{"id":"34a59734-a8f9-4398-8787-cd839f17f768","arxiv_id":"2508.19141","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For a shared channel where each node knows the value of its own data, the optimal distributed access rule is a threshold, and the authors provide algorithms that reach it.","lead":"This paper builds a mathematical model for many sensors sharing one wireless channel, where each sensor sends only when its data is valuable to a receiver. It shows the best rule is a simple threshold, and gives two algorithms that find it: one for nodes that know the statistics, one for nodes that must learn them.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BETA's quasi-exponential convergence is unproven: Alg. 2's unbiasedness is established only for stable strategies, precisely what learning violates.","rationale":"The paper's most valuable result is the threshold best-response characterization (Theorem 2), and I do not see a flaw there: from (5), the derivative w.r.t. x_{n,v} is p_{n,v}[ζ_n v − ψ − Σ_{m≠n} Σ_u u p_{m,u}x_{m,u} ζ_n/(1−x̄_m)], which yields the threshold in (9), and the KKT partition argument is sound. The subsequent claim that IBR reaches a Nash equilibrium is plausible at least for the finite discrete model. The soft spot I find is in the learning half. The paper itself concedes in §V-B that the counterfactual reward estimates are unbiased only when strategies are stable, but BETA's entire purpose is to learn thresholds, so strategies are not stable during the learning phase. The sentence 'In any case, the semi-bandit feedback respects the conditions for quasi-exponential convergence' is an unsupported leap. This matters because the abstract promises a distributed learning algorithm with guaranteed convergence, and the simulations cannot substitute for the missing unbiasedness argument. If, contrary to my reading, the estimates are unbiased even under adaptation (e.g., because the window W is short or a martingale argument applies), the concrete test above would reveal that. The temporal-independence assumption identified by the reader is a legitimate scope limitation, but it is explicitly acknowledged in Section VII and does not threaten the internal logic of the model. The Theorem 4 gap is internal to the paper's own proof and is therefore the more load-bearing concern. Overall, the threshold theorem and LIBRA results are substantial, so a conditional verdict remains appropriate; the learning half needs an additional proof or a weakened claim.","tokens_in":16147,"tokens_out":15928,"duration_ms":184102,"concrete_test":"Use a small instance (e.g., N=2 or 3, binary VoI, ψ=0.25) and exactly enumerate the true conditional expected reward E[R(θ, θ_{-n}^{(i)}) | past] for every counterfactual threshold θ at each BETA round i. Compare, over many Monte Carlo runs, the mean of Alg. 2's estimates with this true conditional expectation during the first ~20,000 steps while thresholds are still moving. If the mean estimation error is systematically nonzero in this nonstationary phase, the unbiasedness condition of [28] is violated and Theorem 4 cannot be invoked for BETA.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V-B's proof of Theorem 4 ends with: 'All these estimates are unbiased and with finite variance, provided that the strategies of nodes are stable. ... In any case, the semi-bandit feedback respects the conditions for quasi-exponential convergence.' The imported convergence guarantee from [28] requires an unbiased, finite-variance estimator at every learning round. But BETA is explicitly a learning algorithm: thresholds θ_n(i) change while the running averages ρ̄, ρ_n, α_n, β_n, λ_n are accumulating. The 'stable strategies' clause is exactly the condition that fails during the exploration/exploitation phase, and the final 'in any case' sentence is an assertion, not a proof. Without unbiasedness under nonstationary strategies, quasi-exponential convergence to the LIBRA threshold is not established. This does not invalidate Theorem 2, but it undercuts the second half of the central claim: that the problem reduces to learning one threshold per node with a distributed bandit algorithm.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formulates the Goal-oriented Multiple Access (GoMA) problem in a slotted collision channel with N sensor nodes. Each node observes a Value of Information (VoI) drawn from a known distribution (initially) and chooses a transmission probability for each possible VoI level. The main analytical results are: non-convexity of the resulting optimization problem (Lemma 1.1); a closed-form threshold characterization of each node's best response to the others' strategies (Theorem 2); an extension to continuous VoI domains (Theorem 3); and an iterated-best-response protocol, LIBRA, claimed to provably reach a Nash equilibrium. The paper then relaxes the distribution-knowledge assumption and proposes BETA, a distributed ε-Hedge algorithm with semi-bandit counterfactual estimation, claimed to converge quasi-exponentially to the LIBRA solution. Numerical experiments compare both protocols with a pull-based dominant-node scheduler and report gains in reward, energy consumption, and fairness. Reproducible simulation code is provided.","tokens_in":16341,"tokens_out":12531,"duration_ms":120580,"significance":"If the convergence claims are established, this is a valuable first theoretical treatment of distributed goal-oriented medium access. The threshold best-response theorem is clean, parameter-free, and self-contained; the non-convexity proof via Hessian minors is valid; and the paper gives a concrete algorithmic route