Pith. sign in

REVIEW 6 major objections 5 minor 9 references

A three-module neural model of mental rotation matches human behavior; its key idea is that people compare objects through four 90-degree symbolic descriptions rather than exact angles.

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 ↗

T0 review

2026-08-03 16:24 UTC pith:W6PKPPM6

load-bearing objection The architecture is novel and honest, but the headline behavioral match is largely designed into the label space, not discovered. the 6 major comments →

arxiv 2512.13517 v2 pith:W6PKPPM6 submitted 2025-12-15 q-bio.NC cs.LG

A Deep Learning Model of Mental Rotation Informed by Interactive VR Experiments

Raymond Khazoum , Daniela Fernandes , Aleksandr Krylov , Qin Li , Stephane Deny This is my paper
classification q-bio.NC cs.LG
keywords mental rotationShepard-Metzler objectsneuro-symbolic representationequivariant neural networksQuadrant Hypothesisvirtual realityspatial reasoningdeep learning
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 paper tries to establish that mental rotation—comparing three-dimensional cube-block objects seen from different angles, as in the classic Shepard-Metzler task—can be explained by a specific three-stage neural architecture: an equivariant encoder that builds a manipulable 3D spatial representation from a single image, a neuro-symbolic transformer that converts that representation into a short symbolic description of the object, and a decision agent that uses pairs of such descriptions to choose coarse rotation steps or a same/mirror answer. The authors ran new VR experiments in which people could rotate one of the two objects, and found that people typically take about one ballistic action and settle for rough alignment within plus or minus 45 degrees rather than precise alignment. Their model reproduces human accuracy, this near-single-action behavior, and the classic linear reaction-time trend from 0 to 120 degrees, on shapes held out from training. If the paper is right, it is the first mechanistic neural account of mental rotation that both solves the task and explains why humans rotate objects in the few discrete jumps they do, and it supports a hybrid view: spatial representations are manipulated, but the decisions that guide the manipulation run on symbolic descriptions.

Core claim

The paper's central claim: mental rotation is a loop over three learned components — an equivariant encoder that infers a rotatable 3D spatial latent from an image, a neuro-symbolic transformer that maps the latent to a nine-letter object description, and an MLP agent that declares same/mirror or prescribes a 90°/180° rotation. Rotations act on the spatial latent until a decision. On held-out shapes the model reaches 96% accuracy and takes about as many actions as humans. Ablations confirm each component. The paper reads humans' parsimonious actions as evidence for its Quadrant Hypothesis: objects get four 90°-wide symbolic descriptions, and decisions follow quadrant membership, not exact an

What carries the argument

The key object is the Quadrant-Dependent Symbolic Representation Hypothesis: the 360° circle splits into four 90° quadrants, and each object is described by one of four sequences — nine letters over {U,D,B,F,L,R} listing cube-transition directions. The description is invariant within a quadrant and changes only at 45° boundaries, permuting under a cyclic group. This partition makes Module III's five output classes (same, mirror, one quadrant either way, two quadrants apart) exhaustive, lets a single 90° or 180° rotation land most pairs in the same quadrant, and yields the coarse ±45° alignment humans show. Rotations act on the equivariant spatial latent; decisions are symbolic.

Load-bearing premise

The load-bearing premise is the Quadrant Hypothesis — that people encode each object as one of four 90-degree-wide symbolic descriptions and decide from quadrant membership rather than exact angle; if human mental rotation instead runs on continuous trajectories or another abstraction, the model's match to human action counts is a designed consequence of its five-class discretization, not an independent discovery.

