{"id":"1af20695-c65c-45d2-9ac1-ef201e329298","arxiv_id":"2607.25327","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"BAIT, a belief-aware MPPI controller with a hierarchical particle filter, sustains long-term influence over humans in repeated lane merges while preserving measurably higher trust than influence-only baselines.","lead":"A new robot controller called BAIT decides when to act aggressively and when to act predictably during repeated lane-merging encounters, aiming to keep influencing humans to yield without destroying their trust. The paper reports simulations, a 30-person user study, and real vehicle tests showing it can match the influence of aggressive baselines while preserving more trust.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Belief-estimation channel is unvalidated and possibly inert: BAIT-adapt's reported benefits may be driven by the success-rate heuristic alone, undermining the 'belief-aware' contribution.","rationale":"The reader's weakest assumption focuses on whether the generative human model matches real human belief dynamics. I agree that the lack of external validation is central, but I would sharpen it: the paper's validation is not merely 'same model family'; it is circular because the ground-truth belief in simulation and the pseudo-labels in the user study are both constructed from the exact geometric proxy equation (Eq. 10) that the filter uses as evidence. This circularity means the reported belief-tracking accuracy (Table I) cannot be taken as evidence that the filter tracks actual human beliefs. More importantly, even if the model were faithful, the arbitration rule also includes a success-rate threshold that could alone produce the observed task and trust outcomes. The paper does not ablate the belief threshold, so the 'belief-aware' mechanism may be superfluous. This is a load-bearing concern because the paper's novelty claim is specifically 'Belief-Aware' influence, not just a heuristic mode-switcher. The concrete test is feasible: remove the belief threshold in simulation and measure outcomes; if results do not change, the central contribution reduces to a simple threshold rule. I retain the reader's CONDITIONAL verdict because these are serious but addressable concerns: the empirical comparisons (BAIT-adapt vs BAIT-infl on task performance and trust) remain meaningful, and an ablation could strengthen the paper. I do not see grounds for outright rejection, but the belief-aware mechanism needs independent validation.","tokens_in":12943,"tokens_out":3970,"duration_ms":63329,"concrete_test":"Re-run the Section V simulation with BAIT-adapt modified to ignore \\hat{φ}: switch β=1 iff the 5-episode success rate <60%, else β=0 (no belief threshold). Compare human lane progress and, in a follow-up user study, trust ratings against full BAIT-adapt. If outcomes are statistically indistinguishable, the belief channel is not load-bearing. Additionally, in the user study, collect per-episode self-reported human belief about robot assertiveness and correlate with the filter's \\hat{φ}; low correlation would confirm the belief estimates are unvalidated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is belief-aware arbitration: BAIT-adapt switches β using a 5-episode success-rate threshold OR a belief threshold (Section V-C). The belief estimates are validated circularly: the simulated human's ground-truth belief is generated by the same geometric proxy b_t^y (Eq. 10) used in the robot's filter (Section V-A), and the real-human pseudo-labels are anchored to that same proxy via Eq. (16). Thus the filter's accuracy numbers (Table I) do not demonstrate that \\hat{φ} tracks genuine human internal belief. If the belief channel is inaccurate or redundant, the reported trust/task outcomes may be entirely attributable to the simple success-rate rule, which requires no belief tracking. The paper presents no ablation removing the belief threshold, and the real-vehicle experiment is qualitative only, so it cannot discriminate. This leaves the 'Belief-Aware' claim unsupported as a mechanism, even if the headline empirical trade-off replicates.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BAIT, a controller for repeated human-robot interaction that tracks a two-timescale latent state — a fast strategy variable z_t and a slow belief variable phi_t — with a hierarchical particle filter, and uses a deterministic-surrogate MPPI planner to explicitly trade off long-horizon influence against human trust while enforcing task performance as a CVaR constraint. The authors evaluate BAIT-trust, BAIT-infl, and an adaptive BAIT-adapt variant in a CARLO-based simulated lane-merging task, in a 30-participant in-person user study, and on Polaris GEM vehicles. The central claim is that adaptive arbitration between influence and trust achieves cumulative task performance comparable to influence-only baselines while preserving significantly higher user trust. The paper also claims real-time scalability (~0.11 s/step) by replacing full MOMDP belief-tree evaluation with a posterior-moment surrogate.","tokens_in":13152,"tokens_out":6856,"duration_ms":86752,"significance":"If the belief-estimation channel were independently validated, BAIT would be a meaningful advance: it provides a concrete, real-time architecture for the influence-trust trade-off in repeated HRI, with explicit equations, a user study, and a hardware demonstration. The deterministic-surrogate