{"id":"6c6478b3-1774-405b-b4fa-59ae60cc0cf1","arxiv_id":"2501.06698","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper claims Bayesian Efficient Coding best simulates human navigation in VR, but its loss functions fit each model to a noise-corrupted version of its own output, so the comparison is close to circular.","lead":"This paper compares three mathematical models of sensory coding on virtual reality navigation data and reports that Bayesian Efficient Coding fits best. The comparison is weakened by a self-referential fitting setup and by evaluating only one participant against a theoretical benchmark.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BEC/FMC evaluation is circular: in Eqs. (1) and (2) the target y_noise is generated by the same nonlinearity g1 being optimized, so the reported superiority over LNP is an artifact of self-generated data.","rationale":"I agree with the reader's identification of the circular target as the weakest assumption. The paper's abstract claim depends on Table II, which compares fitted curves to an Infomax benchmark, but the BEC/FMC targets are generated by the same parametric function being optimized. This is not a question of disagreement with scientific consensus; it is an internal check on the evaluation. The Limitations section acknowledges model simplifications but does not flag this self-referential construction. A simple re-fit against real physiological data would settle the issue. Because the central empirical claim is unsupported by the reported analysis, the rejection verdict stands.","tokens_in":8299,"tokens_out":4455,"duration_ms":44791,"concrete_test":"Re-run the BEC and FMC optimizations with y_noise,i set to the actual measured phasic EDA signal (or, if that is not the intended target, to a held-out participant's navigation and physiological data) instead of g1(xstd,i; θ) + ε, keeping the loss functions otherwise unchanged, and evaluate on a held-out test split. If the BEC/FMC MSE to the Infomax response rises toward LNP levels, or if BEC no longer ranks first, the reported superiority is an artifact of the self-generated target.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison is fixed by Eqs. (1) and (2): in both BEC and the BEC+FMC variant, the target y_noise,i is defined as g(xstd,i; θ) + ε, where g is the same parametric tanh nonlinearity g1(·; θ) whose parameters are being optimized. For FMC, the same synthetic target also appears inside the weight w(y_noise,i) = |y_noise,i|. Minimizing either loss therefore recovers the generating parameters from noisy samples of the same function; there is no independent physiological or behavioral signal being predicted. The MSE values in Table II are computed against the Infomax response, a theoretical curve derived from the stimulus CDF, not against held-out physiological or behavioral data. The LNP model, by contrast, is fit to a different quantity (the mean response yi) with an additional penalty term, so the comparison is not like-for-like. Consequently, the abstract's claim that BEC 'generally outperforms the others' is an artifact of evaluating a model on noisy samples of itself. The single-participant analysis and the acknowledged model simplifications in the Limitations section do not repair this structural circularity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares Bayesian Efficient Coding (BEC), Fitness Maximizing Coding (FMC), and Linear-Nonlinear Poisson (LNP) models for simulating human navigation behavior in virtual reality. Using standardized navigation speed from one participant in a public VR dataset, the authors optimize each model's response and compute mean squared error (MSE) against the Infomax response curve. The central claim is that BEC generally outperforms the other models, with BEC+FMC competitive at small error penalties. The evaluation is based entirely on MSE to the Infomax theoretical benchmark for a single participant; physiological signals are described but not used as the prediction target.","tokens_in":8620,"tokens_out":4446,"duration_ms":37749,"significance":"If the comparison were valid, the paper would offer a straightforward benchmark of normative models for VR sensory encoding. The manuscript has some strengths: it provides explicit loss functions, uses a public dataset, and reports numbers for all p values. However, the evaluation is circular for BEC and BEC+FMC (Eqs. (1)-(2)), and no physiological or behavioral ground truth is used as the target. Consequently, the reported superiority of BEC is an artifact of fitting a model to its own output. The single-participant analysis further limits any general