., Wei, Y., and Bai, S. Stratified rule-aware network for abstract visual reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pp.\ 1567--1574, 2021

    2021

  26. [34]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  27. [35]

    R., Lewandowsky, S., and Griffiths, T

    Little, D. R., Lewandowsky, S., and Griffiths, T. L. A bayesian model of rule induction in raven's progressive matrices. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 34, 2012

    2012

  28. [36]

    Object-centric learning with slot attention

    Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., and Kipf, T. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33: 0 11525--11538, 2020

    2020

  29. [37]

    A structure-mapping model of raven's progressive matrices

    Lovett, A., Forbus, K., and Usher, J. A structure-mapping model of raven's progressive matrices. In Proceedings of the Annual Meeting of the Cognitive Science Society, 2010

    2010

  30. [38]

    and Ma \'n dziuk, J

    Ma ki \'n ski, M. and Ma \'n dziuk, J. Deep learning methods for abstract visual reasoning: A survey on raven's progressive matrices. arXiv preprint arXiv:2201.12382, 2022 a

    2022 arXiv

  31. [39]

    and Ma \'n dziuk, J

    Ma ki \'n ski, M. and Ma \'n dziuk, J. A review of emerging research directions in abstract visual reasoning. arXiv preprint arXiv:2202.10284, 2022 b

    2022 arXiv

  32. [40]

    Intuitive physics

    McCloskey, M. Intuitive physics. Scientific american, 248 0 (4): 0 122--131, 1983

    1983

  33. [41]

    The colossal book of short puzzles and problems, 2007

    McMillen, J. The colossal book of short puzzles and problems, 2007

    2007

  34. [42]

    Conditional generative adversarial nets

    Mirza, M. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014

    2014 arXiv

  35. [43]

    B., Zhu, Y., and Anandkumar, A

    Nie, W., Yu, Z., Mao, L., Patel, A. B., Zhu, Y., and Anandkumar, A. Bongard-logo: A new benchmark for human-level concept learning and reasoning. Advances in Neural Information Processing Systems, 33: 0 16468--16480, 2020

    2020

  36. [44]

    Generating correct answers for progressive matrices intelligence tests

    Pekar, N., Benny, Y., and Wolf, L. Generating correct answers for progressive matrices intelligence tests. arXiv preprint arXiv:2011.00496, 2020

    2011 arXiv

  37. [45]

    Raven, J. C. and Court, J. Raven's progressive matrices. Western Psychological Services Los Angeles, CA, 1938

    1938

  38. [46]

    Generating diverse high-fidelity images with vq-vae-2

    Razavi, A., Van den Oord, A., and Vinyals, O. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019

    2019

  39. [47]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  40. [48]

    Raven's progressive matrices completion with latent gaussian process priors

    Shi, F., Li, B., and Xue, X. Raven's progressive matrices completion with latent gaussian process priors. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pp.\ 9612--9620, 2021

    2021

  41. [49]

    Compositional law parsing with latent random functions

    Shi, F., Li, B., and Xue, X. Compositional law parsing with latent random functions. In The Eleventh International Conference on Learning Representations, 2023

    2023

  42. [50]

    Towards generative abstract reasoning: Completing raven’s progressive matrix via rule abstraction and selection

    Shi, F., Li, B., and Xue, X. Towards generative abstract reasoning: Completing raven’s progressive matrix via rule abstraction and selection. In The Twelfth International Conference on Learning Representations, 2024

    2024

  43. [51]

    Learning structured output representation using deep conditional generative models

    Sohn, K., Lee, H., and Yan, X. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015

    2015

  44. [52]

    Are disentangled representations helpful for abstract visual reasoning? Advances in Neural Information Processing Systems, 32, 2019

    Van Steenkiste, S., Locatello, F., Schmidhuber, J., and Bachem, O. Are disentangled representations helpful for abstract visual reasoning? Advances in Neural Information Processing Systems, 32, 2019

    2019

  45. [53]

    Abstract diagrammatic reasoning with multiplex graph networks

    Wang, D., Jamnik, M., and Lio, P. Abstract diagrammatic reasoning with multiplex graph networks. arXiv preprint arXiv:2006.11197, 2020

    2006 arXiv

  46. [54]

    S., and Cohen, J

    Webb, T., Mondal, S. S., and Cohen, J. D. Systematic visual reasoning through object-centric relational abstraction. Advances in Neural Information Processing Systems, 36, 2024

    2024

  47. [55]

    The scattering compositional learner: Discovering objects, attributes, relationships in analogical reasoning

    Wu, Y., Dong, H., Grosse, R., and Ba, J. The scattering compositional learner: Discovering objects, attributes, relationships in analogical reasoning. arXiv preprint arXiv:2007.04212, 2020

    2007 arXiv

  48. [56]

    Videogpt: Video generation using vq-vae and transformers

    Yan, W., Zhang, Y., Abbeel, P., and Srinivas, A. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021

    2021 arXiv

  49. [57]

    Compositional scene representation learning via reconstruction: A survey

    Yuan, J., Chen, T., Li, B., and Xue, X. Compositional scene representation learning via reconstruction: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (10): 0 11540--11560, 2023

    2023

  50. [58]

    A deeper look at bongard problems

    Yun, X., Bohn, T., and Ling, C. A deeper look at bongard problems. In Advances in Artificial Intelligence: 33rd Canadian Conference on Artificial Intelligence, Canadian AI 2020, Ottawa, ON, Canada, May 13--15, 2020, Proceedings 33, pp.\ 528--539. Springer, 2020

    2020

  51. [59]

    Abstract spatial-temporal reasoning via probabilistic abduction and execution

    Zhang, C., Jia, B., Zhu, S.-C., and Zhu, Y. Abstract spatial-temporal reasoning via probabilistic abduction and execution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 9736--9746, 2021 a

    2021

  52. [60]

    N., Zhu, S.-C., and Zhu, Y

    Zhang, C., Xie, S., Jia, B., Wu, Y. N., Zhu, S.-C., and Zhu, Y. Learning algebraic representation for systematic generalization in abstract reasoning. arXiv preprint arXiv:2111.12990, 2021 b

    2021 arXiv

  53. [61]

    Mmicl: Empowering vision-language model with multi-modal in-context learning

    Zhao, H., Cai, Z., Si, S., Ma, X., An, K., Chen, L., Liu, Z., Wang, S., Han, W., and Chang, B. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023

    2023 arXiv

  54. [62]

    Abstract reasoning with distracting features

    Zheng, K., Zha, Z.-J., and Wei, W. Abstract reasoning with distracting features. Advances in Neural Information Processing Systems, 32, 2019

    2019

  55. [63]

    and Kankanhalli, M

    Zhuo, T. and Kankanhalli, M. Effective abstract reasoning with dual-contrast network. In International Conference on Learning Representations, 2021

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.