Pith. sign in

REVIEW 3 major objections 6 minor 32 references

Causal-PIK: Causality-based Physical Reasoning with a Physics-Informed Kernel

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Causal-PIK claims that rating actions by their predicted physical effects, instead of their geometric distance, lets Bayesian optimization solve physical puzzles in fewer attempts.

desk verdict The physics-informed kernel is not a valid GP kernel, so the BO justification collapses; the empirical heuristic may still be worth a serious look. read the letter →

arxiv 2505.22861 v2 pith:VBQSHZZD submitted 2025-05-28 cs.LG

Carlota Parés-Morlans , Michelle Yi , Claire Chen , Sarah A. Wu , Rika Antonova , Tobias Gerstenberg , Jeannette Bohg This is my paper
classification cs.LG
keywords BayesianoptimizationphysicalreasoningcausalsimilarityGaussianprocessphysics-informedkernelVirtualToolsbenchmarkPHYREdynamicsmodel
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

Single-intervention physical reasoning tasks are hard because an agent cannot plan precisely in advance; it has to act, observe, and adjust. Causal-PIK claims a recipe for doing this efficiently: wrap a Gaussian-process Bayesian optimizer in a kernel that rates how similar two actions are by the physical effects they are predicted to cause, rather than by how close they are in action space. On the Virtual Tools benchmark it reports the highest AUCCESS among compared agents, and on PHYRE it reports AUCCESS on par with the best reduced-action-space method while operating over the full action space. If the recipe holds, sample-efficient search on such tasks would not require retraining an agent on every new puzzle; a dynamics model plus an effect-similarity kernel would guide the search directly.

What carries the argument

The load-bearing object is the Physics-Informed Kernel, the function $\mathrm{sim}_{csl}(a,b)$ defined by the paper's Eq. 6. It uses a learned dynamics model to predict the next $n_{\text{pred}}$ states after an action is executed; extracts, for each dynamic object $O$, the state change $\dot{s}^O$ at the first causal event $t_{\text{event}}$; and combines directional similarity (cosine of the state-change vectors) with magnitude similarity (a reciprocal-distance score) into a per-object score in $[0,1]$. Averaging over all dynamic objects and applying an exponential scaling yields the final kernel. This replaces the usual geometric-distance assumption of the RBF kernel with an effect-space assumption: two actions are similar when they push objects in similar directions and by similar amounts, so a failed attempt updates the model's belief about every untried action predicted to have the same causal effect.

What would settle it

Compute the Gram matrix of $\mathrm{sim}_{csl}$ for the candidate actions of one Virtual Tools attempt and inspect its smallest eigenvalue: if any such matrix has a negative eigenvalue, the function is not positive semi-definite, so the Gaussian-process posterior variance can go negative and the upper-confidence-bound acquisition becomes ill-defined.

Watch

Extended reading notes

Core claim

The central claim is that the objective function of a physical puzzle is modeled well by a Gaussian process whose kernel encodes causal-effect similarity. For two actions a and b, the kernel averages, over all dynamic objects, the product of a cosine similarity and a magnitude similarity between the state changes predicted by a learned dynamics model, then rescales the average by an exponential to sharpen the contrast; the paper asserts that this function qualifies as a valid kernel because it is symmetric and non-negative. Running this kernel inside an upper-confidence-bound acquisition loop, Causal-PIK reports AUCCESS of 65.0 versus 58.0 for SSUP on the twenty Virtual Tools puzzles, and on PHYRE-1B Cross it reports 41.6, comparable to the best reduced-action-space baseline while searching the full action space. The authors also report that per-puzzle difficulty scores correlate strongly with human performance on PHYRE (r = 0.73).

Load-bearing premise

The method assumes that the similarity function used as the Gaussian-process kernel is a legitimate covariance function; the paper verifies only symmetry and non-negativity, not the stronger positive semi-definiteness property that keeps posterior variances and the acquisition function well defined.