MPPI planner is a practical contribution over prior MOMDP-based approaches, and the paper makes falsifiable quantitative predictions (e.g., lane-progress differences across modes). However, the current validation of the belief channel is internally circular, and the adaptive controller is never compared against a version that removes the belief-based switching signal. As a result, the distinctive 'belief-aware' mechanism is not yet supported as the cause of the reported trust/task outcomes. These gaps are fixable in revision and do not necessarily invalidate the empirical phenomena, but they do prevent acceptance in the current form.","major_comments":[{"comment":"The estimation robustness results in Table I do not establish that the filter tracks a genuine human belief. In Section V-A, the simulated human's ground-truth belief phi_GT is updated using 'the same kinematic moving-average formulation defined for the robot's geometric proxy (b_t^y in Eq. (10))'. The filter's geometric channel uses the same proxy. Thus the reported phi-BCE and phi-MSE numbers measure how well the filter recovers a smoothed kinematic feature generated from the same equation, not whether that feature corresponds to the human's internal belief. An independent generative model or a held-out behavioral prediction test (e.g., using hat_phi to predict real-human actions out of sample) is needed.","section":"V-A / Eq. (10)"},{"comment":"The in-person tracking results in Table I are also not independent evidence. Because no ground-truth internal state exists for real humans, the authors extract pseudo-labels by optimizing Eq. (16), which includes the term lambda_phi (hat_phi_t - b_t^y)^2, where b_t^y is the same geometric proxy from Eq. (10). Consequently, the pseudo-label evaluation can be satisfied by tracking the robot's own kinematic proxy under another name. This circularity should be acknowledged, and the belief channel should be validated against a measurement that does not share Eq. (10).","section":"V-A / Eq. (16)"},{"comment":"BAIT-adapt's switching rule is 'success rate below 60% over the last 5 episodes OR hat_phi > 0.7'. The paper never ablates the belief condition; it only compares BAIT-adapt against BAIT-trust, BAIT-infl, and Stackelberg baselines. Therefore the reported improvements could be entirely due to the success-rate heuristic, with the belief estimate being inert or redundant. Since the title and central contribution emphasize 'Belief-Aware' behavior, an ablation is load-bearing: compare BAIT-adapt against (i) success-rate-only switching, (ii) belief-only switching, and (iii) a mismatch condition where the belief signal is corrupted. Without this, the mechanism behind the trust/task trade-off is unidentified.","section":"V-C / VI-B"},{"comment":"The abstract claims task performance 'comparable' to influence-only baselines, but the simulation results appear to contradict this. In Section V-D, BAIT-adapt's human lane progress (1878.0) is significantly higher than BAIT-infl (1628.7) and Belief-Entropy Stackelberg (1598.9), both p<0.001. In Section VI-B, however, BAIT-adapt is reported as equivalent to BAIT-infl (p=0.811). If human lane progress is the task-performance metric, the simulation result is inconsistent with the abstract; if a different success metric is intended, it should be explicitly defined and reported for all conditions.","section":"V-D / VI-B"}],"minor_comments":[{"comment":"The transition in Eq. (1) writes z_{t+1} = f_s(s_t, a_t, z_t, phi_t), but Eq. (6) defines z_{t+1} as depending on the predicted phi_{t+1}. Please align the notation to avoid confusion about the conditioning order in the hierarchical filter.","section":"III / IV-A"},{"comment":"The terms 'task performance' and 'human lane progress' are used interchangeably. Please define the primary outcome metric explicitly and report merge success rates alongside lane progress, since the two may diverge.","section":"V-D / VI-B"},{"comment":"The statistical reporting is imprecise: 'across all four pairwise comparisons' should be three for three controllers, and the pairwise p-value comparing BAIT-adapt versus BAIT-trust on the trust item is not reported. Also, 'comfort median 2, discomfort median 5' appears to describe reversed-coded scales; clarify the direction.","section":"VI-B"},{"comment":"The real-time claim '~0.11 s/step' would be more informative with a wall-clock comparison to an online MOMDP solver (e.g., POMCPOW) under the same task, rather than a standalone number.","section":"V-D"},{"comment":"The real-world experiment is qualitative only. Please report the number of trials, the number of participants/drivers, and quantitative outcome measures (e.g., merge success, lane progress, trust ratings) to support the claim of 'consistent' yielding across repeated interactions.","section":"VII"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's own Section V-A explicitly states that the simulated human's ground-truth belief uses the same kinematic moving-average formula as the robot's geometric proxy, and Eq. (16) anchors real-human pseudo-labels to that same proxy. The missing ablation of the belief-based switching rule is a concrete, fixable gap. I recommend requiring the authors to add the ablation and to either independently validate the belief channel or substantially soften the 'belief-aware' causal claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read of BAIT. The paper's core contribution is a controller that combines a hierarchical particle filter over a two-timescale human model (fast strategy, slow belief) with an MPPI planner that trades influence against trust, enforced with a CVaR safety constraint. That integration is new, and the engineering is competent: real-time ~0.11 s/step, consistent results in simulation, a 30-person within-subject user study with counterbalancing, and a real GEM vehicle demo.