{"id":"3b2b20e3-92e0-479a-be62-665538245304","arxiv_id":"1908.04446","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A context-aware timeliness metric multiplies estimation error by a context weight, and a Lyapunov scheduling policy that minimizes it cuts threshold violations and improves CartPole control versus Age-of-Information scheduling.","lead":"This paper proposes a new metric, context-aware information lapse, that multiplies the error in a remote status estimate by a context-dependent importance weight. It also derives a multi-user scheduling policy based on Lyapunov optimization and reports substantial gains over Age-of-Information scheduling in simulations and a CartPole control task.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof silently requires A_i(t) and omega_i(t+1) to be independent; the CartPole weight shares the same random force as A_i(t), so the headline experiment may fall outside the theorem's assumptions.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: the theorem's proof needs independence between the error increment A_i(t) and the next-slot context weight omega_i(t+1), and the CartPole definition of omega via x*vx and alpha*alphadot makes that independence implausible because the same random force enters both quantities. This is the most load-bearing concern because the paper's central contribution consists of a provable performance bound plus a numerical demonstration in CartPole; if the theorem does not apply to the headline experiment, the claim that the context-aware policy has a provable bound in that setting is unsupported. I do not think this forces rejection: the Lyapunov algebra is coherent once the missing independence is assumed, and the synthetic i.i.d. simulation in Fig. 5 satisfies the assumptions, so the theorem stands for that model. The paper should either state the stronger independence assumption explicitly or qualify the CartPole results as a heuristic application of the policy. The numerical claims also lack code, error bars, and the closest effective-age baseline, which the reader noted; those are secondary but reinforce the conditional posture. Overall, the reader's CONDITIONAL verdict remains appropriate, hence UNCHANGED.","tokens_in":9149,"tokens_out":15209,"duration_ms":148160,"concrete_test":"Re-run the CartPole simulation with logging enabled and compute, per state dimension, the sample value of E[omega_i(t+1) A_i(t)^2] - E[omega_i] E[A_i^2] and the correlation between A_i(t) and omega_i(t+1). If this quantity is nonzero at a practically significant level, the theorem's final step fails in the CartPole setting; then re-run Fig. 7 with a context weight computed from the last delivered state only, so that omega(t+1) is independent of the undelivered random force, and check whether the improvement over AoI persists.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 1 is valid only if, in addition to the stated assumptions, A_i(t) is independent of (Q_i(t), U_i(t), omega_i(t+1)) and, for the step 'By Eq. (10)', E[omega_i(t+1) A_i(t)^2] = E[omega_i] E[A_i^2]. Section III states only that A_i(t) is independent of Q_i(t) and that omega_i(t) is independent of Q_i(t); it never states independence between A_i(t) and omega_i(t+1). In the CartPole experiment (Section IV), omega_i(t+1) is defined by sign(x*vx) and sign(alpha*alphadot) at the next step, while A_i(t) is produced by the same random force that changes x and alpha, so the two are driven by a common disturbance. The final averaging step therefore fails for the headline experiment, and the theorem's bound cannot be invoked there; moreover, in the nonlinear CartPole dynamics A_i(t) may also fail to be zero-mean or independent of Q_i(t), so the drift cancellation in Appendix A is not justified. The synthetic i.i.d. simulation satisfies the independence assumptions, so Theorem 1 is not disproved; the gap is that the model assumptions are not stated at the level needed to cover the paper's flagship application.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the context-aware information lapse F(t)=ω(t)δ(Q(t)) as a timeliness metric for remote control systems, where ω captures context-dependent importance and δ(Q) captures the error between actual and estimated status. For a multi-user status update system with K of N users served per slot, it formulates the minimization of average squared lapse, proposes a drift-plus-penalty scheduling policy (Eq. (6)) that serves the K users with the largest values of (E[ω_i](1/(p_iπ_i)-1)+ω_i(t+1))p_iQ_i(t)^2, and proves an upper bound on the average lapse under this policy (Theorem 1). Numerical results compare the proposed policy with round-robin, AoI-based, and context-aware AoI baselines in an i.i.d. model and in an OpenAI CartPole control task, reporting substantial gains in threshold-violation probability and control performance.","tokens_in":9430,"tokens_out":10565,"duration_ms":93513,"significance":"The paper addresses a real limitation of AoI by coupling staleness with both the stochastic evolution