{"id":"5ea0dc18-0a22-468f-81e1-9bfd2b4f330f","arxiv_id":"2506.05648","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A reconfigurable, fluidic-logic haptic display paired with a preference-weighted information-theoretic selector reduced user error in two robot guidance tasks.","lead":"This paper combines modular pneumatic haptic displays with an information-theoretic algorithm that selects which signal types (pressure, frequency, or contact area) to render for a given user and task. In a 13-person user study, the algorithm-recommended configurations produced lower task error and higher subjective ratings than the alternatives.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The information-theoretic objective is internally inconsistent: the second term in Eq. (2) is zero for the deterministic signal mappings used in the user study, so the algorithm does not maximize the stated mutual information.","rationale":"I read the paper as making an empirical claim that its recommended configuration improves task performance, with the recommendation driven by an information-theoretic objective. The empirical result (Rank 1 better error) is plausible and the within-subjects design with counterbalancing supports it. However, the load-bearing link between the algorithm and the claimed information maximization is broken: the equations do not compute mutual information as defined. In the user study, the signal is deterministically generated from the task state, so the conditional entropy in Eq. (5) is zero; the human model in Eq. (6) is a separate preference term, not part of I(θ;s). This is more fundamental than the unsupported proportionality in Eq. (4) — indeed Eq. (4) is only problematic because H(s|θ)=0 leaves H(s|S,ρ) as the entire objective, and that entropy is not accurately characterized by a logarithm of cardinalities. The paper could be repaired by reframing the algorithm as a heuristic score combining signal resolution with user preference; the empirical study would then support that heuristic, but not 'information-theoretic interfaces' as claimed. Since the flaw is conceptual but addressable, and the hardware contribution is substantial, conditional acceptance (with required revision) remains the right call — same as the reader's verdict. I partially agree with the reader's weakest assumption: Eq. (4) is indeed unsupported, but the deeper issue is that the mutual-information decomposition itself is misapplied, making the second term vanish.","tokens_in":63,"tokens_out":8503,"duration_ms":210348,"concrete_test":"Recompute the ranking for all 13 participants using the true mutual information of the actual deterministic mappings: for each configuration, set H(s|θ,S,ρ)=0 and compute I(θ;s) directly from the task prior ρ and the known level-quantization function used in the study (e.g., 4 pressure levels for x-axis). If the resulting rankings differ from the paper's Rank 1/2/3 assignments (because Eq. (6) is dropped), then the algorithm does not maximize the stated objective, confirming the concern. Additionally, independently derive H(s|S,ρ) from Eq. (3) for the uniform 4x4 grid prior and compare to log(|s|·|S|) in Eq. (4); if they disagree, the proportionality is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section IV, the paper defines S* = argmax I(θ; s | S, ρ) and decomposes I = H(s|S,ρ) − H(s|θ,S,ρ) (Eq. 2). In the implemented system (Section V), the displayed signal s is a deterministic function of the hidden state θ: e.g., in Search, the target x-coordinate is mapped to one of four pressure levels; in Assembly, the ingredient index is mapped to a fixed pressure/area/frequency level. Therefore p(s|θ,S,ρ) is a delta distribution and H(s|θ,S,ρ) = 0 for every configuration tested. Equation (5) and the maximum-entropy human model in Eq. (6) instead describe the signal the user expects for θ, which is not the channel likelihood needed for I(θ; s). Consequently, the user-specific term in Eq. (6) is not part of the stated mutual information; it enters as an ad hoc additive penalty. Equation (4) is a symptom: since H(s|θ)=0, the objective reduces to H(s|S,ρ), which Eq. (4) approximates by signal-set cardinality without justification from Eq. (3). The claimed 'maximization of information transfer between the intended message and the user's interpretation' is therefore not what the algorithm computes. The empirical finding that Rank 1 configurations yield lower error is not undermined, but the paper's theoretical rationale for the ranking is.