{"id":"a60c90c9-db45-43b2-a856-7fec91bdad8e","arxiv_id":"2507.07692","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A leader-follower game framing with two neural networks predicts next-step haptic signals, reporting 80-95% human-side and 70-90% robot-side accuracy, with a Taylor loss bound presented as a robustness guarantee.","lead":"A short conference paper proposes a game-theoretic prediction scheme for tactile internet, where a leader-follower setup has one network predict the robot's next signal on the human side and another network predict the human's next signal on the robot side. The paper reports 70% to 95% prediction accuracy on a public haptics dataset and derives a Taylor-expansion loss bound, but does not compare with existing predictors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 21 KNN mutual-information estimator is piecewise constant, so the leader 'maximize I' step has no usable gradient and Algorithm 1 cannot be run as written.","rationale":"The reader's weakest_assumption identifies exactly this concern: the KNN mutual-information estimator in Eq. 21 is not shown to be differentiable or optimizable, so the leader's 'maximize I' step cannot be executed from the text. My reading confirms this and goes further: the estimator is piecewise constant, so the obstacle is not merely missing gradient formulas but the structure of the estimator itself. I considered other weaknesses, including the Taylor bound in Eq. 29 bounding training-loss change rather than prediction error from packet loss, the undefined accuracy metric, and the placeholder code link. These are serious, but they are secondary: even if Eq. 29 were corrected and baselines added, the central algorithm would still need a differentiable or otherwise well-defined leader objective. The paper provides none. Therefore, the reported leader accuracies cannot be traced to the minimax game-theoretic training loop, and the central claim of the paper is unsupported as written. Since this is the same load-bearing concern the reader identified, my stress-test does not change the reader's REJECT verdict.","tokens_in":8599,"tokens_out":3284,"duration_ms":40963,"concrete_test":"Take a trained leader-network checkpoint for any reported dataset, fix a batch of (SR, ŜR) pairs, and compute Eq. 21. Perturb one predicted sample Ŝ_j by δ with decreasing magnitudes (e.g., 1e-2 through 1e-8) and record I. If the numerical derivative of I with respect to Ŝ_j is exactly zero for all δ smaller than the distance to the selected k-th neighbor, the estimator is not differentiable in the direction required for SGD. Equivalently, an autodiff pass through Eqs. 15–21 will show zero gradient for every predicted output. If a nonzero gradient is claimed, the paper must specify the smoothing or relaxation used; without it, the leader update in Algorithm 1 does not exist.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing gap is the leader's maximization step. Eq. 21 estimates I(SR; ŜR) using hard k-nearest-neighbor counts n_f, n_v, n_p from Eqs. 18–20. These counts are piecewise constant functions of the network outputs ŜR: for almost every output vector, a sufficiently small perturbation changes no neighbor relations, so ∂I/∂Ŝ_j = 0 almost everywhere. Algorithm 1 instructs the agents to 'Compute I with Eq. 21' and then 'Solve MiniMax optimization with Eq. 9', but it provides no surrogate loss, no relaxation, and no differentiable approximation through which gradients can flow. The training section states that SGD with momentum is used for the neural networks, yet no training objective for the leader network is specified other than maximizing this mutual information. Consequently, the leader-follower training loop described in the text is not an executable algorithm, and the reported leader accuracies (80.62–95.03%) cannot be attributed to the game-theoretic maximization of mutual information. This is central: without the leader's I-maximization step, Eq. 9 is a formal utility expression that is never optimized, and the claimed cooperative Stackelberg solution is unsupported. This is not merely a missing implementation detail; the structure of the estimator makes the natural gradient-based update degenerate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LeFo, a Leader-Follower game-theoretic framework for predicting haptic signals in Tactile Internet to mitigate the effects of packet loss and delay. The human side (Leader) is trained to maximize mutual information between actual and predicted robot signals, while the robot side (Follower) minimizes KL divergence between actual and predicted human signals. The authors report prediction accuracies of 80.62--95.03% for robot signals at the human side and 