of the status and its context-dependent importance. The Lyapunov drift-plus-penalty framework is a standard and, apart from the independence issue detailed below, algebraically sound vehicle for deriving an index policy. The CartPole evaluation is a useful step beyond synthetic queueing simulations, and the paper explicitly compares against several baselines. If the missing independence assumptions are added and the CartPole experiment is appropriately qualified as a heuristic case study, the proposed metric and policy would be a meaningful contribution to context-aware status update design.","major_comments":[{"comment":"The proof of Theorem 1 silently requires that A_i(t) is independent of ω_i(t+1) (and of the scheduling/channel state) in addition to the assumptions stated in Section III. Specifically, the drift calculation cancels the cross term E[(1-D_i(t))Q_i(t)A_i(t)], and the final step from Eq. (10) to the bound replaces E[ω_i(t+1)A_i(t)^2] with E[ω_i]E[A_i^2]. Section III only assumes A_i(t) independent of Q_i(t) and ω_i(t) independent of Q_i(t); it never states independence between A_i(t) and ω_i(t+1). As written, Theorem 1 is therefore not proven under the stated model assumptions.","section":"Section III / Appendix A"},{"comment":"The CartPole experiment defines the context weight as ω_i(t+1)=9 when x(t+1)v_x(t+1)>0 (and similarly for α, α̇), where x, v_x, α, α̇ evolve under the same random force that generates A_i(t). Hence A_i(t) and ω_i(t+1) are dependent, violating the condition needed for Theorem 1. Additionally, because the weight depends on the actual state while Q_i(t) is the estimation error, ω_i(t) is not independent of Q_i(t), contrary to the Section III assumption. The paper should either modify the CartPole setup to satisfy the model assumptions or explicitly state that the CartPole results are a heuristic demonstration outside the scope of the theoretical guarantee.","section":"Section IV (CartPole)"}],"minor_comments":[{"comment":"The text spells 'Karush-Kuhn-Tucker' as 'Karush-Kuhn-Tucher'; this should be corrected.","section":"Section III"},{"comment":"The telescoping step after Eq. (12) silently shifts the time index in the sum over E[ω_i(t+1)Q_i(t+1)^2] and drops boundary terms. The authors should state the limsup argument and the bounded-second-moment condition needed for the boundary terms to vanish.","section":"Appendix A"},{"comment":"The threshold-violation probabilities are only described qualitatively as 'approximately 10^-4' and 'approximately 1/50'; reporting the precise numbers or confidence intervals would make the improvement easier to assess.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The central theorem is likely repairable by adding the missing independence assumption and qualifying the CartPole experiment, but as it stands the paper's main theoretical guarantee does not cover its headline numerical application. The authors should also verify whether the CartPole noise model makes A_i(t) zero-mean and independent of Q_i(t); if not, a brief discussion of the breakdown would be appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper's metric is a nice, simple extension of AoI and the Lyapunov scheduling policy is standard but effective. The math is mostly sound under the assumptions the authors actually state. The main problem is that the theorem's proof silently needs independence that the headline CartPole experiment almost certainly violates, so the central evidence is weaker than the text claims.\n\nWhat is genuinely new and good: F(t)=omega(t)*delta(Q(t)) cleanly separates context importance from estimation error, and reduces to AoI and non-linear AoI in special cases. That is a useful framing for the status-update literature. The scheduling policy (6) is a straightforward drift-plus-penalty design, but the context-weight term gives it a natural interpretation and the numerical gains over plain AoI are substantial in the i.i.d. synthetic scenario. The paper also correctly notes that predicting one-step-ahead weight helps only a little, which is honest.\n\nThe soft spot is real and load-bearing. Appendix A's final averaging step needs E[omega_i(t+1) A_i(t)^2] = E[omega_i] E[A_i^2]. Section III only states that A_i(t) is independent of Q_i(t) and that omega_i(t) is independent of Q_i(t); it never states independence between A_i(t) and omega_i(t+1). In the CartPole experiment, omega_i(t+1) is defined by sign(x*vx) and sign(alpha*alphadot) at the next step, while A_i(t) comes from the same random force that changes x and alpha, so the two are correlated. The theorem's bound therefore does not apply to the paper's flagship application. In addition, the nonlinear CartPole dynamics may violate the zero-mean and independence assumptions for A_i(t) itself, so even the drift decomposition is shaky there. The i.i.d. simulation does satisfy the assumptions, so the theorem itself is not wrong—just narrower than presented.