What would settle it

Measure the chronometric curve and action counts at 15-degree angular steps from 0° to 180°: the Quadrant Hypothesis predicts flat within-quadrant reaction times with abrupt jumps at the 45° and 135° boundaries, while a continuous-rotation account predicts a smooth ramp. In the Action setup, additionally ask participants after their final rotation whether the two objects are exactly aligned — the hypothesis predicts they cannot perceive residual misalignment inside ±45°, so an ability to detect and correct small internal angles would falsify it.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Feed-forward Siamese networks trained end-to-end on the same match/mismatch task fall to chance on held-out shapes rotated in depth, whereas the proposed loop generalizes; the paper's ablations conclude that the equivariant spatial latent and the symbolic readout are each necessary for depth-rotation generalization.
  • A symbolic abstraction with four 90-degree classes reproduces the paper's decisive behavioral measurement: people take about one rotation action per trial and stop within ±45° of alignment, and the classic linear reaction-time trend up to 120° follows from the probability that two views land in different quadrants.
  • Because rotation actions are applied to a spatial representation while decisions are made from symbolic descriptions, the model can accommodate both continuous and discrete accounts of mental rotation, and it unifies manual and mental rotation within one decision loop.
  • The paper does not explain the human reaction-time difference between 120° and 180°; it attributes that residual to a hypothesized neural rotation mechanism whose duration grows with angle, implemented via gated recurrent circuits, and notes that the model's action counts are less variable than humans'.

Where Pith is reading between the lines

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

  • Editorial inference: the Quadrant Hypothesis predicts a stepped chronometric curve — reaction times and action counts flat within each quadrant and jumping at the 45°/135° boundaries — which would distinguish it from a smooth continuous-rotation ramp; re-plotting existing reaction-time datasets at fine angular steps would test this, and the paper does not run that test.
  • Editorial inference: the hypothesis has a direct behavioural signature — after their final rotation, participants should be unable to detect residual misalignment inside ±45° and should regain that ability sharply at quadrant boundaries; a VR alignment-judgment probe would settle whether the abstraction is genuinely what humans use.
  • Editorial inference: the same three-module loop could transfer to other spatial-reasoning tasks that admit quadrant-like equivalence classes (part assembly, comparing articulated or folded structures), and the explicit symbolic stage would make the model's rotation choices human-readable; the paper leaves such applications open.
  • Editorial inference: the action-count match is partly built in — the decision module is trained to output exactly the four quadrant-relation classes, so the parsimony result demonstrates that the quadrant discretization is sufficient to reproduce human behavior, not that humans provably use it.

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

6 major / 5 minor

Summary. The paper proposes a three-component neural architecture for mental rotation: (1) an equivariant neural renderer (EqNR) that builds a 3D spatial latent representation from 2D views; (2) a Vision Symbolic Model (VSM) that maps that latent into a sequential symbolic description; and (3) an MLP decision agent that compares symbolic descriptions to either make a same/mirror judgment or choose a quadrant rotation action. The model is trained on Shepard-Metzler shapes and tested on held-out shapes, where it reaches 96.13% accuracy. The authors complement the modeling with a new VR experiment (N=19, 15 analyzed) contrasting a No-Action condition with an Action condition in which participants can rotate one object manually. From the VR data they formulate the "Quadrant-Dependent Symbolic Representation Hypothesis" — that humans encode objects by 90-degree quadrant membership and use coarse quadrant-based rotation actions. The paper claims to capture human performance, response times, and action behavior, and to demonstrate the necessity of each component through ablations. The work is presented as the first mechanistic neural model of human mental rotation.

Significance. If the central claims held, this would be a valuable contribution: it would be one of the first deep-learning models to solve Shepard-Metzler mental rotation on held-out shapes while also making contact with human behavioral data, and it would concretely propose a symbolic-spatial hybrid mechanism for mental rotation. The authors are commendably transparent about important limitations (Section 5: the model cannot explain 120° vs 180° reaction-time differences; the absence of single-trial consistency in human data; the speculative nature of the neural implementation). The release of code and the planned release of the VR dataset are also strengths. However, as discussed below, the behavioral and cognitive-interpretation claims are currently overstated because the action-count behavior is largely an immediate consequence of the chosen discrete label space, and the reaction-time claim is not backed by a fitted RT model.

