{"id":"e699d803-f81d-416d-b699-695870dd90bc","arxiv_id":"2608.05642","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Bayesian model equating the value of trying a design option with expected uncertainty reduction predicts, and experiments confirm, that exploration is most valuable when knowledge generalizes moderately across neighboring options.","lead":"This paper proposes a model that values a design trial by how much it is expected to shrink uncertainty about the relationship between design choices and outcomes. Simulations and human experiments show that the most valuable trials come from environments with moderate generalizability, where knowledge from one trial extends to similar but not identical options.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'value of trying' rating is an unvalidated proxy for V_epi, and its failure to show the predicted sigma effect (H2-1) means the experimental confirmation of the B-EUR mechanism is not established; the Section 4.4 explanation is post hoc.","rationale":"The reader's verdict is CONDITIONAL, and my read does not change that. The paper's central claim has two components: a formal identity (Eq. 3) and an empirical pattern. The formal identity is a standard active-inference definition and is not in dispute. The empirical component rests on the assumption that the 'value of trying' rating measures V_epi. That assumption is the least secure link because it is unvalidated and because one of the model's signature predictions, the monotonic effect of sigma on V_epi, is absent in the primary rating measure. If the rating tracks curiosity, visual complexity, or perceived uncertainty instead, the observed inverted-U with l is predicted by classical arousal theory (Berlyne, 1970), which the paper itself cites, so the data would not uniquely support B-EUR. A trial-level model comparison, using the already-collected data and the paper's own kernel, would settle whether ratings track the model's mutual information. A secondary concern is that the simulation's update width is set proportional to the environmental l (Appendix Table 4), so the agent is given the very quantity it is supposed to discover; however, this affects the strength of the theoretical prediction rather than the empirical claim, and it is also testable by fixing l_model. Both concerns are addressable without rejecting the framework, so CONDITIONAL remains the appropriate verdict.","tokens_in":20396,"tokens_out":10585,"duration_ms":100953,"concrete_test":"Re-analyze Experiment 1 at the trial level: for each participant, action, and step, compute the B-EUR model's V_epi(u_t) (Eq. 3) using the true sigma and l, the participant's observation history, and the Gaussian-weighted update kernel (l_model = beta*l, Appendix Table 4). Fit a mixed model predicting the within-participant standardized 'value of trying' rating from V_epi, controlling for perceived uncertainty, local predictive variance, distance to the nearest observed point, and a visual-salience proxy (e.g., vertical separation from neighboring points). If V_epi is not positive and significant after controls, or if a control predicts ratings better, then the rating does not specifically measure expected uncertainty reduction and the experimental confirmation of the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim requires that the rating 'How valuable do you feel it is to know the value of this point?' (Section 3.1.4) measure the model's V_epi (Eq. 3). No trial-level validation is provided: the analyses only test aggregate hypotheses, not whether ratings track the model's mutual information. The strongest discriminator between V_epi and a generic uncertainty/interest response is the sigma effect: V_epi increases with sigma in simulation (S3, Figure 3(B)), and H2-1 predicts this in ratings, yet the effect is null (beta = -0.030, SE = 0.054, p = .572). The paper's Section 4.4 explanation, that effects of l dominated ratings, is post hoc and does not follow from the model, since Figure 3(B) shows sigma raises V_epi across all l values, including the l levels used in Experiment 1. If the rating instead tracks perceived uncertainty or visual 'interestingness', the confirmed inverted-U with l is equally consistent with a general preference for intermediate complexity (Berlyne, 1970), not with expected uncertainty reduction. The sigma effect in Experiment 2 choice (H4-1) is also ambiguous because participants chose based on static graph examples, so amplitude salience, not epistemic value, may drive it. Without a validated measure, the experiments do not test the B-EUR mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Bayesian Expected Uncertainty Reduction (B-EUR) model, which formalizes the value of trying a design action as the expected mutual information between the action-outcome mapping and the predicted observation (Eq. 3). Using a graph-shape guessing task with Gaussian-process