{"id":"fc2b89b1-4afc-483e-ae69-369091ec8b77","arxiv_id":"2505.05842","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A Bayesian-persuasion and dynamic-pricing mechanism for online federated learning under two-sided incomplete information, with a claimed 2ξ approximate optimality gap.","lead":"DaringFed is a proposed incentive mechanism for online federated learning where the server and clients each hide resource information from the other. It combines Bayesian persuasion signals with dynamic pricing to set client rewards, and the authors report faster convergence and higher accuracy in experiments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's 2ξ gap proof reverses the monotonicity comparison and drops the survival term, so the paper's central approximate-optimality guarantee is unproven.","rationale":"The reader's verdict of REJECT is supported. I focused on Theorem 3 rather than Theorem 1 because the 2ξ approximate-optimality bound is the paper's strongest quantitative claim and the proof of that bound contains a concrete invalid inequality chain (Eq. 30). The monotonicity error is not a matter of missing regularity; it reverses the relevant comparison and ignores the fact that larger rewards increase, not decrease, the survival probability. A counterexample in the paper's own experimental parameterization would directly falsify the bound. I do not manufacture an objection: the proof as written genuinely does not establish Theorem 3, and no independent support (machine-checked proofs, code, or a parameter-free derivation) is present. The reader's identified weakness in Theorem 1 is also real and material, which is why agreement is only partial; it is, however, secondary to the failure of the central optimality-gap theorem.","tokens_in":15882,"tokens_out":6270,"duration_ms":64182,"concrete_test":"Run the exact synthetic setup of Section 6 (c=(1−θ)^2(1.2−μ)^2, s(θ)=1−((θ−0.1)/0.8)^8, ξ=0.01, θ,γ∈[0.1,0.9]) and compute two quantities: C* = min over a fine continuous grid of γ and all Bayes-plausible posterior means of cs(γ,ρ), and C+ = min over the discrete grid R={0.1,0.11,...,0.9} using the paper's ρ+ construction from Eq. (17). If C+ − C* exceeds 0.02, Theorem 3 is contradicted. Because the invalid sign in Eq. (30) is the only step producing the bound, a single such counterexample settles the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is Theorem 3's bound cs(γ+,ρ+) − cs(γ*,ρ*) ≤ 2ξ. In the Appendix proof, the authors state that higher reward leads to a lower computation-resource threshold and conclude s(γ*,μ) ≥ s(γ+,μ). This inequality is backwards: for γ+ ≥ γ*, the threshold θhat = min{θ : c(θ,μ) ≤ γ} satisfies θhat(γ+) ≤ θhat(γ*), and because s is non-increasing in θ, s(γ+,μ) ≥ s(γ*,μ). With the correct direction, the display in Eq. (30) cannot be closed: the term γ*[∫∫ρ+s(γ+,μ) − ∫∫ρ*s(γ*,μ)] is nonnegative, not nonpositive, and no Lipschitz or continuity bound on s in γ is provided, nor is any relation between ρ+ and ρ* established. The gap can therefore grow with the participation probability, not stay O(ξ). Because the 2ξ guarantee is the headline result supporting approximate optimality under two-sided incomplete information, the core argument is unsupported. The reader's two-point-prior objection to Theorem 1 is also valid, but it concerns the optimal-signal derivation; the Theorem 3 failure would invalidate the bound even if Theorem 1 were repaired.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies incentive design for online federated learning under two-sided incomplete information: arriving clients know only a signal about the server's communication resource, and the server does not know the clients' computation-resource distribution. The proposed DaringFed mechanism combines a Bayesian-persuasion signal rule with a dynamic pricing rule; the server estimates the survival function of client computation resources by a UCB procedure and searches a discretized reward and signal space. The theoretical claims are a unique Bayesian-persuasion Nash equilibrium (Lemma 1), an optimal signal and reward for one-sided incomplete information (Theorem 1), an approximate signal for the two-sided case (Theorem 2), and a 2ξ gap between the approximate and optimal server cost (Theorem 3). Experiments on MNIST, Fashion-MNIST, FEMNIST, and CIFAR-10 compare DaringFed with three ablations, and synthetic experiments study convergence of the threshold and reward.","tokens_in":16158,"tokens_out":11104,"duration_ms":115933,"significance":"The modeling choice is relevant: both sides having private, dynamically changing resource information is a real obstacle for federated-learning incentives, and framing the interaction as Bayesian persuasion with a bandit-based estimation of the client-type distribution is a plausible design. The paper also provides a UCB estimator and an algorithmic instantiation in Algorithm 1. However, the core theoretical contributions are not currently supported: the equilibrium definition has reversed inequalities for a cost-minimizing server, the proof of uniqueness does not establish uniqueness, the optimal-signal derivation implicitly assumes a two-point prior, and the proof of the 2ξ bound contains a direction error that prevents the inequality chain from closing. Because these issues are load-bearing for the claimed approximate-optimality