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a modular soft pneumatic haptic display whose signal types (pressure, frequency, area) can be reconfigured through fluidic logic circuits without software changes, together with an information-theoretic algorithm that ranks interface configurations by estimated mutual information between task state and rendered signal. The authors validate the approach in a within-subjects user study (n=13) across two tasks (Search, Assembly), reporting that configurations ranked first by the algorithm yield significantly lower task error and higher subjective ratings than lower-ranked configurations. The paper also analyzes hardware and software contributions, showing different users benefit from different physical configurations and that the recommended configuration is not simply the most preferred one.","tokens_in":19135,"tokens_out":4335,"duration_ms":44453,"significance":"The hardware contribution is substantial and reproducible: detailed fabrication instructions, valve characterization, and the use of the Soft Compiler give the fluidic-logic layer a concrete engineering value. The user study is carefully designed with counterbalancing, objective measures from a robot arm, and Likert-scale questionnaires, and the empirical finding that the algorithm's top-ranked configuration produces lower error and is preferred is a useful result for personalized haptics. However, the paper's central theoretical claim that it 'maximizes information transfer between the intended message and the user's interpretation' is not supported by the derivation in Section IV; the ranking is better described as a preference-weighted signal-cardinality heuristic. Because this theoretical rationale is load-bearing for the contributions, the manuscript needs a major revision rather than minor edits.","major_comments":[{"comment":"The objective actually computed is not the stated mutual information. In the implemented system (Section V), the displayed signal s is a deterministic function of the hidden state θ (for example, in Search the target x-coordinate is mapped to one of four pressure levels, and in Assembly the ingredient index is mapped to a fixed pressure/area/frequency level). Therefore p(s|θ,S,ρ) is a delta distribution and H(s|θ,S,ρ)=0 for every configuration tested. The maximum-entropy model in Eq. (6) describes the signal the user expects for θ, which is a decoding model, not the channel likelihood required in Eq. (5). Substituting Eq. (6) into Eq. (5) makes the user-specific term an ad hoc penalty rather than a component of I(θ;s|S,ρ). The paper should either redefine Eqs. (1)-(2) as a design score, or model a genuinely stochastic channel (for example, noisy encoding plus a decoding model) so that H(s|θ,S,ρ)>0.","section":"Section IV, Eqs. (2)-(6)"},{"comment":"The asserted proportionality H(s|S,ρ) ∝ log(|s| × |S|) does not follow from Eq. (3). The entropy of the rendered signal depends on how the task prior ρ induces a distribution over s through the θ-to-s mapping, not merely on the sizes of the signal sets. For example, a configuration with many levels that are never used under the given task prior has the same H(s|S,ρ) as one with fewer effective levels. Because Eq. (4) is used to compute the 'task' term in the ranking, the rank order is not the information-theoretic optimum claimed in Eq. (1). The authors should derive H(s|S,ρ) from ρ and the mapping, or explicitly present Eq. (4) as a simplifying heuristic rather than as an information-theoretic result.","section":"Section IV, Eq. (4)"}],"minor_comments":[{"comment":"The phrase 'along a α-1 scale' appears to be a typo for '0-1 scale'.","section":"Section IV, Implementation"},{"comment":"The normalization claim 'range [α, 1]' is incorrect if P_i are in [0,1]; the diagonal entries (P_i+α)/(1+α) range from α/(1+α) to 1, not from α to 1. Please correct or clarify the intended scaling.","section":"Section IV, Eq. (7)"},{"comment":"The Assembly post-hoc p-value for Rank 1 versus Rank 2 is reported as p = .10 in the text but p = .197 in the Figure 5 caption; please reconcile the two values.","section":"Section V, Results"},{"comment":"The text states that all five scales were reliable with Cronbach's α > 0.7, but Table I lists Adapt α = .69, which is below 0.7; please adjust either the claim or the table.","section":"Section V, Results and Table I"}],"recommendation":"major_revision","confidential_remarks":"The paper fits cs.HC well, and the empirical study is solid. The main risk is that the information-theoretic framing is overclaimed: the ranking derivation in Section IV is not a valid mutual-information computation, so the contributions should be reworded or the math repaired. Given that this section is central to the claims, I recommend major revision rather than minor revision. If the authors reframe the ranking as a preference-weighted heuristic validated empirically, or derive a correct stochastic-channel formulation, the manuscript would likely be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the hardware and user study are genuinely good, but the information-theoretic framing doesn't hold up to the paper's own implementation. The ranking algorithm appears to work empirically, but the stated math is not what the algorithm computes.