environments, the authors simulate how epistemic value depends on correlation length l (generalizability) and amplitude sigma (outcome discriminability), predicting an inverted-U in l and a monotonic increase in sigma. Two human experiments test these predictions: Experiment 1 measures perceived uncertainty, subjective value of trying, and enjoyment; Experiment 2 measures choice among eight graph conditions. The human data support the inverted-U relationship with l for value of trying, enjoyment, and choice probability, and the sigma effect appears in perceived uncertainty and choice behavior, but not in the subjective value-of-trying or enjoyment ratings. The paper discusses implications for prototyping, problem framing, and design management, and acknowledges limitations concerning task abstraction and the exclusion of pragmatic value.","tokens_in":20663,"tokens_out":3056,"duration_ms":35141,"significance":"If the empirical validation holds, the paper offers a useful computational articulation of one mechanism in the UDA framework, connecting active-inference epistemic value to design exploration. The strengths include a clearly specified model, explicit simulation-to-hypothesis mapping, a self-contained abstract task, and apparent availability of code and data in a repository. The inverted-U relationship with generalizability is internally consistent and independently supported in both ratings and choice behavior, which is a meaningful addition to design-research accounts of exploration. However, the central experimental confirmation of the B-EUR mechanism depends on an unvalidated rating measure, and the sigma-effect predictions fail in the key rating data; the current manuscript therefore falls short of establishing that the human behavior is driven by expected uncertainty reduction rather than by a more general preference for intermediate complexity or visual salience.","major_comments":[{"comment":"The main empirical test of the B-EUR mechanism rests on the 'value of trying' rating as a subjective indicator of V_epi (Eq. 3), but the paper provides no validation that this rating actually tracks the model's trial-level mutual information. The model generates trial-level predictions for V_epi(u_t) under the participant's evolving belief state, and the experiment collects a rating at each step, yet the analyses aggregate over steps and conditions and test only ordinal hypotheses. This gap is load-bearing because the model's sigma prediction S3 directly fails in the rating data: H2-1 is not supported (beta = -0.030, SE = 0.054, p = .572), even though Figure 3(B) shows that sigma raises V_epi across the full l range used in Experiment 1. The Section 4.4 explanation that 'effects of l dominated' is post hoc and does not follow from the model. A trial-level analysis of ratings against model-derived V_epi values, or a pre-registered alternative validation of the rating measure, is necessary before the experiments can be said to confirm the B-EUR mechanism rather than a generic preference for intermediate complexity.","section":"Section 3.1.4 and Section 3.2.1"},{"comment":"The simulation's environmental l manipulation is partially confounded with the agent's learning kernel through the setting l_model = beta * l. Because the agent is given the environmental smoothness in its own Gaussian-weighted update rule, the simulation's inverted-U relationship between l and V_epi (S2, Figure 3(B)) may be a built-in consequence of the update-width calibration rather than an emergent property of the task environment alone. To make the simulation claim robust, the authors should report a mismatch analysis, for example fixing l_model while varying l, or estimating l_model from the observed rating data. The human inverted-U remains informative, but the current simulation setup overstates the independence of the prediction from the model's own assumptions.","section":"Appendix A, Table 4"},{"comment":"The choice experiment (Experiment 2) presents examples of all eight graph conditions simultaneously and asks participants to select the one that 'looks most enjoyable.' The observed sigma effect on choice probability (H4-1: beta = 1.035, p < .001) may therefore reflect visual amplitude salience or hedonic preference rather than expected uncertainty reduction, and it is not a process-level measure of V_epi. The Discussion (Section 4.4) itself suggests that sigma became more 'salient' when conditions were compared at once, which is consistent with this alternative explanation. A manipulation that decouples amplitude salience from epistemic value, or a model-based choice analysis using the participant's inferred belief state, is needed before H4-1 can be interpreted as support for S3.","section":"Section 3.1.4 and Section 3.2.2"}],"minor_comments":[{"comment":"The relationship between prior entropy H[q(B_t)] and V_epi is described as an inverted-U but is only shown graphically; no statistical