70.44--89.77% for human signals at the robot side across several haptic datasets, and they claim a Taylor-expansion-based upper bound on the loss for robustness. The paper includes an algorithm, experimental results, and an inference-time table.","tokens_in":8826,"tokens_out":5015,"duration_ms":55249,"significance":"If the claims were substantiated, the LeFo framework would offer a new framing for predictive loss mitigation in Tactile Internet, with potential practical value for relaxing latency constraints in teleoperation. The paper has some strengths: it uses real haptic interaction data, provides a detailed network architecture, reports standard deviations over multiple runs, and honestly discloses that inference times currently exceed the 1 ms target. However, the central game-theoretic contribution is largely nominal, the reported algorithm is not executable as written, and the theoretical bound does not actually address signal loss. The absence of baselines and an undefined accuracy metric further prevent the results from being interpretable.","major_comments":[{"comment":"The minimax formulation in Eq. (9) is not a well-posed game because the two terms are independent: the leader's utility I(SR(n+1); ŜR(n+1)) depends only on the human-side prediction network, while the follower's utility KL(SH(n+1); ŜH(n+1)) depends only on the robot-side prediction network. There is no coupling between the strategies of H and R in the objective, so the problem reduces to two separate supervised learning tasks. The claimed Stackelberg equilibrium and cooperative interaction are therefore not realized by this equation.","section":"Eq. (9)"},{"comment":"Algorithm 1 is not executable because the KNN mutual-information estimator in Eq. (21) is a piecewise-constant function of the network outputs: the neighbor counts n_f, n_v, n_p in Eqs. (18)--(20) change only when the Chebyshev distances cross discrete thresholds, so the gradient of I with respect to the network parameters is zero almost everywhere. The instruction to 'Compute I with Eq. 21' and then 'Solve MiniMax optimization with Eq. 9' provides no differentiable surrogate or relaxation through which gradients can flow. Consequently, the leader's maximization step in Algorithm 1 has no defined gradient-based update, and the reported leader accuracies cannot be attributed to maximizing mutual information.","section":"Algorithm 1, Sec. 2.3"},{"comment":"The Taylor expansion bound in Eq. (29) is a generic local convexity bound on the change in training loss L_{n-1} between two parameter vectors, not a bound on signal loss caused by packet loss or missing samples. The derivation assumes θ_{n-1} is a local minimum (zero gradient), that the Hessian is positive semi-definite, and that higher-order Taylor terms are negligible; none of these conditions is established for the non-convex neural networks used. The empirical verification in Fig. 3 plots the same loss-change quantity against the upper bound, so it does not independently validate the bound as a prediction-error guarantee.","section":"Sec. 2.4, Eq. (29)"},{"comment":"The accuracy metric is never defined. The paper reports 'prediction accuracy' for continuous-valued signals (force, velocity, position) but does not state what threshold or tolerance qualifies as a correct prediction, nor how the percentage is computed. Without this definition, the headline numbers (80.62--95.03% and 70.44--89.77%) are uninterpretable, and the comparison across features and between Leader and Follower is not meaningful.","section":"Fig. 2 and abstract"},{"comment":"No baseline comparison is provided. The paper concludes that LeFo is 'effective as a baseline method' but does not compare against standard time-series forecasting baselines such as ARIMA, LSTM, or a simple feedforward network trained with mean squared error. The reported accuracies, even if correctly defined, cannot support the claimed effectiveness or the benefit of the game-theoretic formulation without such comparisons.","section":"Results section"}],"minor_comments":[{"comment":"The notation ||zi − ˆzi||∞ is confusing because zi is already defined as the pair (S_i, ˆS_i); the subscript on the second term is not clear. This should be rewritten to define distances between the actual and predicted feature vectors.","section":"Eq. (13)"},{"comment":"The text says 'We use simplification in notation by removing the subscripts and simply using S and ˆS' but then continues to use subscripts such as S_i and ˆS_i, making the notation inconsistent.","section":"Sec. 2.3"},{"comment":"The definitions of e_f, e_v, e_p and the neighbor counts n_f, n_v, n_p appear to deviate from the standard KSG estimator form (e.g., distances in the joint space and counts in