Editorial extensions

If this is right

  • On Virtual Tools with at most ten attempts, Causal-PIK reports AUCCESS 65.0, above SSUP's 58.0 and the RBF-kernel ablation's 42.0.
  • On PHYRE-1B Cross with at most one hundred attempts, Causal-PIK reports AUCCESS 41.6, comparable to the best reduced-action-space baseline (42.2) while searching the full action space, and above the best full-action-space baseline (30.24).
  • Replacing the Physics-Informed Kernel with an RBF kernel drops AUCCESS by 23 points on Virtual Tools and by roughly 14 points on PHYRE, indicating that the causal-similarity kernel is the main driver of the reported gain.
  • When the dynamics model is trained on similar puzzles, the L2 prediction error improves from 19.3 to 3.56 but AUCCESS only rises from 41.6 to 45, so the method degrades gracefully as prediction noise grows.
  • Causal-PIK with one hundred attempts scores above the human baselines on both benchmarks, and its per-puzzle AUCCESS correlates with human scores at r = 0.73 on PHYRE.

Reading between the lines

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

  • A consequence the paper leaves implicit is that a cheap repair exists if the kernel fails to be positive semi-definite: apply an RBF or Matérn kernel to the predicted state-change vectors, or add a diagonal jitter, and the rest of the pipeline would still stand.
  • The same effect-space similarity idea should transfer to any Bayesian-optimization domain where a learned forward model can predict the outcome of an action, such as robot pushing, object rearrangement, or material design, since the kernel equations only require the predicted state changes.
  • The paper's near-miss objective (closest distance to the goal at any timestep) suggests a concrete testable extension: humans' subjective judgments of which failed attempts were 'close' should match the ordering of that objective, a prediction the current aggregate correlation does not yet address.
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. The paper proposes Causal-PIK, a Bayesian optimization (BO) method for single-intervention physical reasoning tasks on the Virtual Tools and PHYRE benchmarks. A learned dynamics model predicts short-horizon effects of candidate actions, and a 'Physics-Informed Kernel' compares actions by these predicted causal effects. This kernel is used in a Gaussian process (GP) surrogate, and an Upper Confidence Bound (UCB) acquisition function selects the next action to execute. The authors report AUCCESS gains over SSUP and prior PHYRE methods, compare against human performance including a new PHYRE user study, and provide an RBF-kernel ablation plus a robustness analysis with noisy dynamics predictions.

Significance. The core idea of using learned causal-effect predictions to define an action-similarity kernel for BO is appealing and potentially useful for sample-efficient physical reasoning. The new human PHYRE study, the cross-fold training protocol, the RBF ablation, and the noisy-prediction analysis are concrete contributions. However, the central mathematical claim that the proposed kernel is a valid GP kernel is incorrect, and the reported performance advantages over the strongest baselines are within one standard error. If the kernel is redefined to be positive semi-definite and the evaluation is tightened, the approach could still be a useful contribution; in its current form, the theoretical foundation and the headline empirical claim are not established.

