{"id":"453077bf-d53b-415c-9b15-ec48e474dafb","arxiv_id":"2502.05118","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A pilot study found a non-significant trend (p=0.061) for people to give more positive feedback to a cuter-looking robot, plus a stochastic TAMER variant to cope with biased feedback.","lead":"This paper reports a 14-person pilot study testing whether people give feedback differently to a robot with a cute face versus a plain face. The authors found a trend towards more positive feedback for the cuter robot and propose a modified TAMER learning algorithm to resist that bias.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The feedback-ratio effect may be confounded by presentation order: the paper never states whether cute/control order was counterbalanced, and its own limitations report that the first-seen robot got higher ratings, so the p=0.061 difference could reflect order rather than cuteness.","rationale":"The reader's weakest assumption identified the stochastic TAMER oracle reward as the key issue, but that concern pertains to the algorithm evaluation section, which is secondary to the paper's central empirical claim about cuteness and feedback. The most load-bearing problem is that the central claim's evidence is vulnerable to a presentation-order confound that the authors themselves document but do not control or analyze. This is a higher-stakes threat because it undermines the causal interpretation of the main result, regardless of the algorithm's oracle assumption. I therefore partially agree with the reader: both of us recognize the statistical fragility and overclaim, but I prioritize the order confound over the simulation oracle as the decisive weakness. The reader's verdict of CONDITIONAL remains appropriate—the paper is an honest, transparent pilot, and if the authors can show the effect survives order adjustment (or clearly report counterbalancing), the central claim could stand as a preliminary finding. The recommended conditions should be expanded to include an order-confound analysis and correction of the reported degrees of freedom, not merely better statistical wording and algorithm variance. No finding of misconduct or dishonesty is implied; the issue is an incomplete analysis of an acknowledged confound.","tokens_in":9400,"tokens_out":3359,"duration_ms":35985,"concrete_test":"Obtain the raw per-participant feedback data and check the presentation order of the cute vs control robot for every participant. Fit a mixed-effects model on log((positive+0.5)/(negative+0.5)) with fixed effects for condition and presentation order and a random intercept per participant. If the condition effect becomes non-significant after controlling for order, or if the order × condition interaction is significant, re-analyze only the first-presented robot; if the cute-vs-control difference disappears, the central conclusion must be softened to a hypothesis-generating trend.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that perceived cuteness has a statistical impact on the ratio of positive to negative feedback—rests on a within-subject comparison that is potentially confounded by the order in which robots were presented. Section V explicitly reports order effects on overall ratings: the first-seen robot received a higher overall rating (3.86 vs 3.35), and cuteness ratings also shifted depending on order. Yet the Methods (Section III.C) never state whether the cute/control order was counterbalanced across participants, and no analysis of feedback ratio by order is reported. If the cute robot was systematically seen second (or first), the observed feedback-ratio difference (2.65 vs 1.71, p=0.061) could be an artifact of fatigue, anchoring, or baseline adjustment rather than cuteness per se. This is a direct threat to internal validity, not merely a statistical nuance. Additionally, the reported t(14) for the feedback ratio is inconsistent with n=14 paired data, which should give df=13; this signals a reporting error that could affect the exact p-value. The paper's own language in Section IV.B ('cannot confidently state that H2 is supported') contradicts the conclusion's 'does have a statistical impact,' so the overclaim is compounded by the unresolved order confound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports a within-subject pilot study (n=14) investigating whether a robot's perceived cuteness affects the ratio of positive to negative feedback that users give while critiquing pick-and-place trajectories. The authors find that their baby-schema face manipulation is effective (H1: t(13)=3.19, p<0.01), and they report a higher positive/negative feedback ratio for the cute robot than the control (2.65 vs 1.71) with a one-tailed paired t-test reported as t(14)=1.66, p=0.061. They report that H2 is not supported at the conventional threshold in Section IV.B, but the conclusion states that perceived cuteness \"does have a statistical impact\" on feedback ratio. The paper also proposes a Stochastic TAMER variant that judges the optimality of user feedback using the environmental reward and stochastically swaps positive feedback for negative when a user score falls below a threshold, with a small Wumpus World evaluation across three trials.","tokens_in":9651,"tokens_out":2937,"duration_ms":31447,"significance":"If the main empirical claim were established, the paper would address a relevant and understudied question in HRI and interactive robot learning: whether aesthetic cuteness systematically skews the reward signal that humans provide to learning agents. The manuscript is honest in reporting the primary result as p=0.061, and the cuteness manipulation is well supported by the data. The use of prerecorded trajectories controls for variability in robot behavior, and the authors explicitly enumerate confounds in the limitations