Thousand-Brains Systems: Sensorimotor Intelligence for Rapid, Robust Learning and Inference

Hojae Lee (1); Jeff Hawkins (1) ((1) Thousand Brains Project); Niels Leadholm (1); Scott Knudstrup (1); Viviane Clay (1)

REVIEW 4 major objections 4 minor 1 cited by

Thousand-Brains Systems: Sensorimotor Intelligence for Rapid, Robust Learning and Inference

T0 review · 4 major / 4 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 Monty, the first thousand-brains sensorimotor system, learns 3D objects from few views and generalizes to novel poses.

desk verdict First real evaluation of Monty: the classification and efficiency claims hold up, but the pose-error metric relies on Monty's own symmetry detector and needs external validation before the headline numbers are taken at face value. read the letter →

arxiv 2507.04494 v1 pith:P7U6AHQN submitted 2025-07-06 cs.AI cs.CVcs.LGcs.RO

Niels Leadholm (1) , Viviane Clay (1) , Scott Knudstrup (1) , Hojae Lee (1) , Jeff Hawkins (1) ((1) Thousand Brains Project) This is my paper

classification cs.AIcs.CVcs.LGcs.RO

keywords thousand-brainssystemssensorimotorlearningreferenceframes3Dobjectrecognitionposeestimationcontinualsymmetrydetectionactiveperception

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

The paper evaluates Monty, the first implementation of a thousand-brains system, on combined 3D object recognition and pose estimation using the YCB household-object dataset. The central claim is that a sensorimotor agent that moves a narrow sensor over an object and binds each local observation into an explicit object-centered reference frame can learn structured, shape-based representations from very few views. Because pose hypotheses are inferred from the first observation rather than sampled from training views, the system generalizes to rotations never seen during learning, while the associative, local nature of its updates makes learning rapid, continual, and computationally cheap. These results argue that embodied, sensorimotor learning with reference frames deserves attention as an alternative path to robust AI.

What carries the argument

The central object is the learning module (LM), modeled on a cortical column: a semi-independent sensorimotor unit that stores an object model $M_m$ as a set of points $\{ (x_i, R_i, n_i) \}$ in an object-centered reference frame, where each point binds a location, a local surface pose, and non-pose features. Inference runs a set of $K$ hypotheses $H_t = \{ (m_k, R_k, x_{k,t}, e_{k,t}) \}$, each combining an object identity, a hypothesized object rotation, a location in that object's frame, and an evidence score; movement between observations is integrated into each location hypothesis by path integration, and evidence rises or falls when the sensed surface pose matches or contradicts the stored point near that location. Rotation hypotheses are initialized directly from the first local observation via $R = R_{\mathrm{obs}} \cdot (R_{\mathrm{stored}})^T$, which is what allows never-seen poses to be recognized without a discrete pose grid. The system also includes a goal-state generator implementing a hypothesis-testing policy that proposes movements to the most discriminating location between the two top hypotheses, and a voting mechanism through which lateral learning modules share evidence transformed by their relative sensor displacement.

What would settle it

Compare Monty's declared sensorimotor-symmetric rotation sets against ground-truth symmetry axes computed from the YCB CAD meshes (or from objects with known symmetries): if declared-symmetric rotations are not true symmetries, the reported rotation errors are understated because Eq. 12 scores the closest declared rotation rather than the true pose.

Watch

Extended reading notes

Core claim

Monty is presented as the first working thousand-brains system, replicating cortical-column-like learning modules, and the paper's discovery is that this architecture delivers robust performance on a standard 3D perception benchmark without large-scale passive training. On the YCB dataset of 77 household objects, Monty reaches 98.6% classification accuracy with median zero-degree rotation error on rotations seen during training, 93.0% accuracy on novel uniform $\mathrm{SO}(3)$ rotations, and 73.1% accuracy when all color is removed while noise and novel rotations are also present. The mechanism is that each learning module maintains an object model $M_m$ as a set of points $\{ (x_i, R_i, n_i) \}$ in a 3D reference frame, each storing a local surface pose and features; at inference, a particle-filter-like set of hypotheses $H_t = \{ (m_k, R_k, x_{k,t}, e_{k,t}) \}$ is updated by path-integrating movement and matching local observations to stored points. This yields natural symmetry detection: the sensorimotor-symmetric set of rotations whose evidence stays within threshold of the most likely hypothesis, which the paper validates by Chamfer distance on its own learned models. The same local associative binding enables about 50% recognition accuracy after a single view per object, continual learning across 77 sequential tasks without catastrophic forgetting, and training that uses several orders of magnitude fewer FLOPs than vision transformer baselines.