\n\nOther issues are minor by comparison: no code, no error bars, and the baseline set omits the closest prior-metric competitor (effective age from [29]). Self-citation is minimal, and the literature coverage seems fair.\n\nWho gets value: anyone working on status updates or remote control scheduling, especially on context-aware or cost-aware freshness metrics. The metric alone is worth knowing.\n\nRecommendation: worth serious peer review. I would send it out, but the authors should be asked to either state the full independence conditions, prove the CartPole weight satisfies them, or explicitly separate the i.i.d. simulation (where the theorem holds) from the CartPole experiment (empirical only). Also request simulation artifacts.","headline":"A clean, useful metric—context-aware lapse—with a working Lyapunov policy for status updates; but the main theorem carries an unstated independence assumption that the CartPole experiment probably violates, so the headline evidence needs qualification.","tokens_in":9970,"tokens_out":1657,"would_cite":true,"duration_ms":19452,"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":"Scheduling status updates by a context-weighted error metric, not by age, improves remote control and provably bounds average lapse.","keywords":["context-aware information lapse","Age of Information","remote control","status updates","user scheduling","Lyapunov optimization","CartPole"],"falsifier":"Run the policy of Eq. (6) in a simulation where $\\omega_i(t+1)$ is deliberately set to a function of $Q_i(t)$ or $A_i(t)$, for example $\\omega_i(t+1)=|Q_i(t)A_i(t)|$, with the same channel and increment statistics; if the average lapse still respects the Theorem 1 bound, the independence assumption is not load-bearing, and if it does not, the theorem's scope is exactly the independent case.","tokens_in":8934,"feed_emoji":"📡","tokens_out":6929,"duration_ms":66979,"temperature":0.7,"pith_summary":"This paper argues that Age of Information (AoI), which counts only the time since the newest update, misses what actually matters in remote control: how wrong the controller's picture is and how much that error matters right now. It proposes a metric it calls context-aware information lapse, $F(t)=\\omega(t)\\delta(Q(t))$, the product of a context weight and an error-based lapse, and studies minimizing its average over a multi-user wireless system. The central result is a Lyapunov scheduling policy, policy (6), that schedules $K$ users according to a context- and error-dependent index and comes with an upper bound on the long-run average context-aware lapse. The paper also reports that this policy lowers threshold-violation probability to about one fiftieth of the AoI-based policy's value and extends the average CartPole episode in a simulated remote-control benchmark.","feed_headline":"Scheduling by context-weighted error beats AoI in remote control","feed_subtitle":"A new metric multiplies estimation error by urgency; threshold violations drop to about 1/50 of an AoI-based policy.","key_machinery":"The central object is the product $F(t)=\\omega(t)\\delta(Q(t))$ with squared error; $Q(t)$ is the difference between true and monitor-known status, and $\\omega(t)$ is a context weight expressing urgency. The argument is carried by a Lyapunov drift-plus-penalty decomposition: with the penalty chosen as $f(t)=\\sum_i \\omega_i(t+1)(Q_i(t+1)^2-A_i(t)^2)$, minimizing drift plus penalty at each slot reduces to an index rule that uses the predicted next-slot weight $\\omega_i(t+1)$ alongside the current error. The one-step lookahead is what lets context information enter the decision.","core_discovery":"The discovery is that a single product metric, context-aware information lapse, can absorb both the stochastic evolution of the source and the context-dependent importance of its status, and that minimizing it in a multi-user system reduces to an index rule. Under squared error, the paper defines $Q_i(t)$ as the monitor's estimation error, $A_i(t)$ as its per-slot increment, and $F_i(t)=\\omega_i(t)Q_i(t)^2$ as the lapse. It proves (Theorem 1) that the policy scheduling $K$ users with the largest $(\\mathbb{E}[\\omega_i](1/(p_i\\pi_i)-1)+\\omega_i(t+1))p_iQ_i(t)^2$ achieves average lapse at most $\\sum_i \\mathbb{E}[\\omega_i]/(p_i\\pi_i)\\cdot\\mathbb{E}[A_i^2]$, and it chooses the randomized stationary schedule $\\pi$ by a water-filling solution to minimize the bound. Simulations show the resulting policy outperforms AoI-based scheduling in average lapse, threshold-violation probability, and CartPole control.","pith_inferences":["The paper's bound relies on the independence of $\\omega_i(t+1)$ from $Q_i(t)$ and $A_i(t)$; in the CartPole experiment the weight is derived from predicted motion driven by the same random force that creates error, so a direct test of whether correlation breaks the bound would sharpen the scope of the result.","The same index structure should extend to vector-valued or higher-order error lapses as long as the lapse is quadratic; a non-quadratic $\\delta$ would require a new drift-penalty choice rather than the simple square used in