marginal spaces). The paper should provide a derivation or a precise citation that justifies this multi-feature formulation.","section":"Eqs. (15)--(20)"},{"comment":"The inference times (6.5--22.1 ms) exceed the 1 ms target, and while the paper acknowledges this, it weakens the claim that the method addresses ultra-low-latency TI requirements. The statement that inference acceleration can be applied 'with marginal accuracy loss' is speculative.","section":"Sec. 3, Table 1"},{"comment":"The paper relies on the authors' own preprints ([10], [12]) for the MI estimator and for background on Hessians. Independent references or a more detailed derivation would strengthen the presentation. The code link appears only as 'GitHub Repository' without an actual URL.","section":"References"},{"comment":"The axes of Fig. 3 are labeled '(Loss Change)' and 'Upper Bound' with numerical values 0.0--1.0, but the units and normalization are not explained.","section":"Fig. 3"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be a preliminary draft. The central game-theoretic formulation is nominal (the utility functions are independent), the proposed algorithm is not implementable due to the non-differentiability of the KNN MI estimator, and the Taylor bound does not address signal loss. The reliance on the authors' own unpublished references for key components, and the verification of the bound on the very quantity it bounds, raise additional concerns about the substantiation of the claims. The paper would need a fundamental reformulation (e.g., a differentiable MI estimator, a genuinely coupled game objective, a proper accuracy definition, and baseline comparisons) to be viable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — bottom line: the paper's headline claim rests on a training loop that cannot be run as written. The KNN mutual-information estimator in Eq. 21 is a piecewise-constant function of the leader network's outputs; small perturbations of the outputs change no neighbor relations, so any gradient-based maximization of I has zero gradient almost everywhere. Algorithm 1 tells the agents to 'Compute I' and then 'Solve MiniMax optimization with Eq. 9', but gives no surrogate, relaxation, or differentiable approximation. So the reported leader accuracies (80.62–95.03%) cannot be attributed to the game-theoretic MI maximization; at best they come from an unspecified training objective.\n\nThat is the load-bearing flaw, and it's not a minor implementation detail. The Taylor bound in Eq. 29 is also mislabeled: it is the standard local-convexity bound on the change in training loss under a stationary-point assumption, not a bound on signal loss from missing haptic packets. So the robustness guarantee does not do the advertised work.\n\nWhat the paper does have going for it: real haptic datasets (Novint Falcon / Chai3D), a table of prediction accuracies with standard deviations, and an honest acknowledgment that the measured inference times (6.2–22.1 ms) miss the 1 ms target. The Stackelberg framing—leader maximizes MI, follower minimizes KL—is a reasonable way to describe two separate supervised losses, but it is a relabeling, not a new method. Eq. 4 is a standard ARMA predictor, and the two neural networks are standard feedforward regressors.\n\nOther soft spots, in proportion: no baseline comparison (the whole point of a prediction method), an undefined accuracy metric (the percentages in Fig. 2 are never defined as, e.g., thresholded relative error), and a code link that is a placeholder. Self-citations to [10] and [12] are fine, but they don't fill these gaps.\n\nWho is this for? Someone working on haptic teleoperation might skim the experimental results, but the method section needs a complete rewrite before the numbers mean anything. I would not send this to peer review in its current form; the central algorithm is not executable from the text. If the authors replace the KNN estimator with a differentiable MI lower bound (e.g., InfoNCE or a variational estimator) and add baselines, it could become a decent workshop paper. For now, desk reject.","headline":"The LeFo training loop is not executable as written—the KNN MI estimator has zero gradient almost everywhere—and the Taylor bound is a local-loss bound, not a packet-loss bound; the empirical numbers lack a defined metric and baselines.","tokens_in":9373,"tokens_out":2874,"would_cite":false,"duration_ms":31897,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that modeling human-robot interaction in a tactile Internet as a leader-follower Stackelberg game, where the human maximizes mutual information and the robot minimizes KL divergence, enables accurate one-step-ahead…","keywords":["Tactile Internet","Signal