major comments (6)
  1. [Section 5 ("Linear reaction times...") and Abstract] The abstract states the model "captures well ... response times," but the only RT link in Section 5 is: "If we attribute a fixed duration to each action, we predict an approximately linear reaction time as a function of angle from 0° to 120°." No duration is estimated from human data, no quantitative comparison is made between predicted and observed RTs, and the model is explicitly unable to explain the 120° vs 180° difference in the No-Action setup. RT variability at fixed disparity (Fig. 2, App. Fig. 15) is also not modeled. The RT claim should be either removed from the abstract or supported by an actual RT model (e.g., fitting per-action durations plus baseline times, predicting human RT distributions, and reporting goodness of fit).
  2. [Sections 3.3, 4.1.3, and Fig. 10] The action-count match is a designed consequence of the model architecture, not an emergent behavioral prediction. Module III (Section 4.1.3) is trained on exactly the five classes: SAME, MIRROR, 1QDRT CW, 1QDRT CCW, and 2QDRT. At inference, rotation actions are exactly these quadrant jumps and the trial stops when descriptions fall in the same quadrant. Therefore, for a well-trained classifier, the model must take 0 actions at 0° and ≈1 action at all other tested disparities whenever the initial input is classified correctly. The human observation of parsimonious actions and final disparities within [-45°,+45°] was itself used to define the Quadrant Hypothesis (Section 3.3), so using the model's action counts as evidence for that hypothesis is circular. A continuous-action control, an alternative partition (e.g., 3 or 6 bins), or a model that can produce variable action counts is needed
  3. [Section 5.1 (no-action ablation)] The ablation "We tested whether actions were needed at all" shows that a feed-forward MLP trained directly on same/mirror from the same symbolic descriptions achieves 97.03% accuracy, outperforming the full agentic model (96.13%). The authors acknowledge this and argue that actions are needed to explain human behavior. However, given Major Comment 2, the only behavioral evidence for the action loop is the engineered quadrant-class action distribution. Thus the paper does not currently demonstrate that the recurrent action-taking component is necessary for the claimed cognitive explanation; it demonstrates only that such a loop can produce quadrant-jump statistics if the label space is designed that way. The central claim of "necessity of each component" in the abstract is therefore not established for Module III/action loop.
  4. [Section 3 and Table 1/Figs. 2-4] The human VR experiment is the empirical foundation for the Quadrant Hypothesis and for the behavioral validation, but the analysis is purely descriptive. With 15 analyzed participants and some action-count bins containing 1–5 trials (Fig. 4), the claimed trends (parsimony of actions, final disparities within ±45°, absence of RT-action correlation in App. Fig. 16) need inferential support: mixed-effects models, confidence intervals, or at least per-subject summaries. The removal of 4 participants based on their 0°-condition RT is also a potential selection bias that should be discussed or analyzed for robustness.
  5. [Section 4.1.3 and App. B] The mathematical formalism in App. B asserts that each shape yields four unique symbolic descriptions under the C4 action, but this need not hold for objects with rotational symmetries. If a Shepard-Metzler object has a 180° symmetry, e.g., the symbolic description may be invariant under a 2-quadrant rotation, making the "same quadrant" decision ambiguous and the 5-class scheme ill-posed for some inputs. The paper should state what exactly is assumed about the shapes and whether any of the 288 shapes used violate the uniqueness condition, and how Module III handles such cases.
  6. [Sections 5 and Discussion] The authors are appropriately cautious about the absence of single-trial consistency (App. Figs. 17-18) and the unexplained 120° vs 180° RT difference. These limitations should be prominently stated in the abstract and Introduction as well. The current abstract's phrase "captures well ... response times and behavior" is too strong given these acknowledged failures; a more precise summary would say the model captures average action counts at coarse disparity levels and qualitatively reproduces the average RT slope from 0° to 120°.
minor comments (5)
  1. [Section 3.3 heading] Typo: "Quadrant-Dependant" should be "Quadrant-Dependent."
  2. [Section 4.1.3, Eq. (1)-(2)] The notation for symbolic logits is unclear: s_i = [VSM_dec(ˆy_i;<t, m_i)] mixes hidden states and logits. Also, x_dec;t and x_dec[1:t] are used inconsistently. Please define the exact input tensor shapes and whether the MLP receives per-token logits, hidden states, or a concatenation of both.
  3. [Fig. 10 and Fig. 4] Figures 4 and 10 are nearly identical; consider merging them into a single two-row figure or labeling the human/model panels more clearly. The "Error Rate" under each panel should be defined (error rate on all trials of that disparity, or only match trials?).
  4. [Appendix B] The quadrant intervals Q3 and Q4 are expressed as (3π/4,5π/4] and (5π/4,7π/4], which exceed the principal range [−π,π]; state explicitly that these are taken modulo 2π.
  5. [Section 3.1 / App. A.2] The statement "subjects typically perform a few ballistic actions" is followed by data showing a mean of 1.05 actions; this wording is misleading. Suggest "typically one action."

