{"id":"7104d671-34e3-4c5d-96d6-b4d633da0343","arxiv_id":"2504.17939","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A neural-network model using action-outcome prediction, novelty-driven exploration, and motor noise reproduces limb-preference, control-group, and extinction-burst results from the infant mobile paradigm.","lead":"Researchers built a computer model of a baby learning that kicking one leg makes a mobile dance, using prediction, surprise, and motor noise. The model reproduces key infant behaviors, including favoring the connected limb and reacting when the mobile is unplugged.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-limb activity-interest map pre-specifies credit assignment; limb differentiation and the 'numerous muscles' ablation may reflect this limb-level abstraction rather than the proposed infant mechanisms.","rationale":"The paper is a serious modeling effort with public code, explicit mechanisms, and honest discussion of mismatches with infant data. The reader's CONDITIONAL verdict already accounts for the abstraction concerns. My stress-test identifies a more specific mechanism within that abstraction: the per-limb activity-interest map performs direct credit assignment, so the model does not have to solve the harder problem of discovering which muscle-level actions are causally effective. This weakens the inferential leap from 'necessary in the model' to 'involved in infants,' especially for the muscle-count ablation. However, the core qualitative replications (limb differentiation, contingent vs. non-contingent difference, binary vs. conjugate difference) are still produced by a mechanism combining prediction error, novelty-based exploration, and motor noise; the limb-level map is a representational choice that the paper openly describes. The concern is substantive and should be addressed with the proposed re-run, but it does not overturn the conditional acceptance. The verdict remains CONDITIONAL, and since the reader already assigned CONDITIONAL, I mark the resulting recommendation as UNCHANGED.","tokens_in":24500,"tokens_out":8212,"duration_ms":91389,"concrete_test":"Run the model with the activity-interest map defined over the 600-dimensional muscle-command output (or over a single continuous action vector) instead of the per-limb 4x10 maps, keeping all parameters and losses otherwise identical. If connected-limb differentiation still emerges across 20 runs in both binary and non-binary conditions, the limb-level map is not doing the work; if it fails, the model's core result depends on the per-limb credit assignment. In addition, re-run the muscle-count ablation with a dense random matrix M (e.g., uniform entries, no beta sparsity) or with a fixed per-limb one-hot command set; if 50 commands then suffice for the binary condition, the 'numerous muscles' claim is an artifact of the sparse sampling scheme.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is that the model's exploration module pre-specifies credit assignment at the limb level. Section II-C3 defines an activity-interest map of dimensions (4,10): one 10-bin map per limb. When a prediction error (surprise) occurs, the model sets interest to 1 for the activity range of the limb that was just active; when no surprise occurs, it decrements that limb's range. Thus the surprise signal is attributed directly to the specific limb that caused it, and the network is trained to drive each limb toward its own high-interest range. The hard part of the mobile paradigm—discovering which among many possible movements is causally effective—is therefore not solved at the muscle level; the model is told which limb's activity range to reinforce by the structure of the interest map. The 600 muscle commands enter only as a fixed random linear map M from commands to the four limb scalars; credit for surprise never propagates to individual commands. Consequently, the ablation claiming that 'numerous muscle commands' are essential (Section III-E3, Fig. 10) tests only whether the network can drive the four limb scalars through this particular sparse M, not whether high-dimensional motor control is needed for contingency learning. In the binary condition, 100 commands already suffice, and failure at 50 commands likely reflects the beta-sparse sampling (most weights near zero) leaving some limbs poorly controllable, rather than any biological requirement for ~600 muscles. The model's central limb-specificity result may thus be an artifact of the limb-level interest map rather than evidence that prediction, novelty, and motor noise suffice to discover the contingency from muscle-level actions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a neural-network model of the mobile paradigm, comprising 600 abstract muscle commands mapped onto four limb scalars through a fixed random sparse matrix, a sensory prediction module, an