prediction","Stackelberg game","Leader-follower","Mutual information","K-nearest neighbors","Packet loss mitigation","Haptic signals"],"falsifier":"Run the LeFo training loop on a synthetic pair of signals with analytically known mutual information; if the KNN estimate cannot be back-propagated or its value does not increase under the leader update step, then the leader's reported accuracy cannot be attributed to the mutual-information maximization and the central mechanism fails.","tokens_in":8342,"feed_emoji":"🖐️","tokens_out":7969,"duration_ms":71695,"temperature":0.7,"pith_summary":"This paper tries to establish that packet loss in tactile Internet teleoperation can be repaired by one-step-ahead prediction cast as a cooperative Stackelberg game: the human acts as leader and predicts the robot's next force, velocity, and position signal by maximizing mutual information between the robot's actual and predicted feedback, while the robot acts as follower and predicts the human's next signal by minimizing KL divergence. If true, teleoperation systems could tolerate a lost sample by predicting it, relaxing the ultra-low-latency requirement. The paper reports prediction accuracies of 80.62–95.03% for robot signals at the human side and 70.44–89.77% for human signals at the robot side, and derives the loss-change bound $L_{n-1} \\leq \\frac{1}{2}\\lambda^{n-1}_{\\max}\\|\\Delta\\theta\\|^2$ from a second-order Taylor expansion of the previous loss. A sympathetic reading is that the leader-follower asymmetry, not the neural architecture alone, is what produces the accuracy gap, with force signals showing the largest gap.","feed_headline":"Leader-follower game predicts haptic signals up to 95% accuracy","feed_subtitle":"A Stackelberg minimax objective lets humans and robots forecast each other's next move, easing strict delay limits.","key_machinery":"The central mechanism is the leader-follower (LeFo) minimax objective of Eq. 9, which couples the two agents' prediction goals: the human leader drives the shared system by maximizing the mutual information $I(S_R;\\hat{S}_R)$ between the robot's actual and predicted feedback, while the robot follower minimizes the Kullback–Leibler divergence $KL(S_H;\\hat{S}_H)$ between the human's actual and predicted signal. The mutual information is estimated with a K-nearest-neighbor estimator built on Chebyshev distances, neighbor counts, and the digamma function, and the robustness guarantee is the second-order Taylor bound $L_{n-1} \\leq \\frac{1}{2}\\lambda^{n-1}_{\\max}\\|\\Delta\\theta\\|^2$, where $\\lambda_{\\max}$ is the largest eigenvalue of the Hessian of the previous loss.","core_discovery":"The central claim is that bidirectional haptic signal prediction reduces to the minimax objective $\\min_R \\max_H L(H,R,\\theta_n) = \\mathbb{E}_{S_H}[\\max I(S_R;\\hat{S}_R)] - \\mathbb{E}_{S_R}[\\min KL(S_H;\\hat{S}_H)]$, and that training one neural network per role with these utilities produces the reported accuracies. The paper further claims that the human leader's deeper 12-layer network achieves consistently higher accuracy than the robot follower's 8-layer network, that force features exhibit the largest leader-follower gap (8–12%), and that the training loss change is bounded by the largest Hessian eigenvalue times the squared parameter update, giving a robustness guarantee against signal loss.","pith_inferences":["Editorial extension: the reported leader-over-follower accuracy gap is confounded by network size (12 vs 8 layers) and by different utility functions; an ablation with identical architectures and swapped utilities would test whether the Stackelberg role itself explains the gap.","Editorial extension: the Taylor bound's practical value depends on estimating $\\lambda_{\\max}$ cheaply; a Hessian-vector-product routine could turn the bound into an online loss-spike monitor during training.","Editorial extension: the one-sample bound could be composed to cover multi-sample outages, but error accumulation across composed predictions would need separate analysis before claiming longer recovery windows."],"forward_implications":["Teleoperation can tolerate a single missing sample by predicting it, so the effective latency constraint shifts from the packet arrival time to the predictor's inference time.","The human leader side benefits more from prediction than the robot follower side, with force signals showing an 8–12% accuracy gap, indicating force is the most responsive haptic channel for loss mitigation.","The Taylor bound guarantees that small parameter updates between consecutive signals cannot produce an unbounded prediction loss, even when the training objective changes.","With inference times of 6.5–22.1 ms, the current implementation