from the centralized problem to a distributed, learning-based protocol. The empirical evaluation is extensive and the code availability is a strength. However, the theoretical convergence guarantee for BETA is not established under the nonstationary learning dynamics, and Theorem 1's proof relies on insufficiently justified imports. As written, the contribution is a strong problem formulation and a well-validated heuristic protocol rather than a fully proven theory of distributed GoMA learning. The paper is worth publishing after these load-bearing gaps are addressed.","major_comments":[{"comment":"The BETA convergence claim is not supported. The proof of Theorem 4 ends with: 'All these estimates are unbiased and with finite variance, provided that the strategies of nodes are stable.' This is exactly the condition that fails during learning: in BETA the thresholds θ_n(i) are updated continuously, and the running averages ρ̄, ρ_n, α_n, β_n, λ_n accumulate samples collected under different strategy profiles. The imported quasi-exponential convergence result of [28], used in Section V-A, requires an unbiased finite-variance estimator at every learning round. Unbiasedness under stable strategies does not imply unbiasedness with respect to E[R(θ_n(i); θ_−n(i))] while θ changes; the sentence 'In any case, the semi-bandit feedback respects the conditions for quasi-exponential convergence' is an assertion, not a proof. The theorem as stated ('there exists a finite-variance, unbiased estima","section":"Section V-B, Theorem 4 and Section V-A"},{"comment":"The proof of Theorem 1 is not sufficient as written. The game is a common-payoff game with continuous action spaces x_n ∈ [0,1]^{|V|}, so the cited finite best-response path result [23, Thm. 3] does not directly apply, and [24] concerns Markov potential games with independent natural policy gradient, not the static best-response dynamics considered here. The authors should give a direct proof for the finite set of threshold strategies, which is available from Theorem 2 (proved later), or otherwise justify the application of the cited results. Without this, the claim that IBR provably reaches an NE, or an ε-NE in polynomial time, is unsupported. This is load-bearing because LIBRA's optimality certificate rests on it.","section":"Section III, Theorem 1"}],"minor_comments":[{"comment":"The statement of Theorem 3 appears to have typos: the term ψ/θ_n should be ψ/ζ_n, and the integrand should use Q_m(p), not Q_n(p), consistently with Eq. (17) in the proof. Please correct the theorem statement to match the derivation.","section":"Section III-B, Eq. (15)"},{"comment":"There are small presentation issues: 'form which v ≥ θ∗_n' should be 'from which', and the expression E[R|θ∗_n, x_{n,θ∗_n}, x_−n] in Eq. (12) is not formatted clearly. Also, when ζ_n = 0 (some other node always transmits), the formula (9) is singular; the authors should explicitly state the trivial best response x_n = 0 in that case.","section":"Section III-A, proof of Theorem 2"},{"comment":"In the Hessian minor argument, the 2×2 minor is taken with respect to variables (n,v) and (m,u). The text should state explicitly that n ≠ m, since the Hessian is hollow and the off-diagonal formula (6) has factor 1−δ_{m,n}.","section":"Section III-A, Lemma 1.1"},{"comment":"The pseudocode layout of the Collision case is difficult to read, with formulas split across lines. The definitions of α_n, β_n, and λ_n should be stated clearly in the main text before Algorithm 2, and the window W used in the running averages should be integrated into the unbiasedness discussion.","section":"Section V-B, Algorithm 2"},{"comment":"The phrase 'almost all initial point on the left-hand side' and a few other grammatical errors should be corrected. Also, Figure 2 would benefit from marking the axes more clearly when the initial threshold is a quantile versus a VoI value.","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":"The BETA convergence gap identified in the stress-test note is real and is the main obstacle to acceptance. The threshold characterization and the empirical study are solid; I would encourage the authors to either add a genuine nonstationary convergence analysis or explicitly present BETA as an empirically validated heuristic. The paper fits the journal's scope and, with these fixes, would be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: Theorem 2 is solid and worth engaging. The GoMA model, the KKT derivation of the threshold best response, and the non-convexity proof via the 2x2 Hessian minor all check out. The threshold formula in (9) is parameter-free and derived directly from the stated model, and the simulation section matches the theory with Monte Carlo runs and a public code link. That is a genuine step forward for distributed goal-oriented medium access, and the LIBRA algorithm is a plausible application of the framework.\n\nThe soft spots are real but not fatal to the core theorem. The stress-test concern about BETA is correct. In Section V-B, the proof of Theorem 4 ends with the caveat that estimates are unbiased and finite-variance only when strategies are stable, followed by an unsupported 'in any case' sentence. BETA is a learning algorithm: thresholds change while the running averages accumulate, so the stable-strategy condition is exactly what fails during exploration. The quasi-exponential convergence guarantee from [28] is imported without a bridge over this nonstationarity. That does not touch Theorem 2, but it means the second half of the central claim—distributed bandit learning converging to the LIBRA equilibrium—is unproven as written.