Load-bearing premise

The pose-accuracy claims assume that the set of rotations Monty reports as sensorimotor-symmetric really is the object's genuine symmetry; the paper validates this only against its own learned models and visual inspection, not against ground-truth symmetry axes from the object meshes.

Editorial extensions

If this is right

Shape-based generalization is strong enough that Monty keeps 73.1% classification accuracy when HSV color is set to a uniform value, even with feature noise and novel rotations present.
Pose estimation extends to rotations never encountered in training, with 93.0% classification and median 4.5-degree rotation error on novel $\mathrm{SO}(3)$ rotations.
Symmetry detection is intrinsic rather than supervised: the system declares an object sensorimotor-symmetric when the high-evidence rotation set is stable for $\theta_{\mathrm{sym}}$ consecutive steps, with no ground-truth model needed.
Rapid inference scales with modularity: increasing the number of learning modules and sensor patches cuts convergence steps substantially while keeping accuracy roughly constant.
Local Hebbian-like binding within reference frames supports one-shot (about 50% after one view) and continual learning across 77 sequential tasks, and training FLOPs are orders of magnitude below ViT fine-tuning and pretraining.

Reading between the lines

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

If this design holds up, the trick of inferring the object pose from the first local observation rather than sampling it is a general recipe for out-of-distribution generalization that could be ported to other local, high-dimensional modalities such as touch or audio.
The sensorimotor-symmetric set is a candidate mechanism for building invariant-yet-equivariant representations in downstream tasks: symmetric pose modes could be collapsed in lateral communication to avoid redundant relearning, an idea the paper only gestures at with its pencil example.
The comparisons to Vision Transformers are suggestive but restricted to isolated objects and supervised labels; a natural next test is cluttered scenes, occlusions, and unsupervised learning, which would probe whether these advantages survive real-world conditions.
One testable extension is whether the voting protocol generalizes across different sensory modalities (e.g., vision and touch), since votes are agnostic to modality by construction.

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.

Desk Editor's Note

read the letter

Key things to know: this is the first quantitative evaluation of Monty, the Thousand Brains implementation, and the central claims are largely supported. On YCB, Monty gets 98.6% classification on seen rotations, 93% on novel SO(3) rotations, and still 73% when all color is removed. It does this with a few hundred training samples and orders of magnitude fewer FLOPs than the ViT baselines. Those numbers are striking, and the paper is honest about what it does and doesn't show. The code is public, the FLOPs tracking library is public, and the authors explicitly note the supervised setting and that unsupervised learning is future work. That's real evidence, and it should be credited.

The genuinely new stuff: the hypothesis-testing policy, the voting acceleration with multiple learning modules, the symmetry detection, and the direct comparison of continual learning against a fine-tuned ViT. The policy comparisons are not huge gains—96.4% vs 95.6%, median steps 28 vs 30—but the voting scaling result is clean and the symmetry detection is interesting if validated properly.