Circularity Check

2 steps flagged

Action-count match is baked into Module III's quadrant output classes; the model's parsimonious behavior is a designed consequence, not independent evidence for the Quadrant Hypothesis.

specific steps
  1. fitted input called prediction [Section 3.3 (Quadrant Hypothesis) and Section 4.1.3 (Module III); validated in Section 5 / Fig. 10]
    "Actions are consequently limited to rotating the shape by (i) one quadrant clockwise, (ii) one quadrant counterclockwise, or (iii) two quadrants, greatly limiting the typical number of actions needed to solve the task. ... the predicted quadrant disposition is used as a rotation action to be applied on the spatial representation: i) one quadrant clockwise→rotation90◦, ii) one quadrant counterclockwise→rotation−90◦, two quadrants apart→rotation180◦."

    The human observations motivating the Quadrant Hypothesis are that subjects take few actions and end within ±45° (Sec. 3.2). Module III is then trained on exactly the three quadrant-difference classes, and inference maps those classes to ±90°/180° rotations, stopping when the symbolic descriptions fall in the same quadrant. An accurate Module III therefore must produce ~0 actions at 0° and ~1 action at 60/120/180° before terminating; the Fig. 10 action-count match and the 'approximately linear reaction time' in Sec. 5 are entailed by the class/action design, not emergent. The no-action ablation (97.03% accuracy without actions) confirms the matched action count is not needed for task competence, so the 'consistency supports the plausibility of the modeled mechanism' claim rests on a quanti

  2. self definitional [Section 3.2 -> Section 3.3 (Quadrant Hypothesis)]
    "After the last rotation action, the final position of the rotated object is typically found to be distributed within the range [-45°, +45°] of the target object (Fig. 3). ... These observations have led us to formulate a hypothesis on the symbolic nature of the representations ... By dividing the 360◦ rotation space into four equal quadrants, each object can be assigned to one, allowing decisions on rotation actions and similarity judgment to be based on quadrant membership rather than exact angular position."

    The ±45° human final-disparity tolerance is directly re-encoded as the width of the model's four 90° quadrants, and the model's action choices are then defined as quadrant jumps. Consequently, the Quadrant Hypothesis is not independently tested by the model's near-one-action behavior: the model instantiates the human tolerance as its own action labels, and the later match (Fig. 10) is the same observation re-appearing through a different notation. No continuous-action control is provided to show that the quadrant partition is forced by the data rather than assumed.