\n\nThere are two smaller issues. Theorem 1 leans on a natural-policy-gradient result for Markov potential games [24] without demonstrating how that applies to iterated best response in a static potential game; the connection needs spelling out. And the continuous-domain theorem statement appears to contain typos: the denominator in (15) should likely be ζ_n rather than θ_n, and the integral inside the sum should use Q_m, not Q_n (the proof in (17) is coherent). These are fixable, but they make that section harder to trust than it should be.\n\nThe paper also leans on the independence assumption across nodes and time slots, which is stated honestly in Section II and acknowledged as future work in Section VII. That is fine for a first systematic model, but it means the threshold optimality is proven only in the memoryless case.\n\nWho gets value: anyone working on semantic or effective communication, MAC design, or distributed decision-making under value-of-information models. The paper deserves a serious referee—the core theorem is good enough that the BETA convergence gap should be corrected rather than used to reject the whole thing. I would cite Theorem 2 and the LIBRA formulation, and I would send this to review with a request to fix the BETA proof and the Theorem 1 bridge.","headline":"The threshold best-response result is the real contribution; the BETA convergence guarantee does not follow from the supplied proof.","tokens_in":16818,"tokens_out":2143,"would_cite":true,"duration_ms":24576,"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":"This paper tries to establish that distributed goal-oriented medium access reduces to a per-node threshold on the value of information, with best-response and bandit learning reaching a locally optimal equilibrium.","keywords":["goal-oriented communication","value of information","medium access control","threshold strategies","iterated best response","multi-armed bandits","Nash equilibrium","wireless sensor networks"],"falsifier":"Simulate two or more nodes on a collision channel where each node's value of information is a first-order autoregressive process rather than independent over time, then compare the average reward of the threshold policy from Eq. (9) with an optimal history-dependent policy obtained, for instance, by dynamic programming or a learned policy with memory. If the history-dependent policy achieves strictly higher average reward, the central myopic-threshold claim fails in the correlated regime. A second check: in an i.i.d. three-node setting with asymmetric distributions, use exhaustive search to ve","tokens_in":16036,"feed_emoji":"📡","tokens_out":6577,"duration_ms":59319,"temperature":0.7,"pith_summary":"The paper aims to show that when several sensors share a collision channel, the right medium-access rule is not to elect a single dominant transmitter but to let each node transmit only when its measurement's value of information is high enough. It proves that, given the other nodes' strategies, each node's optimal response is a threshold policy, and that the resulting best-response dynamics converge to a locally optimal Nash equilibrium. On top of this it builds LIBRA, an iterative protocol that approaches such an equilibrium from a fair starting point, and BETA, a distributed bandit learner that reaches nearly the same performance without knowing the value distributions in advance. In simulations, the approach outperforms centralized pull-based scheduling by up to 100% in reward while using 20–30% less energy and spreading load more fairly.","feed_headline":"Distributed sensors can schedule themselves by information value","feed_subtitle":"A threshold rule plus best-response learning beats central polling by up to 100%, with lower energy use.","key_machinery":"The load-bearing object is the threshold best-response map of Theorem 2 (and its continuous counterpart Theorem 3). It turns a per-node optimization over a continuum of transmission probabilities into a single number: the threshold theta*_n, equal to the normalized expected value of other nodes' transmissions plus the transmission cost divided by the probability that all others are silent. This map makes the problem an exact potential game, so iterated best response converges; it also collapses the learning problem to one threshold per node, which is what BETA's bandit arms represent.","core_discovery":"The paper's central claim is that the joint problem of maximizing expected received value-of-information minus transmission cost over all nodes can be decoupled into independent threshold rules. Concretely, Theorem 2 states that for fixed strategies of the other nodes, node n's optimal transmission policy is to transmit only if its observed value v exceeds theta*_n = sum_{m≠n} sum_u u p_{m,u} x_{m,u}/(1−x̄_m) + ψ/ζ_n, where ψ is the transmission cost and ζ_n is the probability that all other nodes stay silent. Because the cooperative game is an exact potential game, iterating these best responses yields a local optimum, and the closed-form threshold makes each iteration cheap. The paper also","pith_inferences":["Editorial inference: the threshold formula contains a natural 'load price' — the normalized expected value contributed by other nodes — which could be broadcast as a single scalar so nodes tune their thresholds without knowing each other's distributions; BETA's aggregate feedback is a step in this direction.","Editorial inference: under temporally correlated value of information (the paper's stated open problem), the memoryless threshold is unlikely to remain optimal; the testable conjecture is that optimal policies become