The soft spots, in order of importance. First, the rotation-error metric has a real circularity problem. Eq. 12 takes the minimum over the SMS set, and the SMS set is defined by Monty's own convergence threshold. The validation in 4.1.4 uses Chamfer distance against Monty's learned models, not ground-truth symmetry axes from YCB CAD meshes. The stress-test note is right that a false-positive symmetry declaration would silently inflate pose accuracy. This doesn't sink the classification claims, but it means the "median 0-degree rotation error" is better read as "error modulo symmetries as judged by Monty." They need to validate against CAD ground truth, or at least report error without the min-over-symmetry. Second, there are no error bars or multiple seeds. Given the stochastic exploration policies and the sensitivity of the symmetry threshold, single runs are not enough to trust the finer distinctions. Third, the ViT comparison is not apples-to-apples: full 224x224 RGB-D images vs narrow-patch sensorimotor exploration, and the pre-trained ViT brings in ImageNet-21k. They do constrain Monty to a fixed viewpoint for the FLOPs comparison, which helps, but the training-FLOPs extrapolation from Kaplan et al. plus a custom Python tracker makes the nine-orders-of-magnitude claim a rough estimate, not a precise measurement. Fourth, learning is supervised throughout; the rapid-learning story is associative binding with ground-truth pose and ID, so it doesn't yet demonstrate autonomous acquisition.

Who this is for: anyone working on embodied AI, biologically-inspired architectures, or alternatives to pure deep learning. It deserves a serious referee. I'd send it to review, but I'd require the symmetry validation against CAD ground truth, a few seeds with error bars, and ideally a dedicated 6D-pose baseline before accepting. If those are addressed, this becomes a solid reference point for sensorimotor learning.

Referee Report

4 major / 4 minor

Summary. The paper presents Monty, an open-source implementation of a thousand-brains system, and evaluates it on 3D object recognition and pose estimation using the YCB dataset. Monty learns object models by binding local RGB-D observations into object-centered reference frames while moving over objects, and performs inference by maintaining hypotheses over object identity, pose, and location. The authors report high classification accuracy (98.6% on seen rotations) and low rotation error under noise, novel rotations, and color removal; they show that the learned representations emphasize shape, that sensorimotor symmetric (SMS) pose sets capture object symmetries, and that model-free and model-based policies plus a voting mechanism reduce the number of steps needed for convergence. They also report rapid few-shot learning, continual learning over 77 sequential tasks, and substantially lower FLOPs than ViT baselines. All code and hyperparameters are available in public repositories.

Significance. If the results hold, this is a significant systems demonstration of an alternative, sensorimotor approach to 3D perception. The paper shows that explicit reference frames and associative binding can yield strong out-of-distribution generalization, few-shot and continual learning, and low computational cost from narrow, local observations. Strengths include the public code repositories, the transparent algorithmic descriptions, and the use of YCB objects as an independent benchmark for classification. The main caveats are the symmetry-based pose metric, which is validated only against Monty's own learned models rather than ground-truth CAD symmetries, and the single-run nature of the headline statistics; both are addressed in the major comments.

major comments (4)