activity-interest map driving exploration, motor noise, and a baseline activity term. The model is compared visually with infant data from a binary-condition study [12] and a conjugate-condition study [11]. The authors report that the model replicates preferential movement of the connected limb, the contingent versus non-contingent difference, the binary versus conjugate difference, and occasional extinction bursts. A series of ablation studies claims that prediction error, the exploration loss, motor noise, and a large number of muscle commands are essential for reproducing infant behavior.","tokens_in":24801,"tokens_out":7071,"duration_ms":72342,"significance":"The paper addresses an important developmental phenomenon with an unusual combination of mechanisms (action-outcome prediction, novelty-based exploration, noise-driven variability) and ships open-source code, individual simulation runs, and comparisons against two recent empirical datasets. If the mechanism claims were supported, the model would be a valuable step toward understanding sensorimotor contingency learning in infancy. However, the model's architecture pre-assigns credit to the just-active limb via the per-limb activity-interest map, and the main parameters are tuned on the same infant data, so the central explanatory claims are not currently supported.","major_comments":[{"comment":"The activity interest map is an array of dimensions (4,10), one row per limb, and whenever a limb action generates surprise the interest for that limb's activity range is set to 1. This pre-specifies credit assignment at the level of individual limbs: the model never has to discover which movement is causally effective, because the surprise signal is attributed directly to the limb that was just active. The limb-specificity result of Fig. 4 therefore largely follows from the map's structure, and the prediction-error ablation in Section III-E1 does not demonstrate that the network learns the contingency.","section":"II-C3 and Fig. 3"},{"comment":"The muscle-command ablation varies the number of output neurons while keeping the fixed random sparse mapping M from commands to four limb scalars. Since surprise credit is assigned only at limb level and never propagates to individual commands, the degradation at 50 commands is expected from the beta(0.01, 0.1) sampling leaving some limbs poorly controllable through this particular random matrix. This ablation cannot support the claim in Section IV-A that 'numerous muscle commands' are essential for infant-like contingency learning; it only shows that four scalar readouts require enough output dimensions for this random mapping.","section":"III-E3 and Fig. 10"},{"comment":"The learning rate was chosen 'after empirical testing' to best match the infant studies, and the baseline activity mean and standard deviation were taken from the same studies [11, 12]. The main results are therefore fits to the benchmark data, not predictions. All model-data comparisons (e.g., Figs. 4, 6, 7) are qualitative visual assessments without quantitative fit measures, effect sizes, or statistical tests. The phrase 'correctly simulates' is not supported by the evidence presented.","section":"II-C1 and II-C4"}],"minor_comments":[{"comment":"Model activity is plotted in arbitrary units while infant activity is plotted in gravitational acceleration units (g); provide a normalization or a defined units mapping to make the visual comparisons interpretable.","section":"Figs. 4, 6, 7"},{"comment":"The text reports 'no evidence' for an extinction burst in the binary condition and 'a clear one' in the non-binary condition, but then states that individual runs do not show clear evidence; clarify the model's actual prediction regarding extinction bursts.","section":"III-C"},{"comment":"The novelty threshold ablation is described only qualitatively; quantify variability (e.g., standard deviation of limb activity) to support the claim that the threshold can be adjusted to match infant variability.","section":"III-E5"},{"comment":"The baseline module's linear increase in the maximum activity and the fatigue-from-stillness assumption are ad hoc; report sensitivity analyses for these choices.","section":"II-C4"},{"comment":"The term 'muscle commands' is used as an abstraction; consider explicitly distinguishing between muscles and output neurons in the ablation discussion to avoid implying that the model simulates anatomical muscles.