\n\nWhat's actually new: applying Preston-style soft valves and Soft Compiler circuits to reconfigurable haptic displays is a real new application. The modular multi-DoF pneumatic displays are well engineered, with solid fabrication details and characterization data (snap-through pressures, flow rates, oscillator frequencies). The two-task within-subjects user study (n=13) is competently designed — counterbalanced, participants blind to rank, objective error plus Likert measures — and the repeated-measures ANOVAs support the claim that the Rank 1 configurations reduce error and are preferred.\n\nThe soft spot is the theory. Eq. (4) is asserted, not derived, and assumes a uniform signal distribution while ignoring the actual task prior and the theta-to-signal mapping. More seriously, in the implemented system the displayed signal is a deterministic function of theta, so H(s|theta) = 0 in Eq. (2) and the second term vanishes. The human model in Eq. (6) is actually a model of the user's expected signal for theta, not the channel likelihood p(s|theta); it enters as an ad hoc preference penalty, not as part of the stated mutual information. So in practice the ranking reduces to level-counting plus a preference-weighted distance. That is not 'maximizing information transfer between the intended message and the user's interpretation.'\n\nThe empirical finding is not undermined: the rankings were produced before outcome data, and the effect is real. The heuristic works; the theory just doesn't say why. The paper would be stronger if the authors either reframed the selection as a task-aware preference heuristic or properly defined a channel model with perceptual noise.\n\nThis is for haptics and HRI researchers, especially those working on soft wearable displays. It deserves a serious referee, but a major revision of Section IV is required. No code or data are provided, which is a moderate reproducibility concern for the algorithm.\n\nRead it for the hardware, cite it, and push back on the theory.","headline":"Solid modular haptic hardware and a clean user study wrapped in an information-theoretic framing that does not match what the algorithm actually computes.","tokens_in":19580,"tokens_out":5038,"would_cite":true,"duration_ms":49623,"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":"This paper claims that personalizing a reconfigurable soft haptic display with an information-theoretic ranking of signal configurations reduces user error in robot-guidance tasks beyond any fixed, non-personalized interface.","keywords":["haptic interfaces","personalization","information theory","fluidic logic","reconfigurable hardware","tactile devices","human-robot interaction","soft robotics"],"falsifier":"Run a ranking comparison between two configurations that have the same number of signal levels but different mappings from task states to signals, under a task prior concentrated on a few states: compute the algorithm's approximated score from Equation (4) and the full mutual information from the actual $\\theta$-to-$s$ mapping. If user task error follows the full-mutual-information ranking rather than the approximated one, the paper's stated objective is falsified; if it follows the approximation, the assumption holds. Alternatively, repeat the user study with a strongly non-uniform task prior and check whether the Rank 1 configurations (and their performance ordering) shift as full mutual information would predict.","tokens_in":18574,"feed_emoji":"🖐️","tokens_out":10139,"duration_ms":101434,"temperature":0.7,"pith_summary":"This paper tries to establish that haptic feedback can be personalized in both hardware and signal choice, and that doing so measurably improves how well people perform guidance tasks. The authors build a soft, finger-worn pneumatic display whose tactile cues — pressure, frequency, or contact area — are changed by re-plumbing fluidic-logic circuits instead of rewriting software, and an algorithm that ranks hardware configurations by how much task-relevant information they would convey to a given user. In two user studies with 13 participants, the configuration ranked first by the algorithm produced significantly lower task error and higher subjective ratings of trust, understanding, and preference than the alternative configurations. The paper concludes that both physical modularity and algorithmic selection are necessary: different users genuinely perform best with different displays, and the best display is not simply the one users say they like. This matters because most haptic interfaces are fixed one-size-fits-all systems, and a reconfigurable display plus a short preference probe could adapt feedback to each user and task.","feed_headline":"Personalized haptic display cuts user error in robot guidance","feed_subtitle":"A