section. However, the central new empirical claim rests on a single borderline p-value with an inconsistent degrees-of-freedom report and a plausible order confound that is acknowledged but not resolved. The Stochastic TAMER contribution is preliminary and rests on a reward-availability assumption that conflicts with the original TAMER setup. The paper is suitable as a pilot-study report if the claims are tempered and the load-bearing issues are addressed, but not as definitive evidence of a cuteness effect on feedback.","major_comments":[{"comment":"The order in which participants saw the two robots is a direct threat to the internal validity of the feedback-ratio comparison, yet the Methods do not state whether the cute/control order was counterbalanced or randomized across participants. Section V reports that the first-seen robot received a higher overall rating (3.86 vs 3.35) and that cuteness ratings shifted systematically with order. If the cute robot was not equally often seen first and second, the observed difference in positive/negative feedback ratio (2.65 vs 1.71) could be an artifact of order, anchoring, or fatigue rather than cuteness. Please report the order assignment, provide the feedback-ratio comparison separately by order, and either control for order in the analysis or clearly restrict the conclusion to an exploratory trend.","section":"§III.C and §V"},{"comment":"The reported degrees of freedom for the feedback-ratio test are inconsistent with the sample size. The paper states t(14)=1.66 for a paired test with n=14 participants; the correct degrees of freedom are 13, not 14. This discrepancy could affect the exact p-value and should be corrected. In addition, the paper should report the effect size and a confidence interval for the mean difference, since the p-value alone does not convey the magnitude or precision of the claimed shift.","section":"§IV.B"},{"comment":"The conclusion overstates the evidence for H2. Section IV.B explicitly says \"we cannot confidently state that our second hypothesis (H2) is supported,\" but Section VII states that \"perceived cuteness does have a statistical impact on a human's ratio of positive to negative feedback.\" These statements are contradictory. The conclusion must be aligned with the reported p=0.061 and the acknowledged confounds, or new supporting evidence must be provided.","section":"§VII vs §IV.B"},{"comment":"The Stochastic TAMER user-score mechanism assumes that the algorithm can access the environmental reward for each action when determining whether user feedback was appropriate. This assumption conflicts with the core premise of TAMER described in Section II.A, which is to learn without an environmental reward function. The simulation evaluation is also very limited: only three trials are reported, with no statistical comparison across runs or across simulated users, and the statement that the optimal-feedback condition has \"standard deviation of zero\" is not a substitute for an analysis of variability. Please either reframe the algorithm as an illustrative proof-of-concept that relaxes TAMER's assumptions, or provide a more thorough evaluation and discuss the reward-availability limitation.","section":"§III.D and §IV.C"}],"minor_comments":[{"comment":"The text says the two faces are \"as seen in Figure 2,\" but Figure 2 shows the pick-and-place task; the faces appear in Figure 3. Please correct the cross-reference.","section":"§III.C.1"},{"comment":"The design is described as giving each participant a random sample of four prerecorded trajectories, but the earlier text says half of the 12 trajectories are successes and half are failures. It is unclear whether each participant always received exactly two successes and two failures, or whether the random sample could produce an uneven split. Please clarify.","section":"§III.C.2"},{"comment":"The Wumpus World experiments lack essential details: the state/action space, the user feedback model that generates the specified positive-feedback rates, the number of episodes, and the aggregation across the three trials. Adding these details would make the simulation reproducible.","section":"§IV.C"},{"comment":"The figure captions for Figure 2 and Figure 3 are interchanged in content: Figure 2 is labeled as an illustration of the task but is referenced in the text for the face display, while Figure 3 shows the faces. Please ensure each figure number matches its caption.","section":"§III.C.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a clearly written pilot study with an honest limitations section, but its central claim currently rests on a borderline p-value, an unresolved order confound, and an overstatement in the conclusion. The Stochastic TAMER contribution is too preliminary for the role it plays in the paper's claims. I believe these issues are fixable within the scope of a revision: report the order analysis or counterbalancing, correct the degrees of freedom, align the conclusion with the evidence, and reframe the algorithm as a proof-of-concept. I would not support acceptance without those changes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a small HRI pilot (n=14) that asks a genuinely underexplored question — whether a robot's perceived cuteness shifts the ratio of positive to negative feedback in an interactive learning setting. The good news: the cuteness manipulation clearly works (t(13)=3.19, p<0.01), the study is IRB-approved, and the authors are unusually transparent about their limitations. The bad news: the conclusion overstates what the data show, and there is a real order-confound threat that the paper itself identifies but never resolves.