claim.","major_comments":[{"comment":"The optimization target y_noise,i is defined as g(xstd,i; θ) + ε, where g is exactly the tanh nonlinearity g1(xstd,i; θ) whose parameters θ are being minimized. For p=2, the unregularized least-squares solution recovers g itself (up to noise), so the low MSE in Table II for BEC is expected by construction and cannot support the claim that BEC 'generally outperforms' the alternatives.","section":"Section III-A, Eq. (1)"},{"comment":"The same circularity applies to the BEC+FMC variant, since the weight w(y_noise,i)=|y_noise,i| is computed from the same self-generated target. The comparison with the LNP model is therefore not like-for-like: LNP is fit to the mean response yi (Eq. (3)) while BEC/FMC are fit to noisy samples of their own output.","section":"Section III-B, Eq. (2)"},{"comment":"The MSE values are computed against the Infomax response, a theoretical curve derived from the stimulus CDF (Section IV), not against the phasic EDA ground truth claimed in Section III-D. Thus the metric assesses proximity to a normative curve, not predictive accuracy for physiological responses or navigation behavior.","section":"Section IV, Table II"},{"comment":"The entire analysis uses data from one participant, and the Discussion's interpretation of that participant's SSQ and gaming experience is anecdotal. No cross-validation, hold-out data, or statistical testing across participants is provided, so the abstract's 'superior predictions in most scenarios' is not supported.","section":"Section IV"}],"minor_comments":[{"comment":"The function g1(xstd,i; θ) is said to represent 'LMC response', but LMC is never defined; presumably it stands for 'large monopolar cell' as in the figures, but the acronym is not introduced in the text.","section":"Section III-A"},{"comment":"The variable yi is defined as 'the mean value' without specifying the quantity being averaged; clarify whether it is the mean of the phasic EDA signal or of some other physiological measure.","section":"Section III-C, Eq. (3)"},{"comment":"The statement that 'the Infomax response curve being almost identical to the stimulus CDF curve' is a tautology because the Infomax response is defined as the stimulus CDF; this should be acknowledged rather than presented as an empirical finding.","section":"Section IV-A"},{"comment":"The parameter p is described as an 'error penalty', but in Eqs. (1)-(3) it is the exponent of an Lp loss; calling it a penalty conflates the loss exponent with a regularization weight such as w_penalty used for LNP.","section":"Throughout"},{"comment":"There is a typo in 'reprsents' for 'represents'; similar typos ('impaire', 'make navigation decisions') appear in Sections II-B and VI.","section":"Section III-B"},{"comment":"The irregular spacing of p values (0.10, 0.31, 0.52, ...) is unusual; state how these values were chosen and whether they are used identically for all three models.","section":"Table II"}],"recommendation":"reject","confidential_remarks":"The circularity in Eqs. (1)-(2) is severe and invalidates the main comparison. If the authors have access to the full multi-participant dataset, a re-analysis using the phasic EDA or navigation behavior as the target could be a viable path, but as submitted the central claim is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the headline: the central comparison in this paper is circular. In Eqs. (1) and (2), the target y_noise is defined as g(xstd; θ) + ε, where g is the same parametric tanh nonlinearity whose parameters θ are being optimized. For p=2, the optimal g1 is exactly the generating function, so the low MSE for BEC and FMC is guaranteed by construction. The paper then compares these fitted responses to the Infomax curve, not to any held-out physiological or behavioral measurement. So the abstract's claim that BEC 'generally outperforms' the others is an artifact of evaluating a model on noisy samples of itself.\n\nWhat is new here: applying BEC, FMC, and LNP to VR navigation data is a reasonable idea, and the paper is organized clearly. It uses a public dataset and includes a limitations section. The discussion of cybersickness and efficient coding is a relevant direction. But the execution does not support the conclusions.\n\nThe soft spots are substantial. The circularity is the first. Second, the LNP model is fit to a different quantity (the mean response yi) with an additional penalty, so the three-way comparison is not like-for-like. Third, results are for one participant only. Fourth, the Infomax benchmark is a theoretical curve from the stimulus CDF, not an empirical physiological signal. The limitations section lists model simplifications but does not acknowledge that the BEC/FMC evaluation is structurally tautological.