modular pneumatic display that swaps pressure, frequency, and area cues beats fixed setups in two user studies.","key_machinery":"The load-bearing mechanism has two coupled parts. The hardware is a modular pneumatic display built from soft bistable valves adapted from a ring-oscillator design, arranged in a code-logic stage that acts as a demultiplexer (three pneumatic inputs select one of eight outputs) and an output-logic stage that renders pressure by direct inflation, frequency through a soft oscillator, and contact area through cascaded inflation of chambers; users reconfigure it by unplugging and re-plugging Luer-lock tubes, with no software changes. The software is an information-theoretic selector, $S^\\star = \\arg\\max_{S \\in \\mathcal{S}} I(\\theta; s \\mid S, \\rho)$, which decomposes into a signal-entropy term $H(s \\mid S, \\rho)$ that rewards displays with many distinguishable signal levels and a conditional-entropy term $H(s \\mid \\theta, S, \\rho)$ that penalizes configurations the user struggles to decode. The human model instantiated for the second term is a maximum-entropy distribution $p(s \\mid \\theta, S, \\rho) \\propto \\exp(-\\beta \\|h(\\theta) - s\\|_W^2)$, where $W$ encodes the user's rated preference for each signal axis and $\\beta$ is an overall sensitivity constant; this is what lets the algorithm balance expressiveness against interpretability for a specific person.","core_discovery":"The central discovery is that a haptic interface's effectiveness can be improved by jointly reconfiguring its physical signal rendering and algorithmically choosing which configuration to deploy, modeled as maximizing the mutual information $I(\\theta; s \\mid S, \\rho)$ between the hidden task variable $\\theta$ and the displayed signal $s$. Concretely, the paper reports that in both a $4 \\times 4$ search task and a seven-ingredient assembly task, the configuration selected by the algorithm (Rank 1) led to significantly lower user error than the worst-ranked configuration, without increasing completion time, and that participants rated Rank 1 higher on trust, adaptation, understanding, distinguishability, and preference. The results also show that the algorithm's top pick is not the same as the user's stated preference: the Rank 1 distribution across pressure-area, pressure-frequency, and area-frequency pairings differs from the preference distribution, because the method balances preference against the display's information-carrying capacity for the task. The paper takes this as evidence that both layers of modularity — physical reconfiguration and algorithmic selection — contribute to the performance gain.","pith_inferences":["The paper's Equation (4) asserts $H(s \\mid S, \\rho) \\propto \\log(|s| \\times |S|)$ 'assuming $\\rho$ is independent of interface design,' which drops the mapping from task states to signals; if that mapping matters (as task-specific effects in the study suggest), the objective actually optimized may not be the mutual information the paper states, and the empirical gains could be carried by the human","Because the human model only requires a saliency weight per signal axis, the same ranking procedure should transfer to other modalities such as vibration or thermal cues; a natural test is whether the accuracy gains replicate on non-pneumatic displays with the same preference probe.","The authors' observation that users form internal models of the signals suggests a closed-loop variant in which errors and reaction times refine $W$ and $\\beta$ online and re-rank configurations during a session, going beyond the static personalization demonstrated here.","The study caps signal levels at four for pressure, three for area, and two for frequency; whether the mutual-information ranking keeps predicting performance as the number of levels per modality grows is an open empirical question that would strengthen or bound the framework."],"forward_implications":["If the paper is correct, deploying the algorithm's top-ranked haptic configuration should reduce user error in guidance tasks without slowing users down, as observed in both the search and assembly studies.","Different users require different displays: four of thirteen participants performed best on a configuration other than the most common one, so a single fixed haptic device cannot serve a diverse population.","User preference is informative but insufficient; since the algorithm's top pick differs from stated preference and still outperforms, selection must trade off preference against the display's information capacity for the task.","The same pneumatic code and logic hardware renders pressure, frequency, and area signals with only mechanical reconfiguration, which lowers the practical barrier to field customization of haptic devices.","Because the human model is static and gathered once, the observed