\n\nWhat's actually new: the direct test of cuteness on feedback ratio appears to be missing from the prior work they cite, so this is a legitimate gap. The stochastic TAMER variant is a modest algorithmic tweak, not a major contribution. Its user-score mechanism relies on knowing the environmental reward for each action — that contradicts TAMER's core premise of learning without a reward function. The evaluation is also thin: three trials, no error bars, no parameter values, and the paper admits it was done under time pressure.\n\nSoft spots, in order of severity:\n\n1. Order effects. The Methods never state whether cute/control order was counterbalanced. The Limitations section reports that the first-seen robot got higher overall ratings, and that cuteness ratings shifted depending on order. If order wasn't balanced, the p=0.061 feedback-ratio difference could be partly an artifact of fatigue or anchoring. They should have reported the feedback ratio split by order.\n\n2. Statistical reporting. They report t(14) for a within-subject n=14; paired df should be 13. That error changes the exact p-value and needs fixing.\n\n3. Overclaim. Section IV.B says \"cannot confidently state that H2 is supported,\" but the Conclusion says \"perceived cuteness does have a statistical impact.\" That is internally inconsistent.\n\n4. Stochastic TAMER's oracle assumption is a design flaw, though they are upfront that the work is preliminary.\n\nWho this is for: researchers in HRI and interactive robot learning who care about how aesthetics bias human teaching. The paper is a proof-of-concept, honestly labeled, with a robust manipulation check. With the overclaim corrected, the statistical reporting aligned, and a discussion of the order confound, it would be a reasonable pilot-study paper. As it stands, I'd send it to peer review with a major-revision recommendation — the question is worth asking and the flaws are fixable. Not a desk reject.","headline":"Candid pilot study with a solid cuteness manipulation, but the central feedback-ratio claim is overstated and possibly order-confounded.","tokens_in":10176,"tokens_out":2328,"would_cite":false,"duration_ms":23314,"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":"The paper claims that a robot's perceived cuteness shifts human feedback toward a higher ratio of positive to negative critiques, and that a stochastic version of TAMER can compensate for this bias.","keywords":["cuteness","baby schema","positive feedback bias","TAMER","learning from demonstration","human-robot interaction","reinforcement learning from human feedback","user study design"],"falsifier":"Run the same within-subject procedure with a preregistered sample of at least 60 participants: if the cute robot's positive-to-negative feedback ratio is not reliably higher than the control's at $p < 0.05$, the claimed shift fails. Then run Stochastic TAMER in Wumpus World without providing the per-action environmental reward; if the user score cannot be computed or performance does not beat plain TAMER under 66 percent positive feedback, the proposed mitigation is not as described.","tokens_in":9167,"feed_emoji":"🤖","tokens_out":7434,"duration_ms":58486,"temperature":0.7,"pith_summary":"The paper tries to establish that the perceived cuteness of a robot changes the mix of positive and negative feedback people give while training it, with cuter robots receiving proportionally more praise. This matters because reinforcement-learning-from-feedback systems such as TAMER already suffer from a human tendency to give too much positive feedback, and a cute exterior could worsen that bias and keep the robot stuck in suboptimal behavior. The authors run a within-subject pilot (14 participants) in which people critique recorded pick-and-place trajectories of a robot with either a baby-schema 'cute' face or a control face. They find the cute robot is rated cuter and receives a higher mean positive-to-negative feedback ratio (2.65 vs 1.71), though the effect only reaches $p = 0.061$ on a one-tailed paired t-test, short of the conventional threshold. They also propose a stochastic version of TAMER that, when a user score judges feedback to be suboptimal, gradually swaps praise for criticism to break out of positive reward circuits.","feed_headline":"Cuter robots get more praise from human trainers","feed_subtitle":"A 14-person study finds the cute robot's praise-to-criticism ratio was 2.65, vs 1.71 for the plain-faced control.","key_machinery":"The study's central object is the baby schema face manipulation: a 'cute' face with large eyes low on the face versus a control face with smaller eyes higher up, built on the baby schema principle from developmental psychology. The outcome metric is the ratio of positive to negative keypress critiques given during playback of four prerecorded pick-and-place trajectories per robot. For the algorithmic half, the central mechanism is a user score inside Stochastic TAMER that judges whether each piece of user feedback was appropriate by comparing it against the environmental reward for that action; when the score falls below a threshold, the algorithm swaps positive feedback for negative with a probability that increases as the user score stays low.","core_discovery":"The central claim is that a robot's perceived cuteness is a cause of the well-documented positive feedback bias in interactive robot learning, not just a background property of the platform. The authors establish the perception manipulation first: the baby-schema face (larger eyes lower on the face) is rated significantly cuter than the control (mean 4.0 vs 2.79, $p < 0.01$). They then report that participants gave the cute robot a higher ratio of positive to negative feedback (2.65 ± 2.26 vs 1.71 ± 1.23), with a one-tailed paired t-test at $p = 0.061$, which they interpret as a trend warranting further study. The paper further claims that this skewed feedback degrades learning in TAMER-style algorithms and that a proposed Stochastic TAMER, which