{"id":"48f30ba3-2420-4561-bc29-336a52a7363b","arxiv_id":"2506.18584","paper_version":2,"verdict":"REJECT","confidence":"LOW","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A temperature-aware, stochastic offloading policy (TAO) keeps wearable XR devices within power, thermal, and battery limits while cutting offloading cost by over 35% versus a thermal-ignoring baseline.","lead":"This paper proposes a rule for when wearable XR devices should do computations locally versus sending them to an edge server, accounting for power limits, device temperature, and battery life simultaneously. It matters because overheating is a real safety constraint for head-mounted devices in contact with skin, and most prior offloading work ignores it.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The confidence guarantee is vacuous: Eq. (10) evaluates the CDF at its own ω-quantile and never at P_max, and Eq. (14) inherits the backwards inequality direction, so TAO's core feasibility claim is unproven.","rationale":"The reader's verdict of REJECT is appropriate, but the most load-bearing defect is not primarily the thermal-model fidelity. The paper's defining contribution is a stochastic policy with a tunable confidence guarantee; that guarantee is not established because the chance-constraint direction in Eq. (10)–(14) is either tautological or inverted. Even if the thermal impulse response in Eq. (20) were perfectly accurate and B and θ were given, the optimization would still not be a valid solution to the stated constraints. The thermal-model concern is real and contributes to the lack of auditability, but it is secondary: the sign/quantile issue attacks the internal logic of the formulation. The COMSOL evaluation is also too thin to rescue the claim: no α values, no exact SoTA cost, and the text says TAO increases offloading cost for Google Glass, which conflicts with the abstract's 35% reduction. These are consistency problems, but the chance-constraint issue is the one that must be resolved before any experimental validation can be meaningful. I credit the paper for identifying a genuinely relevant multi-timescale constraint and for using COMSOL geometries of real wearables, and I do not see fraud; the problems are mathematical and presentational. Because the core proof and evaluation are unsupported, the REJECT verdict stands, possibly revisable if the constraints are corrected and the simulation parameters are supplied.","tokens_in":7847,"tokens_out":7324,"duration_ms":75481,"concrete_test":"Take the HoloLens parameters used in Fig. 4 (TDP, π_d, δ, request stream) and one candidate α at time t=500 s. Compute F_p(P_max) from Eq. (13). If F_p(P_max) < ω, then enforcing Eq. (14) with '≤ ω' excludes exactly the safe α values and admits unsafe ones. Repeat with '≥ ω'; the feasible sets differ materially. A second part: independently re-derive Eq. (19) and Eq. (21) from the definitions in Sec. II and check whether the CDF argument is the physical limit (P_max, τ_max, battery capacity) or the ω-quantile. If the arguments are quantiles, the constraints are vacuous.","verdict_should_be":"UNCHANGED","load_bearing_attack":"TAO's central promise is that the stochastic policy meets power, thermal, and battery limits with confidence ω. For that to be true, the optimization must impose Pr(p(d,t) ≤ P_max(d)) ≥ ω (and similarly for τ and b). Eq. (10) instead states F_{p(d,t)}(T_p) ≤ ω, where T_p is defined as the ω-quantile of p(d,t); this is a tautology, since a CDF at its own quantile equals ω (up to atoms). It does not reference P_max. The only place the limit enters is Eq. (13), which correctly computes Pr(p ≤ P_max) as a Poisson CDF. Eq. (14) then replaces P_max/π_d by T_p/π_d with no derivation, and keeps the direction '≤ ω'. If T_p is meant to be P_max, the inequality should be '≥ ω', not '≤ ω'. The same defect propagates to Eq. (19) (where T_b is never defined) and Eq. (21). Consequently the closed-form expressions do not define a feasible policy for the stated chance-constrained problem: no value of α satisfying Eq. (14) has been shown to give the advertised confidence. The COMSOL section does not report the α values or the constraint slack, so the temperature traces cannot be attributed to the optimization. This is an internal-logic failure, independent of whether the LTI thermal model is accurate; it must be fixed before the thermal-model question can even be tested.