test or fitted curve is reported for this claim, despite it being presented as a standalone simulation result.","section":"Section 2.2, Figure 3(D)"},{"comment":"The free parameters beta, Delta, eta, gamma, and alpha0 are listed in Table 4, but the paper does not justify their values or report sensitivity analyses. In particular, the discretization width Delta and the update-scale eta directly affect the computed V_epi magnitudes and the decay rates.","section":"Section 2.1.3 and Appendix A"},{"comment":"The Gaussian-process setup is cited to Angel Deborah et al. (2021) and Han et al. (2012), which are not standard references for GP regression; a canonical textbook or methodological reference would be more appropriate.","section":"Section 2.1.1"},{"comment":"In the analysis of H2-3/H2-4, the variable name 'rep c' appears in the interaction notation; this should be defined or renamed to avoid confusion with 'repetition' across sessions.","section":"Section 3.2.1"},{"comment":"The repository link is indicated as 'an online repository' but no URL or DOI is provided; the reproducibility claim cannot be verified in the current manuscript.","section":"Supplementary materials"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid simulation foundation and the inverted-U with l is empirically robust across ratings, enjoyment, and choice. The main concern is construct validity: the value-of-trying rating is the only direct measure of V_epi, and its failure to show the predicted sigma effect, combined with the absence of any trial-level validation against the model, means the central mechanism is not yet established. The Section 4.4 explanation reads as post hoc. I would advise the editor that revision is warranted, but that the required work—trial-level model-based validation or a new pre-registered experiment—is substantial and goes beyond local rewriting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the take: the formal model is standard active-inference epistemic value (expected information gain). The authors know this and cite the right sources. The new contribution is the operational mapping of two design-relevant environmental properties onto GP parameters, and a clean experimental package that tests the resulting predictions. That package is the reason to take this seriously.\n\nThe simulation work is solid. The inverted-U with l emerges from the mutual-information calculation rather than being baked in, and the temporal decay results (S4) are a nice extra. The human experiments are also well executed: 44 participants, explicit hypotheses, LMM/GLMM with participant random effects, within-participant standardization. The inverted-U for the value of trying and enjoyment, and in choice, is consistent and statistically clear. That is a real empirical finding.\n\nNow the soft spots, in rough order of severity.\n\nFirst, the link between the rating and the model's V_epi is not established. The paper treats \"How valuable do you feel it is to know the value of this point?\" as a direct report of V_epi, but no trial-level analysis shows that the rating tracks the model's mutual information. The place this hurts most is H2-1: the model predicts sigma increases V_epi across all l levels used in Experiment 1, yet the rating shows no sigma effect. The Section 4.4 explanation—that l dominated the ratings—does not follow from the model, since Figure 3(B) shows sigma's effect at every l. So the confirmation of the inverted-U is equally consistent with Berlyne-style preference for intermediate complexity, which the paper itself cites. That leaves the B-EUR mechanism, as opposed to a general arousal potential account, untested by the rating data. The sigma effect in choice (H4-1) is also ambiguous because participants chose from static graph examples; amplitude salience, not expected uncertainty reduction, could drive that.\n\nSecond, the simulation hands the agent the environmental length scale: the belief-update width is l_model = βl (Appendix A). With β fixed, the agent's inductive bias matches the generative process, so the l effect on epistemic value is partly self-referential. A robustness check with mismatched l_model would tell you how much of the inverted-U depends on this gift.\n\nThird, the text says data and code are in an online repository, but no URL appears in the manuscript. If that link is present in the actual submission, fine; if not, a referee should require it.