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper is built around a limb-level curiosity map that hands the model the very contingency it is supposed to learn, and the key parameters are fit to the same data. These are load-bearing problems rather than presentation issues. If the authors wish to resubmit, they would need to (a) remove the limb-specific credit assignment or otherwise demonstrate that the model can discover the connected limb, and (b) provide a quantitative, non-circular evaluation. The current manuscript is not suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a serious modeling paper that reproduces the main mobile-paradigm effects with a mechanism combining predictive surprise, exploration, motor noise, and redundant motor commands. It deserves peer review, but it currently leans on visual curve-matching and on parameters tuned against the very datasets it benchmarks.\n\nThe new thing relative to Butko & Movellan, Zaadnoordijk, and Kelso & Fuchs is the specific combination: a neural network that predicts sensory outcomes, an activity-interest map that drives exploration toward surprising actions, motor noise, and a 600-command overlapping muscle mapping. The model captures limb differentiation, contingent vs. non-contingent differences, the binary vs. conjugate distinction, and occasional extinction bursts. It targets two recent fine-grained infant datasets rather than just classic effects. Code is public, and the paper shows all individual runs plus several ablations.\n\nThe ablations are genuinely informative: removing prediction error, exploration, or motor noise kills limb differentiation, which supports the claim that these components matter in the model. The authors are also candid about mismatches: slower differentiation than infants, lower variability, and no clear extinction burst in the binary condition.\n\nThe main soft spot is parameter circularity. The learning rate was chosen after empirical testing to best match the two infant studies, and the baseline activity statistics came from those same studies. The benchmark plots are therefore not independent predictions. The paper needs either a held-out dataset, a fit-on-one-test-on-the-other procedure, or a systematic sensitivity analysis showing the effects persist across a plausible parameter range. Relatedly, model-data agreement is assessed only visually; no effect sizes, confidence intervals, or statistical tests are reported. That should be fixable with the existing per-run data.\n\nA second, more subtle issue is the limb-level activity-interest map. The surprise signal is assigned to the activity ranges of the limbs that were active, so credit assignment is done at the limb level, not at the muscle-command level. This is a reasonable abstraction, but it partially sidesteps the harder inference problem of discovering which among many muscle commands is causally effective. The \"numerous muscles\" ablation is interesting, but the result depends on the specific random sparse mapping. I would temper the claim that the model demonstrates a biological requirement for roughly 600 muscles; it shows that within this architecture, enough motor redundancy is needed, which is a weaker statement.\n\nBottom line: worth engaging. I'd send it to review, but I'd ask for quantitative fits, an honest treatment of parameter tuning, and more careful wording about what the muscle ablation can and cannot say.","headline":"A solid, honest mobile-paradigm model combining predictive surprise and motor redundancy; the effects reproduce, but parameter circularity and visual-only curve matching keep the claims from being established.","tokens_in":25357,"tokens_out":7769,"would_cite":true,"duration_ms":82376,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural-network model of surprise and noise reproduces the mobile-paradigm results in simulations, and ablations show each component is needed.","keywords":["mobile paradigm","sensorimotor contingency","infant development","action-outcome prediction","novelty-based exploration","motor noise","computational model","predictive coding"],"falsifier":"If a reanalysis of the two infant datasets, matching for overall kick rate and attention time, showed that conjugate and binary conditions produce equally fast connected-limb differentiation, then the model's mechanism (threshold crossing creating a larger sensory change) would be falsified as the explanation for the binary advantage.","tokens_in":1669,"feed_emoji":"👶","tokens_out":2470,"duration_ms":79191,"temperature":0.7,"pith_summary":"The paper sets out to explain a classic developmental finding: when a mobile is tied to one limb, infants come to move that limb more, and they do so only when the mobile's motion really is contingent on their movements. The authors argue that a neural-network model that predicts the sensory outcome of its actions, seeks out surprising actions, and exercises a redundant motor system with many muscle commands and motor noise is sufficient to reproduce the mobile-paradigm data. In simulations with 20 virtual infants per condition, the model distinguishes the connected limb, shows the contingent-versus-non-contingent gap, is more reliable in the binary than the conjugate version, and occasionally produces an extinction burst after the ribbon