guarantee, the significance of the contribution cannot be assessed until the theory is substantially repaired.","major_comments":[{"comment":"The inequalities in the BPNE definition are reversed relative to the paper's own objective. Equation (4) and Eq. (5) define the server's problem as minimizing cs, so a Nash equilibrium for the server should satisfy cs(γ*,ρ*) ≤ cs(γ*,ρ) and cs(γ*,ρ*) ≤ cs(γ,ρ*) for unilateral deviations. As written, Definition 5 requires cs(γ*,ρ*) ≥ cs(γ*,ρ) and cs(γ*,ρ*) ≥ cs(γ,ρ*), which would make the equilibrium the server's worst unilateral outcome rather than a best response. This sign error undermines the meaning of Lemma 1 and the subsequent equilibrium analysis.","section":"Section 4.2, Definition 5 (Eq. (14))"},{"comment":"The proof argues that, for fixed γ, a concavification argument gives an optimal signal, and that for fixed ρ monotonicity and continuity give an optimal reward. This establishes existence of optimal responses separately, not uniqueness of a joint equilibrium (γ*,ρ*) under the two unilateral-deviation conditions in Definition 5. No argument shows that the optimal signal is unique or that the intersection of the two best-response sets is a singleton. The claim that there exists a unique BPNE is therefore unsupported by the proof.","section":"Appendix, Proof of Lemma 1"},{"comment":"The derivation of the optimal signal ρ(μ|τ) uses only the prior masses λ(τ) and λ(τ̄) at the two endpoints of [τ,τ̄]; Eq. (23) is exactly the Bayesian-consistency equation for a prior supported on {τ,τ̄}. The theorem, however, is stated for a general prior over τ ∈ [τ,τ̄], and no two-point-support assumption is given in Section 5.1. If the prior has interior mass, the displayed formula in Eq. (15) does not follow and the claimed optimal signal rule is not established. This is load-bearing because Theorem 1 is the basis for the one-sided optimal design and for the structure of the approximate signals in Theorem 2.","section":"Appendix, Proof of Theorem 1 (Eqs. (23)-(24))"},{"comment":"The inequality chain has a direction error. The proof states that a higher reward leads to a lower computation-resource threshold and concludes s(γ*,μ) ≥ s(γ+,μ). For γ+ ≥ γ*, the threshold θhat(γ+) = min{θ : c(θ,μ) ≤ γ+} is no larger than θhat(γ*), and because s is non-increasing in θ, the correct inequality is s(γ+,μ) ≥ s(γ*,μ). With the correct direction, the term γ*[∫∫ρ+s(γ+,μ)dμdτ − ∫∫ρ*s(γ*,μ)dμdτ] in the final step of Eq. (30) is nonnegative unless a relation between ρ+ and ρ* makes it negative; no such relation, and no Lipschitz or continuity bound on s in γ, is provided. The chain therefore cannot be closed, and the claimed bound cs(γ+,ρ+) − cs(γ*,ρ*) ≤ 2ξ is unproven.","section":"Appendix, Proof of Theorem 3 (Eq. (30))"},{"comment":"Both proofs begin by assuming γ*+ξ ≤ γ+ ≤ γ*+2ξ for the discrete reward γ+ chosen from a grid with spacing ξ. This condition is generally false: for any real γ*, a grid of spacing ξ contains a grid point within ξ of γ*, and if γ* is itself a grid point the nearest point is γ+. The asserted lower bound γ*+ξ is not guaranteed unless one defines γ+ as the first grid point strictly above γ*+ξ, whose existence is not established for a finite grid. Consequently, the 2ξ bound is not a consequence of the discretization in the way the proof claims, and the central approximate-optimality guarantee is unsupported.","section":"Appendix, Proofs of Theorems 2 and 3"},{"comment":"The stated objective is to minimize the expected server cost cs = γ·Pr(participation), as in Eq. (4) and Eq. (13), but BayesBen in Eq. (11) requires the signaling rule to satisfy E_ρ[ψ] ≥ E_λ[ψ], i.e., signaling cannot reduce the participation probability. For a cost-minimizing server, higher participation probability at a fixed positive reward is costly, so this constraint is not aligned with the stated objective. If participation is beneficial only through reaching the accuracy target, that benefit must be modeled explicitly; as written, the formulation is internally inconsistent.","section":"Section 3.3 and Definition 4"}],"minor_comments":[{"comment":"The definition of the signal rule uses an assignment arrow '←' instead of a functional equality; the expression ρ(σ|τ) ← φ(τ|σ)/λ(τ) should be written as an equality, and the densities should be defined on the same measurable space.","section":"Section 4.2, Definition 1, Eq. (6)"},{"comment":"The proof uses p(γ,μ) in the server cost expression, while the theorem statement and the rest of Section 5 use s(γ,μ); either p is a typo for s or it must be defined.","section":"Appendix, Proof of Theorem 1, Eq. (25)"},{"comment":"Equation (16) divides by N_t(θhat), which can be zero for an unselected threshold; the algorithm and the UCB bound need an explicit initialization or pseudo-count treatment.","section":"Lemma 2 and Algorithm 1"},{"comment":"The empirical section reports no error bars or multiple-seed statistics, and Table 1 gives final accuracy only; the claimed 16.99% improvement in accuracy and convergence speed is not broken down, and all baselines are ablations of DaringFed rather than existing incentive mechanisms.","section":"Section 6, Table 1 and Figure 2"}],"recommendation":"reject","confidential_remarks":"The manuscript has several foundational errors in the equilibrium and approximation proofs; even with a generous reading, the main theoretical results do not hold as stated. The reversal of the equilibrium inequalities in Definition 5, the hidden two-point-support assumption in Theorem 1, and the invalid inequality chain in Theorem 3 are independent, load-bearing problems. The empirical claims are also not fully supported by the reported tables. I recommend rejection rather than major revision because the central theoretical contribution would need to be re-derived, not merely patched."