{"id":"89416391-5a03-4a9f-9501-d5cec164771d","arxiv_id":"2512.13517","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"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.","lead":"The authors built a three-part deep learning model that mentally rotates 3D shapes, guided by VR experiments where people compared shapes and sometimes rotated them with a joystick. It reproduces human accuracy and the typical number of rotation actions, but explains reaction times only partly.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"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.","rationale":"The reader's weakest assumption and my own converge: the Quadrant Hypothesis is the load-bearing element. I considered other flaws — the abstract's 'captures response times' is not backed by an actual RT model (the paper itself says reaction times are only 'partly explained' and cannot explain the 120° vs 180° difference), and the IQR-based removal of four participants weakens the human dataset. These are real but either acknowledged by the authors or secondary. The more fundamental issue is that the human-action match is partly manufactured by the choice of Module III's output space. A model that can only rotate in 90° steps and decides to stop based on a 90° symbolic partition will trivially produce 'about one action' for non-zero disparities; Fig. 10 is therefore not a free behavioral prediction. The no-action ablation makes the point starkly: an invariant classifier solves the task at 97.03% without rotating at all, so the model's actions are not necessary for task performance, and the action-count match is a property of the hand-defined control policy rather than of the learned representations alone. This does not invalidate the paper: the model is a credible feasibility proof that equivariant plus neuro-symbolic components can solve held-out mental rotation, the code is available, and the ablations demonstrate that each component contributes. But the claim that the model provides empirical support for the Quadrant Hypothesis in humans should be conditional pending a test against a continuous-action alternative. Hence the verdict remains CONDITIONAL, unchanged from the reader.","tokens_in":25744,"tokens_out":7684,"duration_ms":78208,"concrete_test":"Replace Module III's five-class output with a continuous rotation regressor trained on the same symbolic sequences to predict the exact angular disparity (or a noisy estimate with fitted error sigma), while keeping the EqNR+VSM features, the iterative loop, the same train/test object splits, and the same stopping rule (stop when symbolic sequences are in the same quadrant). Compare the resulting action-count histograms and final-disparity distributions to human Fig. 10 and Fig. 3. If the continuous agent matches human data as well as the quadrant agent does, the quadrant-specific action space is not necessary to explain the behavior; if it does not (e.g., because it aligns too precisely or produces different action counts), the quadrant partition is doing real explanatory work.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central behavioral claim rests on the Quadrant Hypothesis (Sec. 3.3), but the model's action-count match is largely manufactured by the design of Module III (Sec. 4.1.3). Module III is trained to output exactly five classes: SAME, MIRROR, one quadrant clockwise, one quadrant counterclockwise, and two quadrants apart. At inference, rotation actions are these quadrant jumps, and the agent stops when the symbolic descriptions fall in the same quadrant. Thus the model's action-count distribution in Fig. 10 — 0 actions at 0°, about 1 action at 60/120/180° — is not an emergent behavioral prediction; it is an immediate consequence of the discrete label space and the stopping rule. The human observation that participants average 1.05 actions and leave final disparities within [-45°, +45°] is consistent with quadrant coding, but the same coarse statistics would be produced by any actor making a single approximate alignment within a 90° tolerance. The paper provides no continuous-action control. The no-action ablation (Sec. 5.1) sharpens the problem: a feed-forward classifier trained directly on symbolic descriptions reaches 97.03% accuracy without taking any actions, so action-taking is not required for task competence. The only reason the full model 'captures' action counts is that the action space was defined by the quadrant partition. This makes the claimed support for the Quadrant Hypothesis a designed consequence rather than an empirical discovery. The model remains a serious constructive demonstration, but the behavioral match cannot be cited as independent evidence for human quadrant-based symbolic encoding.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":26170,"tokens_out":5738,"duration_ms":57396,"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":[{"comment":"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).","section":"Section 5 (\"Linear reaction times...\") and Abstract"},{"comment":"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","section":"Sections 3.3, 4.1.3, and Fig. 10"},{"comment":"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.","section":"Section 5.1 (no-action ablation)"},{"comment":"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.","section":"Section 3 and Table 1/Figs. 2-4"},{"comment":"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.","section":"Section 4.1.3 and App. B"},{"comment":"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°.","section":"Sections 5 and Discussion"}],"minor_comments":[{"comment":"Typo: \"Quadrant-Dependant\" should be \"Quadrant-Dependent.\"","section":"Section 3.3 heading"},{"comment":"The notation for symbolic logits is unclear: s_i = [VSM_dec(ˆy_i;