[§3.3.2 and §4.1.4] The rotation-error metric Erot in Eq. (12) takes the minimum geodesic distance over the SMS set Rm defined in Eq. (11), where Rm is the set of hypotheses whose evidence remains within θ_converge of the MLH for θsym consecutive steps. This is a strict min-over-declared-symmetries metric, so any false-positive symmetry in Rm directly lowers the reported pose error. The validation in §4.1.4 checks Chamfer distance between Monty's own learned point-cloud models rotated by the SMS poses, plus visual inspection; it does not compare SMS declarations against ground-truth symmetry axes from the YCB CAD meshes. Because learned models are built from a limited set of views (e.g., 14) and can be incomplete or over-smoothed, rotations that look consistent in the learned cloud need not be symmetries of the physical object. I recommend validating the SMS sets against ground-truth symmetries computed from the YCB CAD models (or an independent oracle), reporting the rate of false-positive symmetry declarations per object, and recomputing Erot using only true symmetries. This is load-bearing for the headline pose-estimation claims in Figures 3D and 4C-D.
[§4.1 and Figures 3-7] All accuracy and rotation-error numbers are reported as single runs with no error bars or confidence intervals. Since inference involves random initializations, stochastic noise, and randomized walking policies, it is not clear whether the 98.6%, 93.0%, 88.1%, and 73.1% classification results and the step-to-convergence numbers are stable across repeats. The comparisons between policies (Figure 5B) and between numbers of LMs (Figure 6C-D) are similarly single-run and lack statistical testing. Please report at least 5-10 independent runs with distributions (or bootstrapped confidence intervals) for the headline numbers, and state whether the reported medians are across episodes or across runs.
[§3.7.1 and §4.3.1 (Figure 7A)] The few-shot learning comparison between Monty and the ViTs is not apples-to-apples in the amount of sensory data consumed. For each training 'view', Monty's distant agent collects a sequence of 64×64 local patches together with movement vectors over many steps, whereas the ViT receives a single 224×224 full-object RGB-D image per view. Thus 'one view per object' means very different numbers of sensor observations and effective input pixels for the two systems. This matters because the rapid-learning claim (e.g., 'approximately 50% accuracy after observing only a single view') is comparative. Please either match the total number of local observations/pixels, add a ViT baseline that sees multiple crops per view, or explicitly report the per-view observation counts for both systems so the reader can assess the comparison.
[§4.3.3, Figure 8A] The computational-efficiency claim is sensitive to how FLOPs are attributed. Monty's learning FLOPs are counted with a custom NumPy-level tracker over its actual operations, while the ViT's training FLOPs are extrapolated from a forward-pass count using the 3× multiplier of Kaplan et al. (2020) and include 90 epochs of ImageNet-21k pretraining. The 3× multiplier is a rough convention and the pretraining cost is an estimate. Please report the pretraining and fine-tuning FLOPs separately, provide a sensitivity analysis around the multiplier (e.g., 2-4×), and state whether any overheads (such as the Habitat simulator) are excluded for both systems. Without this, the orders-of-magnitude comparison is hard to audit.

minor comments (4)

[§4.3.3, text after Figure 8A] The phrase 'approximately 34,0000 × fewer FLOPs' appears to contain a typo; it should read '340,000×' or use scientific notation such as 3.4×10^5.
[Abstract and §3.2] The abstract describes Monty as performing sensorimotor learning, but all training experiments use ground-truth object IDs and poses during learning (Section 3.2). Please state this supervision requirement explicitly in the abstract or in the first results paragraph so readers do not infer that the demonstrated learning is unsupervised.
[Figure 3D caption] The caption says 'Gray bars indicate the median,' but the figure appears to show distributions; please clarify what the box and whisker elements represent and whether the median is over objects or over repeated episodes.
[§3.7.3 and Appendix] The main Monty hyperparameters (ε, θ_converge, θ_update, θsym, and the new-point displacement threshold) are referenced but not listed in the main text or appendix; the full table is only in the repository. Please include a hyperparameter table in the appendix for self-contained reproducibility.

Circularity Check

1 steps flagged · score 3.0 of 10

Rotation-error and symmetry claims are evaluated against Monty's own evidence-defined SMS set and its own learned models, making Erot partly self-referential, while classification, continual-learning, and efficiency results remain externally benchmarked.

self definitional [Section 3.3.2 (Eqs. 11-12) and Section 4.1.4]
"This leads to a natural definition of symmetry as poses that cannot be distinguished by sensorimotor exploration of an object, or what we term sensorimotor symmetric (SMS). When an object is deemed SMS, we measure Monty's rotation error as the difference between the ground truth rotation (B M Rgt) and the minimally distant rotation in Robj. That is: Erot = min_{j∈{1,...,J}} Dgeo(B M Rj, B M Rgt) (12) ... To further quantify this, we use Chamfer distance ... Here we rotate learned models by their predicted rotations, and measure the Chamfer distance between the points in each."