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This letter studies edge offloading for XR wearables under instantaneous power, short-term temperature, and long-term battery constraints. The authors formulate a stochastic optimization in which a stationary per-request offloading probability is chosen so that all constraints are met with tunable confidence omega, derive closed-form Poisson-based quantile constraints for the special case of homogeneous requests, and evaluate the resulting policy, TAO, in COMSOL simulations against a thermal-unaware baseline (SoTA) and a plugged-to-AC greedy policy. The paper claims that TAO reduces offloading cost by over 35% while avoiding violations of the 43 C temperature limit.","tokens_in":8121,"tokens_out":4777,"duration_ms":50725,"significance":"The multi-timescale problem framing is well motivated, and the use of detailed COMSOL models of commercial wearables (HoloLens and Google Glass) is a genuine strength that goes beyond abstract thermal abstractions. A tunable-confidence, stationary offloading policy that jointly respects power, thermal, and battery limits would be a useful contribution to the XR offloading literature. However, the central feasibility guarantee is not demonstrated: the quantile constraints are not correctly formulated, the closed-form expressions are internally inconsistent, and the evaluation does not report the quantities needed to connect the simulation results to the proposed optimization. As written, the paper's main claim is therefore unsupported.","major_comments":[{"comment":"The robust constraints in Eq. (10) are tautological and do not reference the actual operational limits. Since T_p is defined as the omega-quantile of p(d,t), the inequality F_{p(d,t)}(T_p) <= omega holds by definition up to atoms, and the same applies to T_tau and T_b. The required power chance constraint is Pr(p(d,t) <= P_max(d)) >= omega, which is never stated or used. Consequently Eq. (10) provides no feasibility guarantee for the problem described in Eqs. (6)-(8). This is a load-bearing defect in the central claim.","section":"Section III, Eq. (10)"},{"comment":"The closed-form constraints have the wrong inequality direction and an unexplained threshold substitution. From Eq. (13), enforcing a confidence omega requires the Poisson CDF evaluated at P_max(d)/pi_d to be >= omega. Eq. (14) instead replaces P_max(d)/pi_d by T_p/pi_d and writes '<= omega'; this both changes the threshold to the omega-quantile and reverses the inequality. No value of alpha satisfying Eq. (14) has been shown to imply Pr(p(d,t) <= P_max(d)) >= omega. The same defect propagates to Eq. (19), where T_b is never defined and the battery constraint is reduced to the deterministic mean inequality in Eq. (18), and to Eq. (21). The closed-form expressions therefore do not implement the advertised chance constraints.","section":"Section III.A, Eqs. (12)-(21)"},{"comment":"The thermal response used in the closed-form example is not justified and is inconsistent with the LTI convolution model in Eq. (5). Eq. (5) defines temperature as the convolution of power with an impulse response h(d,t), whereas Eq. (20) describes the response to a rectangular power pulse as a difference of two first-order step responses. Parameters B and theta are never fitted to the COMSOL models or otherwise justified, and the connection between the per-request power pi_d, the thermal resistance R_th, and the Poisson request count is not specified. Even after correcting the algebraic defect in Eq. (14), Eq. (21) would not express Pr(tau(d,t) <= tau_max) >= omega without a derivation of how the temperature limit maps to a count threshold.","section":"Section III.A, Eq. (20)"},{"comment":"The performance evaluation does not report the alpha values chosen by TAO, the empirical frequency of power/temperature/battery constraint violations, or which of the equations in Section III were actually solved in the simulations. The COMSOL temperature traces and the cost comparisons therefore cannot be attributed to the proposed stochastic optimization. In addition, the abstract's claim of a cost reduction exceeding 35% is not tied to a specific comparison number in the text or in Table II; the table reports only normalized TAO costs, with no corresponding SoTA value.","section":"Section IV"}],"minor_comments":[{"comment":"The constraints in Eqs. (7)-(8) are written as deterministic inequalities on random variables without a precise probabilistic quantifier; the intended meaning of 'with confidence omega' should be stated formally.","section":"Section II, Eqs. (7)-(8)"},{"comment":"The notation f_{p/b/tau}(d,t)(u) in Eq. (9) is difficult to parse; the subscript should be the random variable itself, e.g., f_{p(d,t)}(u), f_{b(d,t)}(u), f_{tau(d,t)}(u).","section":"Section