does not yet meet the 1 ms tactile Internet target, but the method establishes a baseline that lighter models could approach."],"supporting_citations":[{"why":"Defines tactile Internet and its ultra-low latency and reliability requirements, the target the prediction method addresses.","marker":"[1]"},{"why":"Provides the Stackelberg game model on which the leader-follower formulation is built.","marker":"[6]"},{"why":"Supplies the K-nearest-neighbor mutual information estimator (Eq. 21) used for the leader's utility.","marker":"[10]"},{"why":"Supports the KNN-based mutual information estimation approach alongside [10].","marker":"[11]"},{"why":"Cited for the positive-semi-definite Hessian property used in the Taylor expansion bound.","marker":"[12]"},{"why":"Provides the real-world kinaesthetic interaction dataset used for all reported experiments.","marker":"[13]"},{"why":"Gives the He initialization scheme used for the fully connected neural networks.","marker":"[14]"},{"why":"Supplies dropout regularization used in network training.","marker":"[15]"}],"fun_headline_variants":["Stackelberg game predicts haptic signals up to 95% accuracy","Leader-follower AI forecasts haptic moves with 95% accuracy","Game-theoretic haptic prediction hits 95% accuracy","Minimax model predicts tactile signals up to 95% accuracy","Bidirectional haptic forecast via Stackelberg reaches 95%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training loop assumes the K-nearest-neighbor mutual information estimator in Eq. 21 is differentiable with respect to the leader network's parameters and can be maximized by an optimizer, yet no gradient or surrogate loss is provided.","fun_headline_variants_meta":{"raw":{"variants":["Stackelberg game predicts haptic signals up to 95% accuracy","Leader-follower AI forecasts haptic moves with 95% accuracy","Game-theoretic haptic prediction hits 95% accuracy","Minimax model predicts tactile signals up to 95% accuracy","Bidirectional haptic forecast via Stackelberg reaches 95%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1510,"prompt_tokens":847,"completion_tokens":663,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":573}},"tokens_in":463,"tokens_out":663,"duration_ms":6791,"temperature":1.0,"reasoning_tokens":573,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:35:09.341735+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the LeFo training loop on a synthetic pair of signals with analytically known mutual information; if the KNN estimate cannot be back-propagated or its value does not increase under the leader update step, then the leader's reported accuracy cannot be attributed to the mutual-information maximization and the central mechanism fails.","supporting_citations":[{"cited_title":"a network, or a network of networks, for remotely accessing, perceiving, manipu- lating, or controlling real and virtual objects or processes in perceived real-time","cited_arxiv_id":null,"evidence_quote":"Defines tactile Internet and its ultra-low latency and reliability requirements, the target the prediction method addresses."},{"cited_title":"A comprehensive survey of the tactile internet: State-of-the-art and research directions,","cited_arxiv_id":null,"evidence_quote":"Provides the Stackelberg game model on which the leader-follower formulation is built."},{"cited_title":"Green and mobility-aware caching in 5g networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the K-nearest-neighbor mutual information estimator (Eq. 21) used for the leader's utility."},{"cited_title":"A game theory competitive intelligence solution stimulated from a stackelberg game: A three players scenario,","cited_arxiv_id":null,"evidence_quote":"Supports the KNN-based mutual information estimation approach alongside [10]."},{"cited_title":"A game-theoretic analysis of joint mobile edge caching and peer content sharing,","cited_arxiv_id":null,"evidence_quote":"Cited for the positive-semi-definite Hessian property used in the Taylor expansion bound."},{"cited_title":"Deep Reinforcement Learning-based Video-Haptic Radio Resource Slicing in Tactile Internet","cited_arxiv_id":"2503.14066","evidence_quote":"Provides the real-world kinaesthetic interaction dataset used for all reported experiments."},{"cited_title":"Intelligent communications for tactile internet in 6g: Requirements, technologies, and chal- lenges,","cited_arxiv_id":null,"evidence_quote":"Gives the He initialization scheme used for the fully connected neural networks."},{"cited_title":"Shapley-Based Data Valuation with Mutual Information: A Key to Modified K-Nearest Neighbors","cited_arxiv_id":"2312.01991","evidence_quote":"Supplies dropout regularization used in network training."}],"review_version":1}