full rationale

Most of the pipeline is self-contained: Modules I-III are trained on objective image/symbol datasets and evaluated on held-out Shepard-Metzler shapes, and the Siamese/EqNR/VSM ablations are genuine empirical results. The circularity is confined to the behavioral validation of the Quadrant Hypothesis and the action-count 'capture'. Section 3.2 observes that humans make few actions and end within ±45°; Section 3.3 turns that tolerance into a four-quadrant symbolic code; Module II is trained to produce that code and Module III is trained on exactly the quadrant-difference classes, with inference mapping those classes to ±90°/180° rotations and stopping in the same quadrant. An accurate classifier on these labels must produce the reported ~0/1-action pattern, so the claimed match with human action counts is a designed consequence. The paper's assertion that 'this consistency supports the plausibility of the modeled mechanism' is therefore circular for this specific behavioral claim. This does not invalidate the held-out accuracy results or the component ablations, which have independent content, so partial circularity (6) is appropriate.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central behavior claim rests on a representational hypothesis inferred from the same experiments it is validated against. The quadrant partition is a free parameter fitted to the observed [-45,45] degree final-disparity band; the model's action counts largely follow from this partition. Module I's equivariance and the sufficiency of symbolic descriptions are borrowed/assumed rather than proven. Task accuracy on held-out shapes provides independent grounding, but the behavior match is partly by construction.

free parameters (4)
  • Quadrant boundaries for symbolic representation = Quadrants ]-pi/4, pi/4] etc., i.e. 90-degree bins
    Chosen so that human final post-action disparities lie mostly within [-45,45] degrees; this partition defines the action classes of Module III, so action-count predictions are partly a consequence of this choice.
  • Action-count threshold in human VR data = 150 degrees/s angular-velocity crossings
    Threshold for defining joystick 'actions' is arbitrary and changes the human action counts that the model is compared against.
  • Per-action duration for RT prediction = not specified
    The claimed approximate linear RT trend requires assigning a fixed duration to each model action, but no value is fit, tested, or reported.
  • Maximum six actions before model trial is failed = 6
    Arbitrary stopping criterion for model trajectories; affects the reported accuracy definition.
axioms (4)
  • domain assumption A latent representation trained with Dupont et al.'s equivariant neural rendering objective is truly SO(3)-equivariant for the custom Shepard-Metzler dataset and supports arbitrary rotations at inference.
    Invoked in Section 4.1.1; no equivariance error or verification on held-out shapes is reported.
  • ad hoc to paper Quadrant-Dependent Symbolic Representation Hypothesis: human mental rotation can be modeled by four quadrant-specific symbolic descriptions, invariant within each 90-degree quadrant and equivariant under C4.
    Introduced in Section 3.3 after observing human final-disparity distribution; used as training target for VSM and as action classes for Module III.
  • domain assumption The 9-transition symbolic sequence (from nearest cube) is a complete descriptor distinguishing same vs mirrored Shepard-Metzler shapes across all 288 shapes and 6 shape-views.
    Module III's five-way classification assumes symbolic logits carry enough information; no formal proof is given.
  • domain assumption Manual rotation in VR and mental rotation share core cognitive processes.
    Used in Sections 2 and 3.2 to justify treating Action-setup joystick behavior as a window into mental rotation; supported by prior literature but not independently established by this study.
invented entities (1)
  • Quadrant-specific symbolic representation (Quadrant Hypothesis) no independent evidence
    purpose: Compresses continuous azimuth into four symbolic descriptions to guide rotation decisions and similarity judgments.
    Postulated to explain observed parsimony of human actions and final disparities; no independent experiment outside this dataset validates it. The model's action counts are in part a direct implementation of this hypothesis.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A Deep Learning Model of Mental Rotation Informed by Interactive VR Experiments." pith.science (2026). https://pith.science/paper/W6PKPPM6