\n\nWho is this for? Someone interested in cheap model comparisons might skim it, but as a scientific claim it does not hold. The paper deserves a desk reject; if the authors rework the evaluation to predict actual GSR or navigation behavior with held-out data, a resubmission would be worth considering. I would not cite it in its current form.","headline":"BEC/FMC evaluation is circular; the reported superiority is an artifact of fitting a model to its own noisy output.","tokens_in":9076,"tokens_out":2885,"would_cite":false,"duration_ms":26853,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a Bayesian efficient coding model best predicts physiological responses during VR navigation, outperforming fitness-maximizing and LNP models.","keywords":["sensory encoding","Bayesian efficient coding","fitness maximizing code","linear-nonlinear-Poisson model","virtual reality navigation","physiological response","cybersickness","Infomax"],"falsifier":"Run the same three models on the same dataset but replace the target $y_{\\mathrm{noise},i}=g_1(x_{\\mathrm{std},i};\\theta)+\\varepsilon$ with the participant's measured phasic EDA values, or with another participant's values, and recompute the MSE to the Infomax response; if BEC no longer has the lowest MSE, the reported superiority is an artifact of the self-referential target construction.","tokens_in":8138,"feed_emoji":"🥽","tokens_out":6652,"duration_ms":63231,"temperature":0.7,"pith_summary":"This paper asks which sensory encoding model best simulates how a person's physiological state changes while navigating in virtual reality, using navigation speed as the input and phasic electrodermal activity as the ground truth. It compares Bayesian Efficient Coding (BEC), Fitness Maximizing Coding (FMC), and the Linear-Nonlinear Poisson (LNP) model by fitting each model's optimized response to an Infomax benchmark, then ranking them by mean squared error across a range of error penalties $p$. The central claim is that BEC generally produces the most accurate predictions, that BEC with an FMC-style loss is slightly more accurate when the error penalty is small, and that LNP is markedly less accurate and more sensitive to $p$. If correct, this gives VR designers a principled way to anticipate discomfort, fatigue, and cognitive workload from motion data alone.","feed_headline":"Bayesian coding best predicts VR navigation strain","feed_subtitle":"Three sensory models compared on speed and physiological data; BEC wins on accuracy, FMC edges small-error cases.","key_machinery":"The central object is the parametric nonlinearity $g_1(x;\\theta)=\\frac{1}{2}(1+\\tanh(\\theta_1 x+\\theta_2))$, which represents the encoded physiological response to standardized navigation speed. BEC minimizes $\\mathrm{Loss}_{\\mathrm{BEC}}=\\sum_i (y_{\\mathrm{noise},i}-g_1(x_{\\mathrm{std},i};\\theta))^p$, while the FMC variant adds a weight $w(y_{\\mathrm{noise},i})=|y_{\\mathrm{noise},i}|$ to the same $L^p$ error. The fitting target is $y_{\\mathrm{noise},i}=g_1(x_{\\mathrm{std},i};\\theta)+\\varepsilon$ with Gaussian noise $\\varepsilon\\sim N(0,0.05^2)$. The LNP model instead fits a quadratic rectified nonlinearity $g_2$ under a Poisson process assumption. The benchmark is the Infomax response, defined as the cumulative distribution function of the standardized stimulus, which represents maximum information preservation; mean squared error against this benchmark carries the comparison.","core_discovery":"On the paper's own terms, the Bayesian Efficient Coding framework, which minimizes posterior entropy under a resource constraint, produces optimized responses closest to the Infomax response for a participant navigating in VR, with mean squared error around $0.0025$ to $0.0029$ across $p=0.10$ to $2.00$. Combining BEC with the Fitness Maximizing Code loss slightly lowers the error for $p<0.73$ but slightly raises it for $p\\geq 0.73$, which the authors interpret as a trade-off between accuracy and long-term fitness. The LNP model yields much larger and more variable errors, roughly $0.055$ to $0.53$, suggesting that its linear filtering and Poisson spiking assumptions are a poor match for GSR-derived physiological signals. The paper concludes that BEC is the more accurate framework for simulating human navigation behavior in VR.","pith_inferences":["The quantitative comparison uses one participant, so an immediate extension is to fit each of the 53 participants separately and test whether BEC's advantage persists across individuals.","Because the fitting target is a noise-corrupted copy of