state-dependent, for example thresholds that fall after missed transmissions.","Editorial inference: this result is the push-based mirror of index-based scheduling rules for age-of-information pull systems; showing that the threshold ordering coincides with the index ordering would unify the two approaches."],"forward_implications":["Distributed goal-oriented medium access can be solved by threshold rules computed per node, with no central scheduler deciding transmission order.","LIBRA provably reaches a local optimum (ε-Nash equilibrium) and, in the paper's simulations, beats the best single-dominant-node pull-based scheme by up to 100% in reward while cutting energy use by roughly 20–30%.","BETA reaches about 99% of LIBRA's reward using only limited feedback and no prior knowledge of the value-of-information distributions, converging in roughly 20 000 slots.","LIBRA remains within a few percent of its ideal performance even when nodes have imperfect estimates of other nodes' value distributions.","The threshold formulation applies to both discrete and continuous value domains, so the same protocol logic covers sensors reporting from finite alarm sets or from continuous measurement spaces."],"supporting_citations":[{"why":"Supplies the potential-game property that guarantees iterated best response has a convergence target.","marker":"[21]"},{"why":"Provides the finite best-response path theorem used in the proof of Theorem 1.","marker":"[23]"},{"why":"Used to claim polynomial-time convergence of the best-response dynamics to an ε-Nash equilibrium.","marker":"[24]"},{"why":"The Karush-Kuhn-Tucker conditions are the proof tool for the threshold characterization of the best response in Theorem 2.","marker":"[25]"},{"why":"Foundation for the multi-agent semi-bandit formulation that BETA builds on.","marker":"[20]"},{"why":"Proves quasi-exponential convergence of ε-Hedge with unbiased finite-variance semi-bandit estimates, the convergence engine behind BETA.","marker":"[28]"},{"why":"Supplies the ε-Hedge algorithm whose softmax exploration-exploitation rule BETA executes.","marker":"[29]"}],"fun_headline_variants":["Self-scheduling sensors pick winners by data value","Smart transmitters: value-aware threshold bests central control","Goal-oriented access: nodes learn when to talk, beat hub","Distributed GoMA: threshold rule gives 2x gain over central","Sensors decide by info worth, not fixed slots"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Everything rests on the assumption that each node's value of information in a slot is statistically independent of earlier slots and of other nodes' values, so a node's best action depends only on the current slot's expected reward; if values are correlated over time, a myopic threshold need not be optimal.","fun_headline_variants_meta":{"raw":{"variants":["Self-scheduling sensors pick winners by data value","Smart transmitters: value-aware threshold bests central control","Goal-oriented access: nodes learn when to talk, beat hub","Distributed GoMA: threshold rule gives 2x gain over central","Sensors decide by info worth, not fixed slots"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000959,"raw_usage":{"total_tokens":3919,"prompt_tokens":739,"completion_tokens":3180,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":3098}},"tokens_in":483,"tokens_out":3180,"duration_ms":21135,"temperature":1.0,"reasoning_tokens":3098,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:55:57.650330+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate two or more nodes on a collision channel where each node's value of information is a first-order autoregressive process rather than independent over time, then compare the average reward of the threshold policy from Eq. (9) with an optimal history-dependent policy obtained, for instance, by dynamic programming or a learned policy with memory. If the history-dependent policy achieves strictly higher average reward, the central myopic-threshold claim fails in the correlated regime. A second check: in an i.i.d. three-node setting with asymmetric distributions, use exhaustive search to ve","supporting_citations":[{"cited_title":"Potential games,","cited_arxiv_id":null,"evidence_quote":"Supplies the potential-game property that guarantees iterated best response has a convergence target."},{"cited_title":"Congestion games with player-specific payoff functions,","cited_arxiv_id":null,"evidence_quote":"Provides the finite best-response path theorem used in the proof of Theorem 1."},{"cited_title":"Provably fast convergence of independent natural policy gradient for Markov potential games,","cited_arxiv_id":null,"evidence_quote":"Used to claim polynomial-time convergence of the best-response dynamics to an ε-Nash equilibrium."},{"cited_title":"Nonlinear programming,","cited_arxiv_id":null,"evidence_quote":"The Karush-Kuhn-Tucker conditions are the proof tool for the threshold characterization of the best response in Theorem 2."},{"cited_title":"Non-stochastic bandit slate problems,","cited_arxiv_id":null,"evidence_quote":"Foundation for the multi-agent semi-bandit formulation that BETA builds on."},{"cited_title":"Learning with bandit feedback in potential games,","cited_arxiv_id":null,"evidence_quote":"Proves quasi-exponential convergence of ε-Hedge with unbiased finite-variance semi-bandit estimates, the convergence engine behind BETA."},{"cited_title":"A decision-theoretic generalization of on-line learning and an application to boosting,","cited_arxiv_id":null,"evidence_quote":"Supplies the ε-Hedge algorithm whose softmax exploration-exploitation rule BETA executes."}],"review_version":1}