III, Eq. (9)"},{"comment":"The event N_d(t) >= N_d(T) in Eq. (15) is not the event used in the power constraint (12); please clarify what probability this expression is intended to compute.","section":"Section III.A, Eq. (15)"},{"comment":"The left panel and caption describe 'impulse responses,' but the plotted curves are the response to a finite-duration power pulse and are better described as step responses; the terminology should be corrected.","section":"Fig. 3"},{"comment":"The table heading contains the typo 'Normilized'; it should be 'Normalized.' Also, including the corresponding SoTA cost would make the claimed 35% improvement verifiable.","section":"Table II"},{"comment":"The symbol T is used both for the time horizon and for the duration of the pulse stimulus in Eq. (20), which is confusing; a different symbol should be used for the pulse duration.","section":"Eq. (20)"}],"recommendation":"reject","confidential_remarks":"The defects identified above are not presentation-level issues: the core chance-constraint derivation is internally inconsistent, and the closed-form expressions cannot support the stated feasibility guarantee. I see no path to acceptance without re-deriving the stochastic constraints and re-running the evaluation with the corrected formulation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2506.18584. The framing is genuinely useful: offloading for XR wearables spans three time scales—instantaneous power, skin temperature, battery—and adding an explicit thermal constraint to a stochastic offloading policy is a legitimate extension. The COMSOL models of HoloLens and Google Glass are a real asset, and comparing against a temperature-blind baseline is the right experiment. The citation pattern looks fine.\n\nThe central math, though, does not hold up. Eq. (10) says F_p(T_p)≤ω where T_p is the ω-quantile of p. That is true by construction and says nothing about P_max. The actual chance constraint should be F_p(P_max)≥ω, and similarly for temperature and battery. Eq. (14) carries the wrong direction and uses T_p in place of P_max; Eq. (19) uses an undefined T_b; Eq. (21) repeats the same pattern; Eq. (15) is missing its bound. So TAO's advertised confidence guarantee is not derived. This is a load-bearing flaw in the theoretical section.\n\nThe evaluation has two additional soft spots. The thermal impulse response in Eq. (20) depends on B and θ, but the paper never reports their values or justifies the single-exponential form, so the COMSOL temperature traces cannot be tied to the optimization. The abstract's claim of a 35% reduction in offloading cost is contradicted by the body: for Google Glass TAO increases offloading cost relative to SoTA, and for HoloLens the cost is essentially unchanged. That claim appears to compare against something not in the paper.\n\nNone of this makes the central idea worthless. The three-time-scale decomposition is clear, and the COMSOL work is concrete and reproducible in principle. The errors are fixable: correct the inequalities, define the thresholds, report the impulse-response parameters and the α values, and re-state the cost comparison. As it stands, the paper does not establish its main theoretical or numerical claims.\n\nMy recommendation: this deserves peer review rather than desk rejection—the topic is timely and the modeling is a real contribution—but a serious referee should be directed to Sec. III-A first. I would ask for a revision with explicit derivations and a corrected baseline comparison.","headline":"The thermal-aware offloading idea and COMSOL models are worth reading, but the central chance-constrained derivation is vacuous, the inequality directions are backwards, and the 35% cost claim is contradicted by the paper's own body.","tokens_in":8703,"tokens_out":7112,"would_cite":false,"duration_ms":70880,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a single stationary offloading probability can keep an XR wearable within its power, temperature, and battery limits at a tunable confidence, and that doing so cuts offloading cost by more than 35% relative to a…","keywords":["XR offloading","wearable thermal constraints","edge computing","stochastic optimization","Poisson arrivals","temperature-aware offloading","power constraints","battery constraints"],"falsifier":"Run the paper's ten-request, one-hour workload on a physical XR wearable while TAO reports confidence 0.99 and measure skin or surface temperature directly; if temperature exceeds 43°C, or if the measured response to a step in local processing clearly deviates from the single-exponential