\n\nThese are fixable. The paper deserves a serious referee and likely a conditional accept after revision. The core conceptual contribution is modest—it is a translation of known active-inference formalism into a design-research setting—but the translation is careful and the empirical work is a genuine step toward testing uncertainty-driven action selection in design. I'd want a revision that validates the rating measure (e.g., model-based trial-level predictions) or reframes the claims accordingly, and that addresses the l_model circularity.","headline":"A careful translation of active-inference epistemic value into design exploration, with solid simulations and experiments; the empirical validation of the B-EUR mechanism is weakened by an unvalidated rating measure and a self-referential simulation, but the paper deserves serious peer review.","tokens_in":21232,"tokens_out":3279,"would_cite":false,"duration_ms":34550,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that the value of trying a design option is the expected reduction of uncertainty about action–outcome relations, that this value peaks at moderate generalizability, and that it grows with how distinguishable outcomes are.","keywords":["B-EUR model","design exploration","epistemic value","expected uncertainty reduction","Gaussian process","generalizability","outcome discriminability","inverted-U relationship"],"falsifier":"A concrete test would be to fit the B-EUR model's predicted $V_{\\mathrm{epi}}(u_t)$ to each participant's trial-by-trial value-of-trying ratings; if the ratings do not track those predictions once the main effects of $l$, $l^2$, and $\\sigma$ are controlled, or if visual features of the graph predict the ratings better, the claim that subjective value is expected uncertainty reduction fails.","tokens_in":20172,"feed_emoji":"🎯","tokens_out":11522,"duration_ms":104174,"temperature":0.7,"pith_summary":"The paper proposes the B-EUR model, in which the value of trying a design action is defined as the expected reduction of uncertainty about the relationship between actions and outcomes. The model predicts an inverted-U relationship between that value and generalizability: trials are most informative when knowledge from one trial extends to nearby options, but not so much that trials become redundant. It also predicts that value rises when outcome differences are easier to distinguish. Simulations confirm both predictions, and human experiments confirm the inverted-U in ratings of value and enjoyment and in choices, while outcome discriminability appears in perceived uncertainty and choices. If correct, the model explains why some explorations feel worth pursuing and gives concrete guidance for choosing prototypes, framing problems, and collecting feedback.","feed_headline":"Moderate generalizability makes design trials most worth trying","feed_subtitle":"Expected uncertainty reduction is highest at intermediate generalization, and human choices and enjoyment follow that curve.","key_machinery":"The load-bearing object is the B-EUR epistemic value $V_{\\mathrm{epi}}(u_t)=I(B_t; s_{t+1}\\mid u_t)$, the mutual information between the belief $q(B_t)$ about the action–outcome mapping and the next observed state $s_{t+1}$ given action $u_t$. In words, it is how much the learner expects one trial to shrink its uncertainty about that mapping. The environment is generated as a Gaussian process with an RBF (smooth Gaussian) kernel $k(u,u')=\\sigma^2\\exp(-(u-u')^2/2l^2)$, where the correlation length $l$ operationalizes generalizability and the amplitude $\\sigma$ operationalizes outcome discriminability. Observations update beliefs with Gaussian weights whose width is proportional to $l$, so each trial informs neighboring candidate points. This machinery ties one information-theoretic quantity to the two environmental knobs and produces the inverted-U and monotonic predictions tested in humans.","core_discovery":"The paper's central claim is that epistemic value—the worth of trying a candidate design action—is the mutual information between the agent's current belief about the action–outcome mapping and the observation that action would produce. On the paper's own terms, the discovery is that this quantity has a clear inverted-U dependence on generalizability: it is low when observations fail to transfer to neighboring candidates and low again when a few observations exhaust what can be learned, with a maximum at intermediate correlation length. In simulations, the same quantity increases with outcome discriminability because wider output ranges leave more room for learning. In the human graph-shape guessing task, subjective value of trying and enjoyment follow the predicted inverted-U with generalizability, and participants more often choose conditions with intermediate generalizability and larger outcome discriminability; the outcome-discriminability effect appears in choices but not in absolute ratings, which the authors attribute to participants evaluating one graph at a time.","pith_inferences":["An extension the paper leaves implicit: the optimal correlation length should shrink as the trial budget grows, because high-generalizability environments exhaust their learning value quickly, so tasks with more than ten observations should show an inverted-U peak at lower $l$.","A testable consequence of the rating-versus-choice discrepancy: making outcome discriminability