major comments (3)
  1. [Section 3.2, Eqs. (3)-(6)] The validity argument for the Physics-Informed Kernel is incorrect. Symmetry and pointwise non-negativity do not imply positive semi-definiteness, and Eq. (6) is not even pointwise non-negative: writing m = (1/D) * sum_O sim_obj in [0,1], the formula is sim_csl(a,b) = m*exp(m) - 1, which equals -1 when m=0 (e.g., when all rectified cosine similarities are zero). The function is also indefinite. For D=1 and three unit state-change vectors separated by 120 degrees, every off-diagonal entry has sim_cos = max(0, cos(120 deg)) = 0, so off-diagonal sim_csl = -1 while diagonal entries are e - 1, approximately 1.718. The resulting Gram matrix has eigenvalues approximately {2.718, 2.718, -0.282}, including a negative eigenvalue. A GP with an indefinite kernel can yield a non-positive-semidefinite posterior covariance, making the UCB acquisition function in Algorithm 1, line 5, ill-defined. Since the Physics-Informed Kernel is the paper's core contribution, the theoretical basis for Causal-PIK as a BO/GP method is not established; a positive semi-definite kernel (e.g., an RBF over the per-object causal-effect feature vector) would be needed.
  2. [Section 3.3, Eq. (7)] The objective function f(x) contains an unspecified parameter beta. No value, range, or tuning procedure is given in the main text or appendices. Because f(x) is the target of the GP, beta directly affects the posterior and therefore the selected actions and all reported AUCCESS numbers. Without specifying beta, the method is not fully reproducible and the reported results cannot be independently verified.
  3. [Section 5, Tables 1 and 2] The claim of 'significantly outperforming' state-of-the-art models is not supported by the reported statistics. On PHYRE-1B Cross (Table 2), Causal-PIK reaches AUCCESS 41.6 +/- 9.33, which is lower than RPIN (42.2 +/- 7.1) and essentially tied with Ahmed et al. (41.9 +/- 8.8); the 11-point gap over Harter et al. (30.24) is only against a non-state-of-the-art full-action-space baseline. On Virtual Tools (Table 1), the gap over SSUP (65.0 vs. 58.0) is within one standard deviation (25.0 vs. 27.0). The paper needs paired or per-puzzle error bars and significance testing, and the wording should be revised to match the actual evidence.
minor comments (6)
  1. [Section 3.4] The sentence 'In Equation 7, the shortest observed distance dc is normalized by the counterfactual baseline distance dist(s0,sg)' is inaccurate: dist(s0,sg) is the distance between the initial state and the goal, not a counterfactual rollout without an action object, and Eq. (6) contains no counterfactual baseline term. Please correct the description or define what the counterfactual baseline actually is.
  2. [Section 3.2, Eq. (3)] The cosine similarity in Eq. (3) divides by ||sdot_O,a|| * ||sdot_O,b||; for actions that cause no motion, these norms can be zero and the expression is undefined. Please specify how zero state-change vectors are handled.
  3. [Section 3.1 and Section 5] The initial 9 rollouts are excluded from the attempt count ('warm-up samples that do not count towards the total attempt count'). This is consistent with the SSUP baseline, which also uses warm-up samples, but the comparison to humans, who have no warm-up, is misleading. The AUCCESS values labeled '@10+' also exclude these 9 evaluations, so they do not correspond to a literal maximum of 10 environment calls. Please report the total number of environment evaluations including warm-up alongside the attempt count.
  4. [Section 3.2, Eq. (6)] Equation (6) should be parenthesized as m*exp(m) - 1, where m = (1/D) * sum_O sim_obj(O,a,b). As typeset, it is easy to misread the formula as applying the exponential separately to each term or as multiplying before averaging, which affects interpretation of the kernel values.
  5. [Section 5, PHYRE paragraph] The sentence 'As shown in Table 2, Causal-PIK achieves an AUCCESS rate over 10 points higher than the best-performing baseline' conflicts with Table 2, where the reduced-action-space baselines RPIN and Ahmed et al. reach 42.2 and 41.9, respectively. Please specify that the comparison is only to full-action-space baselines or revise the claim.
  6. [Algorithm 1] If any of the n_initial warm-up samples happens to solve the task, the behavior of the while-success == False loop and the handling of successful initial points are unspecified. Please clarify whether success can be detected during initialization.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the core empirical claims are measured against external benchmarks and the kernel's PSD flaw is a soundness issue, not a circular reduction.

full rationale

The derivation chain is not circular. The dynamics model is trained on folds or variations disjoint from the test puzzles (Appendix A), and the Physics-Informed Kernel is defined from predicted state changes (Eqs. 2-6) rather than fitted to the observed rewards. AUCCESS scores in Tables 1-2 are empirical evaluations against external baselines (SSUP, RPIN, DQN, humans), so no fitted parameter is renamed as a prediction and no equation reduces to its own input by construction. The self-citations to Gerstenberg and Tenenbaum (2016) and Antonova et al. (2019) are motivational or related-work only and do not carry the argument. The main mathematical defect is in Section 3.2: the paper asserts that sim_csl is a valid kernel because it is symmetric and non-negative, but non-negativity does not imply positive semi-definiteness, and sim_csl is in fact provably indefinite (e.g., off-diagonal entries of -1 with diagonal entries of e-1 produce a negative eigenvalue). This invalidates the GP/BO posterior justification, but it is a soundness and correctness failure rather than circularity: the empirical results are not forced by the kernel definition. The low score reflects only minor non-load-bearing self-citations, not circular reasoning.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a learned dynamics model and a hand-designed similarity kernel, not on new physical entities. The primary hidden burden is the kernel's positive semi-definiteness, which the paper asserts without proof; secondary is the transfer of the dynamics model from training to test puzzles.