stochastically converts positive feedback into negative when a user score indicates poor feedback quality, can mitigate the problem while retaining identical performance when feedback is optimal.","pith_inferences":["The cuteness effect likely generalizes beyond faces to other baby-schema cues such as body proportions, voice, and movement, since the mechanism is affective response, not visual pixels.","The Stochastic TAMER design as described has an oracle problem, since it requires the very environmental reward that TAMER is meant to replace, so a practical version would need to estimate feedback appropriateness from trajectory outcomes or user consistency.","A between-subjects design with more participants and a washout period could separate the order and anchoring effects the authors observed from the true cuteness effect.","If the effect holds, deploying cuter robots in real feedback-driven learning systems will systematically bias collected data, which RLHF-style pipelines in consumer robots should account for."],"forward_implications":["Aesthetic design choices are not neutral for robot learning: making a robot cuter can inflate the praise signal a learning algorithm receives.","Systems that learn from human feedback should estimate and compensate for the user's positive feedback bias rather than treating every critique as ground truth.","The Stochastic TAMER modification, if validated, would let an agent break out of positive reward circuits without needing a change in the human's behavior.","Researchers designing user studies for interactive robot learning should control for robot appearance as a variable, not just as a confound to mention.","Pilot results at $p = 0.061$ justify larger preregistered studies rather than dismissing the effect."],"supporting_citations":[{"why":"Supplies the baby-schema face design used to make the cute robot's larger eyes and forehead versus the control.","marker":"[3]"},{"why":"Documents the positive feedback bias in human reward and the positive reward circuit failure that the paper builds on.","marker":"[5]"},{"why":"Shows the positive bias increases over time as the robot improves, motivating the feedback-ratio analysis.","marker":"[6]"},{"why":"Defines the original TAMER algorithm that the user study and Stochastic TAMER extend.","marker":"[21]"},{"why":"Shows robot appearance changes how humans phrase feedback, supporting the link between aesthetics and critique type.","marker":"[31]"},{"why":"Provides the TAMER implementation basis for the proposed Stochastic TAMER modification in Methods Part D.","marker":"[34]"}],"fun_headline_variants":["Cuter robots get more praise, skewing robot training","Robot cuteness alters human feedback in RL training","Baby-face robots trigger more positive feedback from users","Cute aesthetics inflate praise in robot learning studies","Perceived cuteness shifts human feedback bias in robotics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The Stochastic TAMER user score assumes the algorithm has access to the true environmental reward for each action so it can tell whether a human's praise was deserved, which is exactly the ground-truth signal that TAMER-style learning is designed to do without.","fun_headline_variants_meta":{"raw":{"variants":["Cuter robots get more praise, skewing robot training","Robot cuteness alters human feedback in RL training","Baby-face robots trigger more positive feedback from users","Cute aesthetics inflate praise in robot learning studies","Perceived cuteness shifts human feedback bias in robotics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1338,"prompt_tokens":904,"completion_tokens":434,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":359}},"tokens_in":520,"tokens_out":434,"duration_ms":4882,"temperature":1.0,"reasoning_tokens":359,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T20:10:13.537707+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same within-subject procedure with a preregistered sample of at least 60 participants: if the cute robot's positive-to-negative feedback ratio is not reliably higher than the control's at $p < 0.05$, the claimed shift fails. Then run Stochastic TAMER in Wumpus World without providing the per-action environmental reward; if the user score cannot be computed or performance does not beat plain TAMER under 66 percent positive feedback, the proposed mitigation is not as described.","supporting_citations":[{"cited_title":"Research on the influence of the baby schema ef- fect on the cuteness and trustworthiness of social robot faces,","cited_arxiv_id":null,"evidence_quote":"Supplies the baby-schema face design used to make the cute robot's larger eyes and forehead versus the control."},{"cited_title":"Reinforcement learning from human reward: Discounting in episodic tasks,","cited_arxiv_id":null,"evidence_quote":"Documents the positive feedback bias in human reward and the positive reward circuit failure that the paper builds on."},{"cited_title":"How humans teach agents: A new experimental perspective,","cited_arxiv_id":null,"evidence_quote":"Shows the positive bias increases over time as the robot improves, motivating the feedback-ratio analysis."},{"cited_title":"TAMER: Training an Agent Manually via Evaluative Reinforcement,","cited_arxiv_id":null,"evidence_quote":"Defines the original TAMER algorithm that the user study and Stochastic TAMER extend."},{"cited_title":"Austermann, S","cited_arxiv_id":null,"evidence_quote":"Shows robot appearance changes how humans phrase feedback, supporting the link between aesthetics and critique type."},{"cited_title":"Interactively shaping agents via human reinforcement: the TAMER framework,","cited_arxiv_id":null,"evidence_quote":"Provides the TAMER implementation basis for the proposed Stochastic TAMER modification in Methods Part D."}],"review_version":1}