Equation 12 evaluates pose accuracy by taking the minimum over Rm, and Equation 11 defines Rm as the rotations Monty's own evidence scores keep within θ_converge of the most-likely hypothesis for the same object. The paper then validates Rm as true symmetry in Section 4.1.4 by computing Chamfer distance between Monty's own learned point-cloud models rotated by those same Rm rotations. Because Rm is generated by matching observations against those exact learned models (Eqs. 4-6), a rotation in Rm will by construction tend to align the learned model points with themselves. Thus the Chamfer check is not an independent ground-truth test (for example, against YCB CAD symmetry axes).

full rationale

The paper's main recognition claims are externally grounded: classification accuracy (98.6%, 93.0%, 73.1% under color removal) is measured against YCB ground-truth object labels and compared to ViT baselines, and the continual-learning and FLOPs results are benchmarked against standard deep-learning protocols. No load-bearing self-citation chain was found: Clay et al. (2024) is cited for the architecture and voting concept, but the quantitative evaluations in this paper are new and do not reduce to that prior work. The one genuine self-referential loop is in the symmetry/pose-error metric: Erot is defined as the minimum over the system's own SMS set (Eq. 12), and the validation of that set as true symmetry uses Chamfer distance between Monty's own learned models rather than ground-truth CAD symmetry. This makes the pose-error numbers partly a statement about Monty's internal consistency rather than an externally verified pose estimate. It is a secondary metric, however, and does not undermine the central classification, robustness, or learning-efficiency results. Accordingly, a score of 3 reflects one partial, secondary circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The paper is an empirical evaluation rather than a derivation, so the ledger captures hand-chosen thresholds, supervised-label assumptions, and the SMS evaluation construct. The most load-bearing items are the supervised-setup axiom, the SMS-based rotation metric, and the unstated matching thresholds (epsilon, D, theta values). These do not break the recognition claims, but they do limit the strength of the pose and symmetry claims.

free parameters (6)

theta_sym (symmetry detection threshold) = 5 steps
Hand-chosen in Section 3.3.2. Directly controls how many poses join the SMS set and hence affects the rotation-error metric in Eq. 12. The authors note a trivial low-error solution if set too low and fix it at 5, but report no sensitivity analysis.
theta_converge (termination evidence gap) = not reported (fixed across experiments)
Section 3.3.2 defines the threshold for an LM to reach a terminal condition. The value is left to the code repository and affects the accuracy versus steps tradeoff.
theta_update (sparse hypothesis update threshold) = not reported (fixed across experiments)
Section 3.3.2 uses a relative threshold to skip low-evidence hypotheses. It affects computational cost and evidence dynamics, but the exact value is not stated.
epsilon (neighborhood radius in Eq. 4) = not reported
Defines which stored points are considered when updating evidence in inference. This is a central matching parameter, but its value appears only in code.
new-point displacement threshold = not reported
Section 3.2 adds a new model point only if it is sufficiently displaced in Cartesian or feature space. This controls model density and memory, but the threshold values are not given.
ViT comparison hyperparameters = lambda=1.0, LR=5e-4, early stop 25 epochs
Section 3.7.1 tunes the ViT baseline on a validation split. These choices affect the accuracy and FLOPs comparison central to Section 4.3, but are not fitted to support Monty's own claims.

assumptions (4)

domain assumption Ground-truth object identity and pose are available during learning
Section 3.2 states: 'In supervised learning, Monty is provided with ground-truth values of the object ID and orientation.' All experiments use this setup; unsupervised learning is explicitly deferred. The sensorimotor learning claim is therefore about supervised binding, not autonomous discovery.
domain assumption A narrow RGB-D patch encoded as surface-normal and principal-curvature pose plus HSV is sufficient for 3D object recognition and pose
Section 3.2 and 3.6 assume that a 64 by 64 pixel patch, reduced to one CMP point, carries enough geometric signal to identify and localize objects. The paper's robustness results support this within the YCB domain, but it is a strong input-modeling assumption.
ad hoc to paper Monty's SMS set approximates true rotational symmetry of objects
Section 3.3.2 defines sensorimotor symmetry as poses Monty cannot distinguish, and Eq. 12 uses the SMS set to compute rotation error. External validation via Chamfer distance uses Monty's learned models, not ground-truth CAD symmetry axes, so the equivalence is assumed rather than proven.
domain assumption Habitat-simulated rigid objects are a valid testbed for sensorimotor object perception
Section 3.6 evaluates only isolated, rigid YCB objects in simulation, without clutter, occlusion, dynamics, or real sensors. Generalization to physical robot perception is assumed rather than demonstrated.