explicit, such as by showing graphs side by side or asking how different outcomes could be, should restore an amplitude effect in subjective ratings, suggesting the paper's null $\\sigma$ ratings reflect framing rather than the mechanism.","The same mutual-information machinery should govern exploration beyond design, so a smoothness-manipulated bandit or information-foraging task should reproduce the inverted-U; this is a generality claim the paper does not test.","Because the model assumes the learner's update width scales with the environmental correlation length, an individual-differences variant could estimate each person's internal generalization width and predict that wider internal widths shift preferred environments toward higher $l$."],"forward_implications":["Trials are most informative when knowledge generalizes moderately; too little generalization isolates each trial and too much makes observations redundant.","Subjective value of trying, enjoyment, and choice probability should all peak at intermediate generalizability; the human experiments report this pattern for value, enjoyment, and choices.","Outcome discriminability raises expected uncertainty reduction in simulations, and people prefer more discriminable exploration conditions when they can compare them; this supports using feedback formats that make outcome differences salient.","The value of trying decays across observations, and the decay is faster in high-generalizability environments, meaning exploration should concentrate early trials in smooth, easily learned spaces.","For design practice, prototype sets should differ along one element with moderate generalizability, problem framing should keep the search space neither too narrow nor too broad, and evaluations should use multiple dimensions to keep outcomes distinguishable."],"supporting_citations":[{"why":"Defines epistemic value as the expected information gain of an action, the quantity the B-EUR model applies to design choices.","marker":"Friston et al., 2015"},{"why":"Links epistemic value to the resolution of uncertainty in action selection, grounding the model's use of mutual information.","marker":"Parr and Friston, 2017"},{"why":"Provides the uncertainty-driven action account of design and the open question about how uncertainty perception selects actions that this paper addresses.","marker":"Cash and Kreye, 2017, 2018"},{"why":"Supplies the conceptual idea that proximity among design options determines how much a trial teaches about neighboring candidates, the basis for generalizability.","marker":"Erat and Kavadias, 2008"},{"why":"Shows that human exploration generalizes observations to nearby options in large decision spaces, supporting the Gaussian-weighted belief update.","marker":"Wu et al., 2018"},{"why":"Demonstrates structure-based generalization in bandit search, the empirical precedent for learners updating beliefs across neighboring inputs.","marker":"Schulz et al., 2020"},{"why":"Provides the Gaussian-process-with-RBF-kernel stimulus generation used to create graphs with controlled correlation length and amplitude.","marker":"Angel Deborah et al., 2021"}],"fun_headline_variants":["Design trial value peaks at moderate generalizability","Mid generalization maximizes worth of trying design options","Bayesian model: epistemic value peaks at intermediate generalization","People choose design trials with moderate generalization most"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on taking participants' answer to 'How valuable do you feel it is to know the value of this point?' as a direct measure of expected uncertainty reduction; if that rating tracks visual appeal, perceived difficulty, or curiosity instead, the human confirmation of the inverted-U would not test the model's mechanism.","fun_headline_variants_meta":{"raw":{"variants":["Design trial value peaks at moderate generalizability","Mid generalization maximizes worth of trying design options","Bayesian model: epistemic value peaks at intermediate generalization","People choose design trials with moderate generalization most"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000328,"raw_usage":{"total_tokens":1825,"prompt_tokens":928,"completion_tokens":897,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":840}},"tokens_in":544,"tokens_out":897,"duration_ms":9279,"temperature":1.0,"reasoning_tokens":840,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T04:59:39.962566+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test would be to fit the B-EUR model's predicted $V_{\\mathrm{epi}}(u_t)$ to each participant's trial-by-trial value-of-trying ratings; if the ratings do not track those predictions once the main effects of $l$, $l^2$, and $\\sigma$ are controlled, or if visual features of the graph predict the ratings better, the claim that subjective value is expected uncertainty reduction fails.","supporting_citations":[],"review_version":1}