\n\nWhat's genuinely good: the user study shows BAIT-adapt sustains task performance comparable to influence-only while scoring significantly higher on trust (median 3 vs 2). That is a useful, concrete result for the influence-trust trade-off in repeated HRI. The paper also does a fair job distinguishing itself from Sagheb et al.'s earlier Stackelberg and MOMDP work, and the deterministic surrogate for belief propagation is a sensible scalability fix.\n\nThe soft spots are real. The belief-tracking validation is circular: the simulated human's ground-truth belief is generated with the same kinematic moving-average equation (Eq. 10) the robot uses as its geometric proxy, and the real-human pseudo-labels are anchored to that same proxy via Eq. (16). So Table I doesn't tell us the filter tracks actual human internal belief; it tells us the filter matches its own proxy. The stress-test concern that the belief channel might be inert is fair: BAIT-adapt switches on either a 5-episode success-rate threshold or a belief threshold, and there's no ablation removing the belief threshold. The success-rate heuristic alone might produce the same task outcomes. That said, the trust results come from subjective ratings, which are independent of the filter, so the headline trade-off isn't invalidated. Also minor: the real-vehicle experiment is qualitative only.\n\nProportionally, the core empirical comparisons (task performance and trust) hold up; the belief-tracking accuracy numbers should be discounted, and the 'belief-aware' mechanism is not yet proven. The paper is written honestly and the limitations it does state are consistent with what I found.\n\nWho is this for? Researchers in long-term HRI and autonomous vehicle planning. It deserves a serious referee: the system is coherent, the user study is a real effort, and the influence-trust trade-off is worth publishing even if the mechanism needs work. I'd send it to peer review with major revision requests: re-frame or heavily caveat the belief-tracking validation, add an ablation of the belief threshold, and make the arbitration rule's sensitivity explicit.\n\nI'd probably bring this to a reading group as a case study in validation circularity, and I'd cite it if I were working on influence-aware planning.","headline":"Solid system paper with real user-study results, but the belief-tracking validation is circular and the adaptive arbitration's belief channel needs an ablation to be credible.","tokens_in":13761,"tokens_out":2609,"would_cite":true,"duration_ms":38645,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A robot can shape a person's belief across repeated encounters to keep its long-term influence without sacrificing trust, by adaptively switching between influence and trust modes.","keywords":["long-term human-robot interaction","trust","influence","belief-aware planning","hierarchical particle filter","model predictive path integral","repeated lane merging","autonomous vehicle navigation"],"falsifier":"In a repeated lane-merging study, collect a per-trial measure of the human's belief that the robot will yield (for example, a short post-merge rating) and compare it with the filter's inferred belief. If the controller achieves high task performance while its inferred belief tracks the reported belief no better than a constant prior, the claim that the robot is actually shaping human belief would be falsified. Alternatively, run the adaptive controller with the belief channel artificially frozen while preserving the same observable actions; if trust ratings are unchanged, the belief-shaping me","tokens_in":12741,"feed_emoji":"🤖","tokens_out":6589,"duration_ms":97839,"temperature":0.7,"pith_summary":"This paper tries to establish that a robot facing the same person over many encounters can actively steer that person's evolving perception of the robot, rather than treating each encounter as isolated. The proposed controller, BAIT, tracks a two-timescale model of the human — a fast-changing driving strategy and a slower belief about whether the robot is assertive or defensive — and plans actions that explicitly balance long-term influence against trust, with immediate task performance enforced as a hard safety constraint. In simulation, a 30-person user study, and a real-world vehicle deployment, BAIT is reported to match the task performance of controllers that maintain influence through unpredictability while producing higher user trust, and to run online at about 0.11 seconds per step. If these results hold, the paper opens a practical route to long-term human-robot cooperation in which robots do not have to choose between being effective and being trusted.","feed_headline":"Adaptive robot keeps influence without losing trust","feed_subtitle":"In repeated lane merges, adaptive switching keeps task success high and user trust above pure-influence control.","key_machinery":"The central object is the two-timescale latent hierarchy: a slow long-term belief φ∈[0,1] (assertive-to-defensive