form of Eq. (20), the central claim is falsified. A cheaper proxy is to compare the multiphysics-derived thermal constants against measured step-response curves for the same device.","tokens_in":7617,"feed_emoji":"🌡️","tokens_out":8865,"duration_ms":90839,"temperature":0.7,"pith_summary":"The paper argues that offloading decisions for XR wearables must be made jointly across three time scales: instantaneous power, short-term temperature, and long-term battery life. It proposes TAO, a stationary stochastic policy that sets the probability each incoming request is processed locally, and shows how to tune that probability so that power, temperature, and battery constraints are each met with a chosen confidence level. Under a linear time-invariant thermal model, the policy has a closed-form solution. In multiphysics 3D simulations of two commercial XR wearable designs, TAO keeps temperature below the 43°C burn-risk threshold while reducing offloading cost by over 35% compared with a state-of-the-art approach that ignores temperature.","feed_headline":"A single offloading policy keeps XR wearables cool and cuts cost 35%","feed_subtitle":"The TAO rule decides which requests run locally, meeting power, heat, and battery limits at a chosen confidence.","key_machinery":"The load-bearing object is the per-device stationary local-execution probability $\\alpha_d$, combined with closed-form Poisson quantile equations for the power, battery, and temperature constraints. The thermal model is the linear time-invariant convolution $\\tau(d,t)=p(d,t)*h(d,t)$, reduced to the single-exponential step response $\\tau(d,t)=B\\bigl[(1-e^{-t/\\theta})-(1-e^{-(t-T)/\\theta})\\bigr]$, which makes the safety threshold a quantile inequality on a Poisson random variable. This machinery collapses all three time scales into cumulative-distribution constraints on the same decision variable $\\alpha_d$, so the stochastic optimization becomes solvable in closed form whenever closed-form impulse responses are available.","core_discovery":"The central claim is that a time-invariant local-execution probability can simultaneously honor instantaneous power, short-term temperature, and long-term battery constraints when those constraints are expressed as quantiles of the request process. Because requests arrive as a Poisson process and each request consumes a fixed power increment, power, temperature, and battery level have Poisson-based distributions, so TAO computes the local-execution probability that maximizes local computation subject to each cumulative-distribution constraint. For the thermal branch, the paper adopts a first-order impulse response with thermal resistance $B$ and time constant $\\theta$, which turns the 43°C limit into an explicit inequality on the Poisson cumulative distribution. In the reported simulations this policy never exceeds the temperature limit, whereas the baseline that drops the thermal constraint violates it for roughly 5% of the operating time.","pith_inferences":["Because the guarantee is probabilistic, an unlucky burst of arrivals can push temperature past 43°C with probability $1-\\omega$; the reported safety is a tunable confidence, not a hard certificate for all possible sample paths.","The 35% cost saving is likely sensitive to the thermal time constant and thermal resistance, so measuring $B$ and $\\theta$ for other wearable models would directly test how far the result generalizes.","The same cumulative-distribution-quantile machinery could be carried over to other skin-contact devices, such as VR headsets or smart glasses with different heat sinks, by replacing only the thermal impulse response.","A natural extension would be to replace the single-exponential thermal response with a measured multi-exponential response and re-derive the quantile constraint numerically, which the paper explicitly leaves open."],"forward_implications":["Lowering the confidence level from 0.99 to 0.80 cuts the normalized offloading cost from 0.84 to 0.33, so the same policy can be tuned to trade safety margin for cost in less critical scenarios.","Because the policy is stationary, it can be implemented as a simple per-device local-execution probability rather than an online per-request scheduler.","The method remains applicable when the thermal impulse response is not known in closed form, since TAO only needs the cumulative distributions, which can be obtained from measurements.","Thermal awareness comes nearly free on devices with ample thermal headroom but can require extra offloading on thermally constrained designs; the evaluation shows both behaviors