free parameters (6)
  • beta in objective function (Eq. 7) = not reported
    Scaling constant for near-miss distance in the reward function f(x). Chosen by hand; directly shapes the BO objective.
  • UCB exploration coefficient = not reported
    The Upper Confidence Bound acquisition function in Algorithm 1 requires an exploration weight; its value is not stated in the paper.
  • npred = 20
    Number of timesteps predicted by the dynamics model; set to capture one collision. Affects the causal state-change estimates.
  • n_initial = 9
    Number of warm-up actions used to initialize the GP; these attempts are not counted in the reported AUCCESS, consistent with the SSUP baseline.
  • n_candidate = 500
    Number of Sobol-sampled candidate actions evaluated by the acquisition function each iteration.
  • n_best = 5
    Number of top candidate actions whose outcomes are approximated via probabilistic simulation before final selection.
assumptions (3)
  • ad hoc to paper A symmetric, non-negative function is a valid kernel for a Gaussian process
    Used in Section 3.2 to declare sim_csl a valid kernel. This is not sufficient; positive semi-definiteness is required and is not proven.
  • domain assumption The learned dynamics model (RPIN) provides sufficiently accurate state-change predictions for the kernel to reflect true causal similarities
    The kernel in Eq. 6 depends entirely on predicted state changes. The paper's robustness analysis shows performance is resilient to some noise, but the method presupposes a meaningful signal.
  • domain assumption Dynamics models trained on puzzle variations or training folds generalize to unseen test puzzles
    Virtual Tools trains on 10 variations per puzzle; PHYRE trains on the training fold. The method is tested on held-out puzzles that share 'similar underlying physical concepts'.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal-PIK: Causality-based Physical Reasoning with a Physics-Informed Kernel." pith.science (2026). https://pith.science/paper/VBQSHZZD

@misc{pith2026250522861,
  author       = {Pith},
  title        = {Pith review of: Causal-PIK: Causality-based Physical Reasoning with a Physics-Informed Kernel},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VBQSHZZD}},
  note         = {Machine review of arXiv:2505.22861}
}
read the original abstract

Tasks that involve complex interactions between objects with unknown dynamics make planning before execution difficult. These tasks require agents to iteratively improve their actions after actively exploring causes and effects in the environment. For these type of tasks, we propose Causal-PIK, a method that leverages Bayesian optimization to reason about causal interactions via a Physics-Informed Kernel to help guide efficient search for the best next action. Experimental results on Virtual Tools and PHYRE physical reasoning benchmarks show that Causal-PIK outperforms state-of-the-art results, requiring fewer actions to reach the goal. We also compare Causal-PIK to human studies, including results from a new user study we conducted on the PHYRE benchmark. We find that Causal-PIK remains competitive on tasks that are very challenging, even for human problem-solvers.

Figures

Figures reproduced from arXiv: 2505.22861 by the authors.

Figure 1
Figure 1. Example puzzles from the Virtual Tools (Allen et al., 2020) and PHYRE (Bakhtin et al., 2019) benchmarks. For Virtual Tools, the objective is to place one of the blue action objects from the left to have a red ball fall into the green area. For PHYRE, the objective is to place a red ball of variable radius in the environment to have the green and blue objects touch for at least 3 seconds. The runs highlighted in red … view at source ↗
Figure 2
Figure 2. An overview of Causal-PIK. The Causality-Based Action Selection module proposes the best action to execute based on the predictions from the Gaussian Process. These predictions are generated considering the rewards from previous actions and the causal similarities between candidate actions computed by the Physics-Informed Kernel. These similarities are derived from the individual effects of actions, as predicted by … view at source ↗
Figure 3
Figure 3. Illustration of how causal similarity is computed. We obviate the object superscript as there is only one object (red ball) in this example. (a): Cosine similarity (Eq. 3). s˙ 1 has a high cosine similarity with s˙ 2 as they point in the same direction. s˙ 3 points in the opposite direction of s˙ 1 , obtaining a negative similarity value. (b): Magnitude similarity (Eq. 4). s˙ 2 has a magnitude that is 40% smaller th… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of posterior distributions f(x) generated by Causal-PIK and a BO agent with an RBF kernel. Given the same initial observations {x 1 , x 2 }, Causal-PIK produces a more expressive posterior, effectively clustering actions based on their causal effects - the r…
Figure 5
Figure 5. Figure 5: Puzzle variations for one of the original puzzles in the Virtual Tools benchmark. The puzzle variations share a similar underlying physical concept with the original puzzle, but have different dynamics due to varying object sizes and relative positions. 12 [PITH_FULL_…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 28 canonical work pages

  1. [1]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Physical reasoning using dynamics-aware models