gains come without any online adaptation; the authors note that updating the model with user errors and reaction times could re-rank configurations over time."],"supporting_citations":[{"why":"Supplies the soft bistable valve design that the fluidic-logic control stages adapt for pneumatic logic.","marker":"[16]"},{"why":"Provides the original soft bistable valve architecture that the paper's valves are based on.","marker":"[14]"},{"why":"Shows how soft valves compose into digital logic gates, the basis of the code-logic demultiplexer.","marker":"[15]"},{"why":"Provides the Soft Compiler tool the paper uses to generate the pneumatic demultiplexer circuits from a truth table.","marker":"[17]"},{"why":"Frames haptic device design as an information-transmission problem and motivates the mutual-information objective.","marker":"[6]"},{"why":"Establishes information transfer rates in human communication that motivate the channel model.","marker":"[18]"},{"why":"Introduces the prior approach of selecting interfaces to maximize information transfer, which this paper's personalization algorithm extends to modular hardware.","marker":"[33]"}],"fun_headline_variants":["Reconfigurable haptics cut errors via info-theoretic tuning","Algorithm picks best haptic mode, slashing user error","Modular haptics: swap cues to boost message transfer","Personalized haptic signals minimize user mistakes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ranking algorithm depends on the assumption that the task prior is independent of interface design, which lets it replace the signal-entropy term with $\\log(|s| \\times |S|)$; if the mapping from task states to the signals a configuration can render actually matters, then the mutual information being optimized is not the one that drives task performance.","fun_headline_variants_meta":{"raw":{"variants":["Reconfigurable haptics cut errors via info-theoretic tuning","Algorithm picks best haptic mode, slashing user error","Modular haptics: swap cues to boost message transfer","Personalized haptic signals minimize user mistakes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00043,"raw_usage":{"total_tokens":2227,"prompt_tokens":1005,"completion_tokens":1222,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":1167}},"tokens_in":621,"tokens_out":1222,"duration_ms":11541,"temperature":1.0,"reasoning_tokens":1167,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:13:26.350188+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a ranking comparison between two configurations that have the same number of signal levels but different mappings from task states to signals, under a task prior concentrated on a few states: compute the algorithm's approximated score from Equation (4) and the full mutual information from the actual $\\theta$-to-$s$ mapping. If user task error follows the full-mutual-information ranking rather than the approximated one, the paper's stated objective is falsified; if it follows the approximation, the assumption holds. Alternatively, repeat the user study with a strongly non-uniform task prior and check whether the Rank 1 configurations (and their performance ordering) shift as full mutual information would predict.","supporting_citations":[{"cited_title":"A soft ring oscillator,","cited_arxiv_id":null,"evidence_quote":"Supplies the soft bistable valve design that the fluidic-logic control stages adapt for pneumatic logic."},{"cited_title":"A soft, bistable valve for autonomous control of soft actuators,","cited_arxiv_id":null,"evidence_quote":"Provides the original soft bistable valve architecture that the paper's valves are based on."},{"cited_title":"Digital logic for soft devices,","cited_arxiv_id":null,"evidence_quote":"Shows how soft valves compose into digital logic gates, the basis of the code-logic demultiplexer."},{"cited_title":"The soft compiler: A web-based tool for the design of modular pneumatic circuits for soft robots,","cited_arxiv_id":null,"evidence_quote":"Provides the Soft Compiler tool the paper uses to generate the pneumatic demultiplexer circuits from a truth table."},{"cited_title":"Methodology for maximizing information transmission of haptic devices: A survey,","cited_arxiv_id":null,"evidence_quote":"Frames haptic device design as an information-transmission problem and motivates the mutual-information objective."},{"cited_title":"Note on information transfer rates in human communication,","cited_arxiv_id":null,"evidence_quote":"Establishes information transfer rates in human communication that motivate the channel model."},{"cited_title":"Limit: Learning interfaces to maximize information transfer,","cited_arxiv_id":null,"evidence_quote":"Introduces the prior approach of selecting interfaces to maximize information transfer, which this paper's personalization algorithm extends to modular hardware."}],"review_version":1}