is cut. Ablation studies are the load-bearing part: removing prediction error, removing novelty-based exploration, shrinking the muscle-command set, or setting motor noise too low or too high prevents the correct behavior. If the model is right, these mechanisms are not just possible aids but necessary ingredients for this form of infant sensorimotor learning.","feed_headline":"Surprise-driven simulation reproduces infants' mobile-paradigm learning","feed_subtitle":"Removing prediction, exploration, motor noise, or muscle redundancy breaks the model's limb learning.","key_machinery":"The engine is the activity interest map, a $4\\times 10$ table that discretizes each limb's activity into ten ranges and assigns each range an interest value. After each step, the sensory prediction error is compared to a novelty threshold; surprise sets the visited range to 1, and lack of surprise decrements it by 0.1, so the agent repeatedly re-enters ranges whose sensory effects it cannot yet predict. A second mechanism, the fixed random projection from 600 muscle commands to the four limb scalars, is sampled from a $\\beta$ distribution with $z_1=0.01$, $z_2=0.1$ so that each limb is affected by only a handful of commands; this makes credit assignment non-trivial because many commands co-activate all limbs, so the network must learn which outputs actually drive the connected limb. The neural network is trained on three mean-squared-error losses (sensory prediction error, distance to the most interesting limb activity, and deviation from a slowly rising baseline), so that exploration and prediction pulling against each other generate the behavior.","core_discovery":"The central claim is that a minimal cognitive architecture built around action-outcome prediction and surprise-driven exploration can account for the four main mobile-paradigm findings without any explicit reward or notion of agency. The model contains no reward; it selects limb activities through an activity interest map in which surprising outcomes raise interest and predictable outcomes lower it. A neural network outputs 600 abstract muscle commands that are projected through a fixed random sparse matrix onto four scalar limb activities, and motor noise is added before the limb state is updated. Each run is treated as one infant; the network learns by backpropagation on three losses: prediction error, distance from the currently most interesting activity, and distance from a preferred baseline. The paper reports that this reproduces preferential movement of the connected limb, higher activity in contingent than in non-contingent replay controls, the larger effect for binary than for conjugate coupling, and a non-systematic extinction burst, and that ablation of any core component except the baseline term destroys the replication.","pith_inferences":["Inference: The model could be read as predicting that infants should detect a time-lagged contingency less well, because the prediction error at the moment of feedback would not align with the action that caused it; a delayed-feedback mobile experiment would test this directly.","Inference: The same surprise-driven architecture might be applied to vocalization paradigms, where the 'limb' is replaced by vocalization effort; the model would predict that a contingent adult response is learned only when the infant's motor variability spans the response threshold.","Inference: The ablation claim depends on the particular random muscle-command projection; an alternative implementation with a structured, low-dimensional body model might require far fewer commands, so the '600 commands' number is likely a property of this architecture rather than a general biological bound.","Inference: The model's low within-run variability suggests that adding attention or habituation mechanisms would make it testable against more granular data, such as individual kick-rate time series."],"forward_implications":["The paper's account implies that the same architecture should transfer to other sensorimotor contingencies, such as vocalization-contingent sounds, without adding any reward or agency machinery.","The ablation results suggest that internal prediction and surprise are not optional refinements; they are the mechanism that makes the connected limb discoverable, so developmental theories of the mobile paradigm should include a predictive component.","The model explains the binary-conjugate asymmetry: threshold crossing turns small activity changes into large sensory changes, making the connected limb easier to identify, so conjugate studies should show weaker and slower differentiation.","The occasional extinction burst is a natural consequence of a surprise-driven explorer; it will not appear in every run, matching the inconsistent reports in the literature.","Reducing the muscle-command count below about 100 in the binary condition