perception of the robot) that shapes a fast short-term strategy z (the human's weighting of progress versus safety), together with the conditional coupling that makes the belief observable only through the strategy. This hierarchy is tracked by a two-layered particle filter — long-term particles drift through a mixture Beta kernel and are updated by both action evidence and a geometric yield-belief proxy, while short-term particles mean-revert through an Ornstein-Uhlenbeck process and are reweighted by Boltzmann action likelihoods. For planning, the full mixed-observability belie","core_discovery":"The central claim is that the influence–trust trade-off in repeated human-robot interaction can be managed online by separating the human's latent state into a fast strategy variable and a slow belief variable, inferring both with a hierarchical particle filter, and then planning with a belief-aware sampling-based controller that propagates posterior moments deterministically. The controller arbitrates between an influence cost, which promotes unpredictable entropy-increasing behavior to keep the robot from being exploited, and a trust cost, which promotes transparent predictable legibility, while a conditional value-at-risk constraint blocks any trajectory that risks the immediate task. An","pith_inferences":["Because the paper reports only binary mode switching, a natural testable extension is continuous arbitration over β; if the trust–influence frontier is well behaved, a smoother schedule could dominate the binary hysteresis rule.","The geometric yield-belief proxy is a stand-in for real perception; replacing it with elicited self-reports of expected robot yielding would test whether the filter is tracking beliefs or just kinematics.","A stress test of the model family: run the same controller against simulated humans whose belief-update kernel differs from the assumed Beta/OU structure (for example, a belief that updates directly from robot actions rather than only through the strategy variable); if trust and task performance degrade sharply, the result may owe more to model fit than to the arbitration logic.","The user study limits generalization to a single scenario and 20 episodes per condition; whether the trust benefit persists across heterogeneous road users or longer exposure remains open."],"forward_implications":["Repeated human-robot interaction can be planned in real time: the deterministic surrogate avoids the exponential branching of full belief-tree solvers, with reported control at about 0.11 seconds per step.","If the claim is right, adaptive mode switching lets a robot keep its long-term influence (preventing humans from settling into exploitative behavior) while preserving significantly more user trust than pure influence controllers.","Transparent, trust-maximizing robots are exploitable over repeated interactions: both the stationary baseline and the pure-trust variant saw human lane progress rise over episodes as drivers learned the robot would yield.","The belief-shaping effect transfers from simulation to a 30-person user study and to physical vehicle deployments, so the mechanism is not limited to one environment.","Immediate task performance can be made non-negotiable: the CVaR constraint filters out belief-shaping trajectories that jeopardize the merge or collision avoidance."],"fun_headline_variants":["Belief-aware controller maintains influence while raising trust","Influence preserved, trust boosted: belief-aware robot control","Adaptive robot keeps task success and user trust high","Trust-preserving influence for repeated human-robot interactions","Belief-aware planning improves trust in repeated lane merges"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central result collapses if a real human driver's belief does not evolve through the assumed two-timescale structure — a slow perceptual belief that only acts through a fast strategy — because the filter, the planner's rollout surrogate, and the simulated humans all share that same model family.","fun_headline_variants_meta":{"raw":{"variants":["Belief-aware controller maintains influence while raising trust","Influence preserved, trust boosted: belief-aware robot control","Adaptive robot keeps task success and user trust high","Trust-preserving influence for repeated human-robot interactions","Belief-aware planning improves trust in repeated lane merges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000305,"raw_usage":{"total_tokens":1566,"prompt_tokens":701,"completion_tokens":865,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":802}},"tokens_in":445,"tokens_out":865,"duration_ms":11592,"temperature":1.0,"reasoning_tokens":802,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T03:24:50.544152+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a repeated lane-merging study, collect a per-trial measure of the human's belief that the robot will yield (for example, a short post-merge rating) and compare it with the filter's inferred belief. If the controller achieves high task performance while its inferred belief tracks the reported belief no better than a constant prior, the claim that the robot is actually shaping human belief would be falsified. Alternatively, run the adaptive controller with the belief channel artificially frozen while preserving the same observable actions; if trust ratings are unchanged, the belief-shaping me","supporting_citations":[],"review_version":2}