across the two devices considered.","A purely battery- and power-aware strategy can violate the temperature limit for a non-negligible fraction of operating time, so thermal modeling is not a refinement but a necessary constraint for skin-contact XR devices."],"supporting_citations":[{"why":"Supplies the convolution-based dynamic compact thermal model on which Eq. (5) rests.","marker":"[12]"},{"why":"Justifies treating heat capacity and thermal resistance as constant in handheld electronic devices.","marker":"[11]"},{"why":"Supports the constant-property thermal approximation specifically for smart glasses.","marker":"[13]"},{"why":"Provides the linear-superposition result used to write the temperature response as two step responses in Eq. (20).","marker":"[14]"},{"why":"Fixes the 43°C threshold beyond which low-temperature burns may occur.","marker":"[10]"},{"why":"Supplies the still-air heat-transfer coefficient used as the boundary condition in the thermal simulations.","marker":"[15]"},{"why":"Motivates minimizing offloading cost as the objective for edge-supported XR systems.","marker":"[9]"}],"fun_headline_variants":["TAO offloading policy cuts XR cost 35% while meeting all limits","One offloading rule keeps XR cool and cuts cost 35%","Temperature-aware offloading slashes XR cost 35% safely","Offloading policy balances power, heat, battery for XR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The no-overheating guarantee rests on assuming that each device heats and cools along a single-exponential linear curve with fixed thermal constants; if real wearables respond nonlinearly or with much slower dynamics, the temperature check is misspecified and the guarantee collapses.","fun_headline_variants_meta":{"raw":{"variants":["TAO offloading policy cuts XR cost 35% while meeting all limits","One offloading rule keeps XR cool and cuts cost 35%","Temperature-aware offloading slashes XR cost 35% safely","Offloading policy balances power, heat, battery for XR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000715,"raw_usage":{"total_tokens":3168,"prompt_tokens":855,"completion_tokens":2313,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":2235}},"tokens_in":471,"tokens_out":2313,"duration_ms":15322,"temperature":1.0,"reasoning_tokens":2235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:45:05.217692+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the paper's ten-request, one-hour workload on a physical XR wearable while TAO reports confidence 0.99 and measure skin or surface temperature directly; if temperature exceeds 43°C, or if the measured response to a step in local processing clearly deviates from the single-exponential form of Eq. (20), the central claim is falsified. A cheaper proxy is to compare the multiphysics-derived thermal constants against measured step-response curves for the same device.","supporting_citations":[{"cited_title":"A deep reinforcement learning-based offloading scheme for multi-access edge computing-supported extended reality systems,","cited_arxiv_id":null,"evidence_quote":"Motivates minimizing offloading cost as the objective for edge-supported XR systems."},{"cited_title":"Dynamic compact thermal models for skin temperature prediction of portable electronic devices based on convolution and fitting methods,","cited_arxiv_id":null,"evidence_quote":"Supplies the convolution-based dynamic compact thermal model on which Eq. (5) rests."},{"cited_title":"Parametric thermal modeling of heat transfer in handheld electronic devices,","cited_arxiv_id":null,"evidence_quote":"Justifies treating heat capacity and thermal resistance as constant in handheld electronic devices."},{"cited_title":"Thermal model and countermeasures for future smart glasses,","cited_arxiv_id":null,"evidence_quote":"Supports the constant-property thermal approximation specifically for smart glasses."},{"cited_title":"Using linear superposition to solve multiple heat source transient thermal problems,","cited_arxiv_id":null,"evidence_quote":"Provides the linear-superposition result used to write the temperature response as two step responses in Eq. (20)."},{"cited_title":"Is It Getting Hot in Here? The Effects of VR Headset Microclimate Temperature on Perceived Thermal Discomfort, VR Sick- ness, and Skin Temperature,","cited_arxiv_id":null,"evidence_quote":"Fixes the 43°C threshold beyond which low-temperature burns may occur."},{"cited_title":"Cengel and A","cited_arxiv_id":null,"evidence_quote":"Supplies the still-air heat-transfer coefficient used as the boundary condition in the thermal simulations."}],"review_version":1}