Theorem 1.","A practical consequence left implicit is that the value of context prediction grows with how predictable the context weight is; if $\\omega(t+1)$ is nearly deterministic given current observations, the lookahead term adds little and a context-unaware lapse policy may suffice."],"forward_implications":["If the bound in Theorem 1 holds, a scheduler that knows next-slot context and current error can guarantee a finite average context-aware lapse without knowing the full distribution of the error process beyond its second moment.","When the error increment is identically one and the weight is constant, the context-aware lapse reduces to the conventional AoI, so AoI-based scheduling becomes a special case of the proposed rule.","The context-aware policy cuts threshold-violation probability by about a factor of fifty compared with the AoI-based policy in the simulated setting, even though the scheduling problem was not designed to minimize violations.","In the CartPole remote-control experiment, the context-aware lapse scheduler lets the controller keep poles balanced for substantially more steps than schedules based on AoI or round robin."],"supporting_citations":[{"why":"Defines Age of Information, the baseline timeliness metric the paper argues is insufficient.","marker":"[1]"},{"why":"Introduces non-linear age penalties, which the context-aware lapse generalizes when weights are constant.","marker":"[6]"},{"why":"Supplies the AoI-optimal scheduling baseline against which the proposed policy is compared.","marker":"[20]"},{"why":"Shows that mean squared error can be expressed as a function of AoI when the source is not observable, motivating the need for an error-based metric when it is observable.","marker":"[27]"},{"why":"Proposes an effective-age concept that tracks estimation error; the paper's lapse differs by adding context weights.","marker":"[29]"},{"why":"Defines the CartPole learning-control problem used to evaluate remote-control performance.","marker":"[30]"}],"fun_headline_variants":["Context-aware lapse scheduling beats AoI in remote control","Error-weighted urgency metric outperforms AoI for status updates","New lapse metric: 50x fewer violations than AoI in remote control","Ditch AoI, use context-aware lapse for smarter remote control","Scheduling by context-weighted error improves remote control over AoI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes the next-slot context weight is statistically independent of the current error and of the error increment; the CartPole demonstration defines the weight from predicted motion driven by the same random force that creates the error, so the demonstration does not strictly satisfy the theorem's assumption.","fun_headline_variants_meta":{"raw":{"variants":["Context-aware lapse scheduling beats AoI in remote control","Error-weighted urgency metric outperforms AoI for status updates","New lapse metric: 50x fewer violations than AoI in remote control","Ditch AoI, use context-aware lapse for smarter remote control","Scheduling by context-weighted error improves remote control over AoI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000845,"raw_usage":{"total_tokens":3683,"prompt_tokens":952,"completion_tokens":2731,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":2644}},"tokens_in":568,"tokens_out":2731,"duration_ms":20715,"temperature":1.0,"reasoning_tokens":2644,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:43:05.403707+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the policy of Eq. (6) in a simulation where $\\omega_i(t+1)$ is deliberately set to a function of $Q_i(t)$ or $A_i(t)$, for example $\\omega_i(t+1)=|Q_i(t)A_i(t)|$, with the same channel and increment statistics; if the average lapse still respects the Theorem 1 bound, the independence assumption is not load-bearing, and if it does not, the theorem's scope is exactly the independent case.","supporting_citations":[{"cited_title":"Minimizing age of infor- mation in vehicular networks,","cited_arxiv_id":null,"evidence_quote":"Defines Age of Information, the baseline timeliness metric the paper argues is insufficient."},{"cited_title":"Age and value of information: Non-linear age case,","cited_arxiv_id":null,"evidence_quote":"Introduces non-linear age penalties, which the context-aware lapse generalizes when weights are constant."},{"cited_title":"A unified sampling and scheduling approach for status update in multiaccess wireless,","cited_arxiv_id":null,"evidence_quote":"Supplies the AoI-optimal scheduling baseline against which the proposed policy is compared."},{"cited_title":"Towards an “effective age","cited_arxiv_id":null,"evidence_quote":"Proposes an effective-age concept that tracks estimation error; the paper's lapse differs by adding context weights."},{"cited_title":"Neuronlike adaptive elements that can solve difficult learning control problems,","cited_arxiv_id":null,"evidence_quote":"Defines the CartPole learning-control problem used to evaluate remote-control performance."}],"review_version":1}