invented entities (1)

Sensorimotor symmetric (SMS) pose set independent evidence
purpose: Defines the set of rotations Monty treats as indistinguishable and is used as the reference for reporting rotation error and for the symmetry-detection claim.
Introduced in Section 3.3.2 (Eqs. 11-12) and used throughout Section 4.1.4. It has partial independent evidence via Chamfer distances and visual inspection, but no validation against ground-truth CAD symmetry, so it is a plausible but not fully secured construct.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Thousand-Brains Systems: Sensorimotor Intelligence for Rapid, Robust Learning and Inference." pith.science (2026). https://pith.science/paper/P7U6AHQN

@misc{pith2026250704494,
  author       = {Pith},
  title        = {Pith review of: Thousand-Brains Systems: Sensorimotor Intelligence for Rapid, Robust Learning and Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P7U6AHQN}},
  note         = {Machine review of arXiv:2507.04494}
}

read the original abstract

Current AI systems achieve impressive performance on many tasks, yet they lack core attributes of biological intelligence, including rapid, continual learning, representations grounded in sensorimotor interactions, and structured knowledge that enables efficient generalization. Neuroscience theory suggests that mammals evolved flexible intelligence through the replication of a semi-independent, sensorimotor module, a functional unit known as a cortical column. To address the disparity between biological and artificial intelligence, thousand-brains systems were proposed as a means of mirroring the architecture of cortical columns and their interactions. In the current work, we evaluate the unique properties of Monty, the first implementation of a thousand-brains system. We focus on 3D object perception, and in particular, the combined task of object recognition and pose estimation. Utilizing the YCB dataset of household objects, we first assess Monty's use of sensorimotor learning to build structured representations, finding that these enable robust generalization. These representations include an emphasis on classifying objects by their global shape, as well as a natural ability to detect object symmetries. We then explore Monty's use of model-free and model-based policies to enable rapid inference by supporting principled movements. We find that such policies complement Monty's modular architecture, a design that can accommodate communication between modules to further accelerate inference speed via a novel `voting' algorithm. Finally, we examine Monty's use of associative, Hebbian-like binding to enable rapid, continual, and computationally efficient learning, properties that compare favorably to current deep learning architectures. While Monty is still in a nascent stage of development, these findings support thousand-brains systems as a powerful and promising new approach to AI.

Figures

Figures reproduced from arXiv: 2507.04494 by the authors.

**Figure 1.** The Architecture of Monty, a Thousand-Brains System. A) Monty consists of a series of modular components, the learning module, sensor module, and motor system, that interact via a common Cortical Messaging Protocol (CMP). CMP signals consist of a pose and features, and can be used for feed-forward communication (blue), lateral voting for achieving rapid consensus (green), and goal states for guiding the motor system… view at source ↗

**Figure 2.** Learning in Monty. A) Monty performs supervised learning by initializing an internal coordinate space (reference frame) for an object if it has not been previously learned. It then moves over the object, associating sensed observations with locations in this internal reference frame. The movement pattern used during learning consists of either an eye-like agent performing a spiral scan (as shown diagrammatically) or… view at source ↗

**Figure 3.** Robust Sensorimotor Inference. A) We examine Monty’s internal representations during inference of a mug, focusing on the evidence values associated with three out of the 77 known objects. B) Incoming sensory information from local patches arrives in the form of a CMP signal. An LM matches this input against information stored in its reference frames for known objects. Evidence accumulates for different location hypo… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Structured Object Representations and Symmetry. A) We present ten objects to Monty that correspond to human categories of cutlery, boxes, and cups. After Monty explores a given object, we examine the evidence counts for all the hypotheses to measure similarity, where s…

**Figure 5.** Figure 5: Rapid Inference with Model-Free and Model-Based Policies. A) Rather than perform a random walk over the surface of the object, Monty can make use of model-free policies, resulting in more rapid and robust inference. Here, the agent moves along the surface of the object…