    Ahmed, E., Bakhtin, A., van der Maaten, L., and Girdhar, R. Physical reasoning using dynamics-aware models. arXiv preprint arXiv:2102.10336, 2021

    arXiv 2021

  3. [3]

    R., Smith, K

    Allen, K. R., Smith, K. A., and Tenenbaum, J. B. Rapid trial-and-error learning with simulation supports flexible tool use and physical reasoning. Proceedings of the National Academy of Sciences, 117 0 (47): 0 29302--29310, 2020

    work page 2020

  4. [4]

    Antonova, R., Rai, A., and Atkeson, C. G. Deep Kernels for Optimizing Locomotion Controllers . In Conference on Robot Learning (CoRL). PMLR, 2017

    work page 2017

  5. [5]

    Bayesian Optimization in Variational Latent Spaces with Dynamic Compression

    Antonova, R., Rai, A., Li, T., and Kragic, D. Bayesian Optimization in Variational Latent Spaces with Dynamic Compression . In Conference on Robot Learning (CoRL). PMLR, 2019

    work page 2019

  6. [6]

    Phyre: A new benchmark for physical reasoning

    Bakhtin, A., van der Maaten, L., Johnson, J., Gustafson, L., and Girshick, R. Phyre: A new benchmark for physical reasoning. Advances in Neural Information Processing Systems, 32, 2019

    work page 2019

  7. [7]

    W., Hamrick, J

    Battaglia, P. W., Hamrick, J. B., and Tenenbaum, J. B. Simulation as an engine of physical scene understanding. Proceedings of the National Academy of Sciences, 110 0 (45): 0 18327--18332, 2013

    work page 2013

  8. [8]

    Berkenkamp, F., Krause, A., and Schoellig, A. P. Bayesian optimization with safety constraints: safe and automatic parameter tuning in robotics. Machine Learning, 112 0 (10): 0 3713--3747, 2023

    work page 2023

Show all 32 references
  1. [9]

    R., Gerstenberg, T., Tenenbaum, J

    Bramley, N. R., Gerstenberg, T., Tenenbaum, J. B., and Gureckis, T. M. Intuitive experimentation in the physical world. Cognitive Psychology, 105: 0 9--38, 2018

    2018

  2. [10]

    Bayesian Modeling for Optimization and Control in Robotics

    Calandra, R. Bayesian Modeling for Optimization and Control in Robotics . PhD thesis, Darmstadt University of Technology, Germany, 2017

    2017

  3. [11]

    Pip: Physical interaction prediction via mental simulation with span selection

    Duan, J., Yu, S., Poria, S., Wen, B., and Tan, C. Pip: Physical interaction prediction via mental simulation with span selection. In European Conference on Computer Vision, pp.\ 405--421. Springer, 2022

    2022

  4. [12]

    Feng, S., Whitman, E., Xinjilefu, X., and Atkeson, C. G. Optimization-based Full Body Control for the DARPA Robotics Challenge . Journal of Field Robotics, 32 0 (2): 0 293--312, 2015

    2015

  5. [13]

    and Tenenbaum, J

    Gerstenberg, T. and Tenenbaum, J. Understanding" almost": Empirical and computational studies of near misses. In CogSci, 2016

    2016

  6. [14]

    Forward prediction for physical reasoning

    Girdhar, R., Gustafson, L., Adcock, A., and van der Maaten, L. Forward prediction for physical reasoning. arXiv preprint arXiv:2006.10734, 2020

    2006 arXiv

  7. [15]

    Solving physics puzzles by reasoning about paths

    Harter, A., Melnik, A., Kumar, G., Agarwal, D., Garg, A., and Ritter, H. Solving physics puzzles by reasoning about paths. arXiv preprint arXiv:2011.07357, 2020

    2011 arXiv

  8. [16]

    Bayesian optimization meets riemannian manifolds in robot learning

    Jaquier, N., Rozo, L., Calinon, S., and B \"u rger, M. Bayesian optimization meets riemannian manifolds in robot learning. In Conference on Robot Learning, pp.\ 233--246. PMLR, 2020

    2020

  9. [17]

    On the learning mechanisms in physical reasoning

    Li, S., Wu, K., Zhang, C., and Zhu, Y. On the learning mechanisms in physical reasoning. Advances in Neural Information Processing Systems, 35: 0 28252--28265, 2022

    2022

  10. [18]

    I-PHYRE: Interactive Physical Reasoning

    Li, S., Wu, K., Zhang, C., and Zhu, Y. I-PHYRE: Interactive Physical Reasoning . In The Twelfth International Conference on Learning Representations, 2024

    2024

  11. [19]

    P., Krause, A., Schaal, S., and Trimpe, S

    Marco, A., Berkenkamp, F., Hennig, P., Schoellig, A. P., Krause, A., Schaal, S., and Trimpe, S. Virtual vs. real: Trading off simulations and physical experiments in reinforcement learning with Bayesian optimization . In 2017 IEEE International Conference on Robotics and Autom...