the same function being optimized, a stronger test would use independently measured phasic EDA as the target; if BEC still wins, the claim would be much more secure.","If the claim holds against independent targets, a practical upshot is that speed-only streams from a headset could feed a real-time discomfort and workload estimator.","The crossover near $p=0.73$ between BEC and FMC is a testable prediction about how error cost shifts encoding strategy when task urgency or discomfort changes."],"forward_implications":["If BEC is the accurate encoding model, VR systems could predict user discomfort from navigation speed alone, before symptoms become visible.","FMC's better small-$p$ performance suggests that when the cost of error is low, fitness-maximizing encoding captures the user's exploration behavior.","LNP's unstable performance across $p$ discourages its use for GSR-derived physiological signals in VR.","The near-constant MSE for BEC across $p$ means the main result does not hinge on a particular penalty value, simplifying practical deployment."],"supporting_citations":[{"why":"Supplies the efficient coding hypothesis that BEC and the Infomax benchmark build on.","marker":"[21]"},{"why":"Defines the Infomax principle used as the benchmark response in the comparison.","marker":"[22]"},{"why":"Provides the Bayesian efficient coding framework whose loss function is adapted in this paper.","marker":"[23]"},{"why":"Supplies the fitness maximizing code framework and the earlier-stage accuracy-versus-fitness finding the paper contrasts with.","marker":"[24]"},{"why":"Supplies the LNP model structure used as the third comparator.","marker":"[27]"},{"why":"Provides the VR navigation and biosignal dataset used for fitting the models.","marker":"[29]"}],"fun_headline_variants":["Bayesian coding best predicts VR navigation strain","BEC tops sensory models for VR navigation","Bayesian efficient coding wins VR navigation test","VR navigation strain best modeled by BEC","BEC beats LNP in VR navigation encoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The noisy physiological target the models are fitted to is generated by adding small Gaussian noise to the same curve the model is trying to learn, so the near-perfect fits for BEC and FMC may reflect self-consistency rather than genuine predictive power.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian coding best predicts VR navigation strain","BEC tops sensory models for VR navigation","Bayesian efficient coding wins VR navigation test","VR navigation strain best modeled by BEC","BEC beats LNP in VR navigation encoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000133,"raw_usage":{"total_tokens":1092,"prompt_tokens":861,"completion_tokens":231,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":164}},"tokens_in":477,"tokens_out":231,"duration_ms":2639,"temperature":1.0,"reasoning_tokens":164,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:54:11.975102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same three models on the same dataset but replace the target $y_{\\mathrm{noise},i}=g_1(x_{\\mathrm{std},i};\\theta)+\\varepsilon$ with the participant's measured phasic EDA values, or with another participant's values, and recompute the MSE to the Infomax response; if BEC no longer has the lowest MSE, the reported superiority is an artifact of the self-referential target construction.","supporting_citations":[{"cited_title":"Possible principles underlying the transforma- tion of sensory messages,","cited_arxiv_id":null,"evidence_quote":"Supplies the efficient coding hypothesis that BEC and the Infomax benchmark build on."},{"cited_title":"Self-organization in a perceptual network,","cited_arxiv_id":null,"evidence_quote":"Defines the Infomax principle used as the benchmark response in the comparison."},{"cited_title":"Bayesian efficient coding,","cited_arxiv_id":null,"evidence_quote":"Provides the Bayesian efficient coding framework whose loss function is adapted in this paper."},{"cited_title":"Sensory perception relies on fitness- maximizing codes,","cited_arxiv_id":null,"evidence_quote":"Supplies the fitness maximizing code framework and the earlier-stage accuracy-versus-fitness finding the paper contrasts with."},{"cited_title":"Characterization of neural responses with stochastic stimuli,","cited_arxiv_id":null,"evidence_quote":"Supplies the LNP model structure used as the third comparator."},{"cited_title":"Dataset for predicting cybersickness from a virtual navigation task","cited_arxiv_id":"2303.13527","evidence_quote":"Provides the VR navigation and biosignal dataset used for fitting the models."}],"review_version":1}