**Figure 6.** Figure 6: Rapid Inference with Voting. A) LMs receiving information from different sensory patches (such as the index fingers of a left and right hand) can vote with one another to rapidly reach consensus about what object is being sensed. In this diagram, LM1 receives informati…

**Figure 7.** Figure 7: Few-Shot and Continual Learning. A) We evaluate performance after training on all 77 objects, but where each object is presented from a small number of views. We compare to a fine-tuned ViT (previously trained on 14 million images), and a ViT trained from scratch. The …

**Figure 8.** Figure 8: Computationally Efficient Learning and Inference. A) During learning, Monty only modifies representations in the current reference frame, at the current location where information is sensed. This results in a massive reduction in the Floating Point Operations (FLOPs) …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

A Deep Learning Model of Mental Rotation Informed by Interactive VR Experiments
q-bio.NC 2025-12 conditional novelty 7.0 of 10

A deep neural model with equivariant, symbolic, and agentic components reproduces human accuracy and action counts on Shepard-Metzler mental rotation, but not full reaction-time behavior.

Reference graph

Works this paper leans on

8 extracted references · 1 canonical work pages · cited by 1 Pith paper

[1]

L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., Avila, R., Babuschkin, I., Balaji, S., Balcom, V ., Baltescu, P., Bao, H., Bavarian, M., Belgum, J., Bello, I.,

Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., Avila, R., Babuschkin, I., Balaji, S., Balcom, V ., Baltescu, P., Bao, H., Bavarian, M., Belgum, J., Bello, I., . . . Zoph, B. (2023). GPT-4 technical report. arXiv [cs.CL]. http://arxiv.org/abs/2303.08774 Ba, J. L., Kiros, J. ...

arXiv 2023
[2]

R., Jankowicz, D., & Becker, S

Howell, S. R., Jankowicz, D., & Becker, S. (2005). A model of grounded language acquisition: Sensorimotor features improve lexical and grammatical learning. Journal of Memory and Language, 53(2), 258–276. Hubel, D. H., & Wiesel, T. N. (1974). Uniformity of monkey striate cortex: A parallel relationship between field size, scatter, and magnification factor...

arXiv 2005
[8]

S., van Steenkiste, S., Sajjadi, M

Zimmermann, R. S., van Steenkiste, S., Sajjadi, M. S., Kipf, T., & Greff, K. (2023). Sensitivity of slot-based object- centric models to their number of slots. arXiv preprint arXiv:2305.18890. 30 A PREPRINT - S EPTEMBER 21, 2025 7 Appendix 7.1 Summary of Mathematical Notation Below, we include Tables 1 and 2 to summarize the mathematical notation we use. ...

arXiv 2023
[115]

Bjorck, J., Castañeda, F., Cherniadev, N., Da, X., Ding, R., Fan, L., Fang, Y ., Fox, D., Hu, F., Huang, S., et al. (2025). Gr00t n1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734. Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Groom, L., Hausman, K., Ichter, B., Jakubczak, S., Jones...

arXiv 2025
[183]

https://doi.org/10.1016/j.cell.2020.10.024 Xiang, Y ., Schmidt, T., Narayanan, V ., & Fox, D. (2017). Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199. Yarbus, A. L. (1967). Eye Movements During Perception of Complex Objects. InEye movements and vision. Ye, X. (2023). Calflops: A fl...

arXiv 2017
[202]

M., & Mountcastle, V

Edelman, G. M., & Mountcastle, V . B. (1982).The mindful brain: Cortical organization and the group-selective theory of higher brain function. MIT press. Fairfield, N., & Wettergreen, D. (2008). Active localization on the ocean floor with multibeam sonar.OCEANS 2008, 1–10. Flesch, T., Balaguer, J., Dekker, R., Nili, H., & Summerfield, C. (2018). Comparing...

arXiv 1982
[872]