@misc{pith2026251213517,
  author       = {Pith},
  title        = {Pith review of: A Deep Learning Model of Mental Rotation Informed by Interactive VR Experiments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W6PKPPM6}},
  note         = {Machine review of arXiv:2512.13517}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Mental rotation -- the ability to compare objects seen from different viewpoints -- is a fundamental example of mental simulation and spatial world modeling in humans. Here we propose a mechanistic model of human mental rotation, leveraging recent advances in deep, equivariant, and neuro-symbolic learning. Our model consists of three stacked components: (1) an equivariant neural encoder, producing 3D spatial representations of objects from images, (2) a neuro-symbolic object encoder, deriving symbolic objects descriptions from these spatial representations, and (3) a neural decision agent, comparing these symbolic descriptions to prescribe rotation simulations in 3D latent space via a recurrent pathway. Our model design is guided by the existing experimental literature on mental rotation, which we complemented with experiments in VR where participants could at times manipulate the objects to compare. Our model captures well the performance, response times and behavior of participants in our and others' experiments, and through ablation studies we demonstrate the necessity of each component. Our work adds to a recent collection of deep neural models of human spatial reasoning, further demonstrating the potency of integrating deep, equivariant, and symbolic representations to model the human mind.

Figures

Figures reproduced from arXiv: 2512.13517 by Aleksandr Krylov, Daniela Fernandes, Qin Li, Raymond Khazoum, Stephane Deny.

Figure 1
Figure 1. Figure 1: Screenshot of the VR app for the Action setup. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Response times as a function of an￾gular disparity, for the Action (blue) and No￾Action (grey) setups. Each point represents one (successful match) trial. Hallmark signatures of mental rotation are present in both the Action and No-Action setup, indicating shared cognitive processes between the two setups. In line with previous reports, subjects achieved a high accuracy on the mental rotation task in both … view at source ↗
Figure 3
Figure 3. Figure 3: Angular disparities be￾tween rotated and target objects following the last action for 180° (top), 120° (middle) and 60° (bot￾tom) conditions. The Action setup provides strong evidence that rotation actions and match/mismatch decisions are guided by sym￾bolic representations of objects. Subjects performed an av￾erage of 1.05 rotation actions in the Action setup before reach￾ing a decision. Even at high angu… view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of number of actions per trial across angular disparities (from left to right: [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Panel (A): Average reaction times in the Action setup, divided into three intervals: plan [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The Quadrant Hypothesis: a shape has 4 possible descriptions depending on the reference frame (or viewpoint). Each de￾scription consists of the sequence of tran￾sition between blocks as observed from the given viewpoint. U: up, D: down, B: back, F: forward, L: left, R: right. Algorithmically, Shepard-Metzler shapes are com￾posed of ten cubes and comprise three elbows. A natural way to describe these shapes… view at source ↗
Figure 7
Figure 7. Figure 7: The encoder of the Equivariant Neural Renderer consists of a cascade of 2D residual [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The Vision Symbolic Model consists of a Vision Transformer as the encoder and an [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The Multi-Layer Perceptron consists of three layers, each followed by a BatchNorm layer [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of number of actions per trial on successful match trials across different [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: GUI screenshot for the No-Action Setup [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Panel (A): Average reaction times in the Action setup, divided into three intervals: plan [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Counting the number of actions in a trial. (Left) Raw angular position of the movable [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Left: Reproduction of the linear reaction times reported by Shepard & Metzler (1971). Right: Distribution of reaction times from the open dataset of Ganis & Kievit (2015). Like in our own experiments, a large variability can be found in reaction times for large angular disparities. At 150°, some responses can be very fast, almost as fast as the fastest responses at 50°, putting in question the constant ro… view at source ↗
Figure 16
Figure 16. Figure 16: Reaction times as a function of number of actions per trial across angular disparities [PITH_FULL_IMAGE:figures/full_fig_p025_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Confusion matrices of number of actions between Cohort I and II for successful match [PITH_FULL_IMAGE:figures/full_fig_p025_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Scatter plot showing reaction time correlation between Cohort I and II for [PITH_FULL_IMAGE:figures/full_fig_p025_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Set of all unique symbolic descriptions we can get from a single Shepard-Metzler object. [PITH_FULL_IMAGE:figures/full_fig_p027_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Distribution of number of actions per trial in the symbolic module ablation where the [PITH_FULL_IMAGE:figures/full_fig_p028_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Siamese networks. (A) With ResNet backbones. (B) With ViT backbones. (C) Results on [PITH_FULL_IMAGE:figures/full_fig_p029_21.png] view at source ↗