    2017

  12. [20]

    I., Garg, A., Ritter, H., et al

    Melnik, A., Schiewer, R., Lange, M., Muresanu, A. I., Garg, A., Ritter, H., et al. Benchmarks for Physical Reasoning AI . Transactions on Machine Learning Research, 2023

    2023

  13. [21]

    K., Romeres, D., van Baar, J., Smith, K

    Ota, K., Jha, D. K., Romeres, D., van Baar, J., Smith, K. A., Semitsu, T., Oiki, T., Sullivan, A., Nikovski, D., and Tenenbaum, J. B. Data-efficient learning for complex and real-time physical problem solving using augmented simulation, 2021

    2021

  14. [22]

    Learning long-term visual dynamics with region proposal interaction networks

    Qi, H., Wang, X., Pathak, D., Ma, Y., and Malik, J. Learning long-term visual dynamics with region proposal interaction networks. In ICLR, 2021

    2021

  15. [23]

    Y., Bernard, M., Lerer, A., Fergus, R., Izard, V., and Dupoux, E

    Riochet, R., Castro, M. Y., Bernard, M., Lerer, A., Fergus, R., Izard, V., and Dupoux, E. Intphys 2019: A benchmark for visual intuitive physics understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44 0 (9): 0 5016--5025, 2021

    2019

  16. [24]

    P., and De Freitas, N

    Shahriari, B., Swersky, K., Wang, Z., Adams, R. P., and De Freitas, N. Taking the human out of the loop: A review of bayesian optimization. Proceedings of the IEEE, 104 0 (1): 0 148--175, 2015

    2015

  17. [25]

    Planck.js, 2017

    Shakiba, A. Planck.js, 2017. URL https://piqnt.com/planck.js/docs/

    2017

  18. [26]

    Different physical intuitions exist between tasks, not domains

    Smith, K., Battaglia, P., and Vul, E. Different physical intuitions exist between tasks, not domains. Computational Brain & Behavior , 1, 07 2018. doi:10.1007/s42113-018-0007-3

    2018 doi

  19. [27]

    Modeling expectation violation in intuitive physics with coarse probabilistic object representations

    Smith, K., Mei, L., Yao, S., Wu, J., Spelke, E., Tenenbaum, J., and Ullman, T. Modeling expectation violation in intuitive physics with coarse probabilistic object representations. Advances in neural information processing systems, 32, 2019

    2019

  20. [28]

    D., Stuhlm \"u ller, A., Goodman, N

    Ullman, T. D., Stuhlm \"u ller, A., Goodman, N. D., and Tenenbaum, J. B. Learning physical parameters from dynamic scenes. Cognitive Psychology, 104: 0 57--82, 2018

    2018

  21. [29]

    Recent advances in bayesian optimization

    Wang, X., Jin, Y., Schmitt, S., and Olhofer, M. Recent advances in bayesian optimization. ACM Computing Surveys, 55 0 (13s): 0 1--36, 2023

    2023

  22. [30]

    Learning to see physics via visual de-animation

    Wu, J., Lu, E., Kohli, P., Freeman, B., and Tenenbaum, J. Learning to see physics via visual de-animation. Advances in neural information processing systems, 30, 2017

    2017

  23. [31]

    Interpretable intuitive physics model

    Ye, T., Wang, X., Davidson, J., and Gupta, A. Interpretable intuitive physics model. In Proceedings of the European Conference on Computer Vision (ECCV), pp.\ 87--102, 2018

    2018

  24. [32]

    A., Tenenbaum, J

    Zhou, L., Smith, K. A., Tenenbaum, J. B., and Gerstenberg, T. Mental jenga: A counterfactual simulation model of causal judgments about physical support. Journal of Experimental Psychology: General, 152 0 (8): 0 2237--2269, 2023

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.