or about 300 in the conjugate condition makes the model fail in some runs, suggesting that motor redundancy itself contributes to reliable learning."],"supporting_citations":[{"why":"Introduces the mobile paradigm and the conjugate-reinforcement setup that the simulations target.","marker":"[5]"},{"why":"Supplies the binary-condition infant data (threshold-triggered mobile activation) against which model output is compared.","marker":"[12]"},{"why":"Supplies the non-binary (conjugate) infant data and the arm-differentiation result the model reproduces.","marker":"[11]"},{"why":"Provides the infomax predictive approach to contingency detection that the model's surprise-based exploration builds on.","marker":"[22]"},{"why":"Presents the earlier minimalist babybot simulation, the main non-predictive baseline the paper argues is insufficient for extinction bursts.","marker":"[25]"},{"why":"Presents the dynamical-systems alternative without prediction, which the paper contrasts with its own predictive account.","marker":"[26]"},{"why":"Documents non-contingent controls and extinction-related findings, used for the control-group and extinction-burst comparisons.","marker":"[18]"}],"fun_headline_variants":["Surprise, not reward, drives infant mobile-learning in simulation","No-reward model replicates infant mobile-paradigm behavior","Simulation reproduces infant limb choice via prediction errors","Model mimics infant learning without any reward signal","Computational infant learns mobile contingency by surprise"],"cache_read_input_tokens":27392,"weakest_assumption_plain":"The whole behavior emerges through a fixed random projection of 600 abstract muscle commands onto a single scalar activity per limb, with no other sensory input; if real limb control is structured differently, the model's similarity to infants may be a coincidence of that abstraction.","fun_headline_variants_meta":{"raw":{"variants":["Surprise, not reward, drives infant mobile-learning in simulation","No-reward model replicates infant mobile-paradigm behavior","Simulation reproduces infant limb choice via prediction errors","Model mimics infant learning without any reward signal","Computational infant learns mobile contingency by surprise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000469,"raw_usage":{"total_tokens":2363,"prompt_tokens":998,"completion_tokens":1365,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":1290}},"tokens_in":614,"tokens_out":1365,"duration_ms":11473,"temperature":1.0,"reasoning_tokens":1290,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:29:01.527539+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a reanalysis of the two infant datasets, matching for overall kick rate and attention time, showed that conjugate and binary conditions produce equally fast connected-limb differentiation, then the model's mechanism (threshold crossing creating a larger sensory change) would be falsified as the explanation for the binary advantage.","supporting_citations":[{"cited_title":"Conjugate reinforcement of infant exploratory behavior,","cited_arxiv_id":null,"evidence_quote":"Introduces the mobile paradigm and the conjugate-reinforcement setup that the simulations target."},{"cited_title":"6-Month-Old Infants’ Sensitivity to Contingency in a Variant of the Mobile Paradigm With Proximal Stimulation Studied at Fine Temporal Resolution in the Laboratory,","cited_arxiv_id":null,"evidence_quote":"Supplies the binary-condition infant data (threshold-triggered mobile activation) against which model output is compared."},{"cited_title":"Development of body knowledge as measured by arm differentiation in infants: From global to local?,","cited_arxiv_id":null,"evidence_quote":"Supplies the non-binary (conjugate) infant data and the arm-differentiation result the model reproduces."},{"cited_title":"Detecting contingencies: an infomax approach,","cited_arxiv_id":null,"evidence_quote":"Provides the infomax predictive approach to contingency detection that the model's surprise-based exploration builds on."},{"cited_title":"Can infants’ sense of agency be found in their behavior? Insights from babybot simulations of the mobile-paradigm,","cited_arxiv_id":null,"evidence_quote":"Presents the earlier minimalist babybot simulation, the main non-predictive baseline the paper argues is insufficient for extinction bursts."},{"cited_title":"The coordination dynamics of mobile conjugate reinforcement,","cited_arxiv_id":null,"evidence_quote":"Presents the dynamical-systems alternative without prediction, which the paper contrasts with its own predictive account."},{"cited_title":"Topographical response differentiation and reversal in 3-month-old infants,","cited_arxiv_id":null,"evidence_quote":"Documents non-contingent controls and extinction-related findings, used for the control-group and extinction-burst comparisons."}],"review_version":1}