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

9 extracted references · 4 linked inside Pith

  1. [1]

    This design enables the formation of two cohorts that are each exposed to the same trials: Cohort I composed of all odd sessions and Cohort II composed of all even sessions

    This meant that the subjects shared a session with the previous subject and the consecutive one. This design enables the formation of two cohorts that are each exposed to the same trials: Cohort I composed of all odd sessions and Cohort II composed of all even sessions. Note that due to an experimental error in No-Action Setup, 6 sub- jects were instead p...

    1998

  2. [384]

    with a residual connection and a layer normalization, and finally a feed-forward network with input dimension 384, hidden dimension 2048, and output dimension 384, using ReLU activations, a residual connection and a layer normalization. The VSM was trained for 250 epochs (early stopping at 201), onW= 8 Nvidia V100 GPUs using DDP with a batch size of 32 pe...

    2048

  3. [1971]

    The performance and reaction time are measured

    but ported in a VR environment. The performance and reaction time are measured. 21 Preprint. The Action SetupThis setup is similar to the no-action setup (Fig. 12), but participants can manipulate the object using the VR thumbstick to rotate it in real-time, integrating real physical control into the mental rotation task. Introducing an ‘action’ component...

    2024

  4. [2006]

    World modeling with probabilistic structure integration.arXiv preprint arXiv:2509.09737,

    Klemen Kotar, Wanhee Lee, Rahul Venkatesh, Honglin Chen, Daniel Bear, Jared Watrous, Simon Kim, Khai Loong Aw, Lilian Naing Chen, Stefan Stojanov, Kevin Feigelis, Imran Thobani, Alex Durango, Khaled Jedoui, Atlas Kazemian, and Dan Yamins. World modeling with probabilistic structure integration.arXiv preprint arXiv:2509.09737,

    Pith/arXiv arXiv

  5. [2010]

    Thousand-brains systems: Sensorimotor intelligence for rapid, robust learning and inference.arXiv preprint arXiv:2507.04494,

    Niels Leadholm, Viviane Clay, Scott Knudstrup, Hojae Lee, and Jeff Hawkins. Thousand-brains systems: Sensorimotor intelligence for rapid, robust learning and inference.arXiv preprint arXiv:2507.04494,

    Pith/arXiv arXiv

  6. [2020]

    Cptr: Full transformer network for image captioning.arXiv preprint arXiv:2101.10804,

    Wei Liu, Sihan Chen, Longteng Guo, Xinxin Zhu, and Jing Liu. Cptr: Full transformer network for image captioning.arXiv preprint arXiv:2101.10804,

    Pith/arXiv arXiv

  7. [2021]

    Large vision models can solve mental rotation problems.arXiv preprint arXiv:2509.15271,

    Sebastian Ray Mason, Anders Gjølbye, Phillip Chavarria Højbjerg, Lenka T ˇetkov´a, and Lars Kai Hansen. Large vision models can solve mental rotation problems.arXiv preprint arXiv:2509.15271,

    arXiv

  8. [2022]

    Mind the gap: Benchmarking spatial reasoning in vision-language models.arXiv preprint arXiv:2503.19707,

    Ilias Stogiannidis, Steven McDonagh, and Sotirios A Tsaftaris. Mind the gap: Benchmarking spatial reasoning in vision-language models.arXiv preprint arXiv:2503.19707,

    Pith/arXiv arXiv

  9. [2024]

    16 Preprint

    6Note that it is debated whether recognition of everyday objects engages the same cognitive processes as mental rotation (Searle & Hamm, 2017). 16 Preprint. Tyler Bonnen, Anthony D Wagner, and Daniel LK Yamins. The medial temporal lobe enables visual perception not possible ‘at a glance’.Cognitive Computational Neuroscience (CCN),

    2017

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.