Higgins, I., Racanière, S., & Rezende, D. (2022). Symmetry-based representations for artificial and biological general intelligence. Frontiers in Computational Neuroscience, 16, 836498. Hinton, G., Krizhevsky, A., Jaitly, N., Tieleman, T., & Tang, Y . (2012). Does the brain do inverse graphics.Brain and Cognitive Sciences Fall Colloquium,

work page 2022
[2025]

McCloskey, M., & Cohen, N. J. (1989). Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem. Psychology of Learning and Motivation - Advances in Research and Theory , 24(100). https : //doi.org/10.1016/S0079-7421(08)60536-8 Mnih, V ., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmil...

arXiv 1989

Pith tools

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

[1] [1]

L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., Avila, R., Babuschkin, I., Balaji, S., Balcom, V ., Baltescu, P., Bao, H., Bavarian, M., Belgum, J., Bello, I.,

Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., Avila, R., Babuschkin, I., Balaji, S., Balcom, V ., Baltescu, P., Bao, H., Bavarian, M., Belgum, J., Bello, I., . . . Zoph, B. (2023). GPT-4 technical report. arXiv [cs.CL]. http://arxiv.org/abs/2303.08774 Ba, J. L., Kiros, J. ...

arXiv 2023

[2] [2]

R., Jankowicz, D., & Becker, S

Howell, S. R., Jankowicz, D., & Becker, S. (2005). A model of grounded language acquisition: Sensorimotor features improve lexical and grammatical learning. Journal of Memory and Language, 53(2), 258–276. Hubel, D. H., & Wiesel, T. N. (1974). Uniformity of monkey striate cortex: A parallel relationship between field size, scatter, and magnification factor...

arXiv 2005

[3] [8]

S., van Steenkiste, S., Sajjadi, M

Zimmermann, R. S., van Steenkiste, S., Sajjadi, M. S., Kipf, T., & Greff, K. (2023). Sensitivity of slot-based object- centric models to their number of slots. arXiv preprint arXiv:2305.18890. 30 A PREPRINT - S EPTEMBER 21, 2025 7 Appendix 7.1 Summary of Mathematical Notation Below, we include Tables 1 and 2 to summarize the mathematical notation we use. ...

arXiv 2023

[4] [115]

Bjorck, J., Castañeda, F., Cherniadev, N., Da, X., Ding, R., Fan, L., Fang, Y ., Fox, D., Hu, F., Huang, S., et al. (2025). Gr00t n1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734. Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Groom, L., Hausman, K., Ichter, B., Jakubczak, S., Jones...

arXiv 2025

[5] [183]

https://doi.org/10.1016/j.cell.2020.10.024 Xiang, Y ., Schmidt, T., Narayanan, V ., & Fox, D. (2017). Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199. Yarbus, A. L. (1967). Eye Movements During Perception of Complex Objects. InEye movements and vision. Ye, X. (2023). Calflops: A fl...

arXiv 2017

[6] [202]

M., & Mountcastle, V

Edelman, G. M., & Mountcastle, V . B. (1982).The mindful brain: Cortical organization and the group-selective theory of higher brain function. MIT press. Fairfield, N., & Wettergreen, D. (2008). Active localization on the ocean floor with multibeam sonar.OCEANS 2008, 1–10. Flesch, T., Balaguer, J., Dekker, R., Nili, H., & Summerfield, C. (2018). Comparing...

arXiv 1982

[7] [872]

Higgins, I., Racanière, S., & Rezende, D. (2022). Symmetry-based representations for artificial and biological general intelligence. Frontiers in Computational Neuroscience, 16, 836498. Hinton, G., Krizhevsky, A., Jaitly, N., Tieleman, T., & Tang, Y . (2012). Does the brain do inverse graphics.Brain and Cognitive Sciences Fall Colloquium,

work page 2022

[8] [2025]

McCloskey, M., & Cohen, N. J. (1989). Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem. Psychology of Learning and Motivation - Advances in Research and Theory , 24(100). https : //doi.org/10.1016/S0079-7421(08)60536-8 Mnih, V ., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmil...

arXiv 1989