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: DaringFed packages two known ideas—Kamenica–Gentzkow signal design and UCB dynamic pricing—into a mechanism for online federated learning under two-sided incomplete information. That application is genuine: OFL with privately known client compute resources and server-controlled communication bandwidth is a real gap in the literature. The paper also ships a concrete algorithm and experiments on standard image datasets, and the problem framing is sensible.\n\nThe bad news is that the theoretical core does not hold up. Lemma 1 claims a unique Bayesian persuasion Nash equilibrium, but the proof only argues existence via concavification; there is no argument eliminating multiple equilibria. Theorem 1 derives the optimal signal as if the prior on communication resources has two-point support at the endpoints; that assumption is never stated in the theorem. More seriously, Theorem 3's 2ξ gap bound is derived through an inequality chain that cannot close. The authors write that higher reward yields a lower computation-resource threshold and conclude s(γ*,μ) ≥ s(γ+,μ). That is backwards: if γ+ ≥ γ*, the threshold is lower, and since s is non-increasing in the threshold, s(γ+,μ) ≥ s(γ*,μ). With the corrected direction, the middle term in Eq. (30) is nonnegative, not nonpositive, and the claimed O(ξ) gap is not established. The concern is not a missing cosmetic step; the bound is the paper's main selling point for approximate optimality under two-sided uncertainty.\n\nThere is also a sense in which the 2ξ bound is less impressive than it looks: with a grid of size ξ, any search over the grid is within O(ξ) of the best grid point by construction. The real work is showing that the grid point is close to the continuous optimum, and the paper does not do that.\n\nExperiments are serviceable but thin: no error bars, no baselines from the FL-incentive literature, only ablations against stripped-down versions of DaringFed. The 16.99% accuracy improvement is over those ablations, not over a non-DaringFed baseline.\n\nThat said, the paper is not a throwaway. The problem is real, the mechanism is a reasonable engineering heuristic, and the UCB estimation is genuinely learned from data. If the authors repair or soften Theorem 3 and add serious baselines, a revised version could be of interest.\n\nFor peer review: I'd send it out, but to a rigorous theory referee who will check the inequalities. The paper deserves engagement; the current version should not be accepted as is.","headline":"A plausible pairing of Bayesian persuasion and UCB pricing for OFL, but the central 2ξ bound and the uniqueness claim are unsupported; the paper needs a serious technical referee.","tokens_in":16681,"tokens_out":2893,"would_cite":false,"duration_ms":29356,"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":"DaringFed shows that an online federated learning server can set near-optimal rewards without knowing client computation resources or revealing its own bandwidth, by combining Bayesian persuasion signals with bandit-style dynamic pricing.","keywords":["Bayesian persuasion","online federated learning","incentive mechanism","two-sided incomplete information","dynamic pricing","upper confidence bound","Nash equilibrium","signal design"],"falsifier":"Choose a prior distribution over communication resources that has substantial mass strictly inside the interval, compute the server's cost under the signaling rule of Theorem 1, and compare it with the true minimum over all feasible signal distributions obtained by brute-force search or convex optimization; if the two differ beyond the discretization gap, the theorem's optimality claim for general priors is false.","tokens_in":15658,"feed_emoji":"🎯","tokens_out":7894,"duration_ms":73216,"temperature":0.7,"pith_summary":"This paper tries to show that a server in online federated learning can set rewards for arriving clients without knowing those clients' computation power, and without clients knowing the bandwidth the server allocates. It models the exchange as a Bayesian persuasion game in which the server sends signals about bandwidth and sets a price, and it claims this game has a unique Bayesian persuasion Nash equilibrium. For the one-sided version the paper derives an explicit optimal signaling rule, and for the two-sided version it shows that a discretized, learnable version of the mechanism stays within a fixed additive gap of the optimal server cost. The paper reports that on four real image datasets the mechanism improves model accuracy and convergence speed, and synthetic tests show the learned reward and computation threshold converge.","feed_headline":"Signals and bandit pricing keep federated learning costs near-optimal","feed_subtitle":"A dynamic Bayesian persuasion rule bounds the server's reward loss by 2ξ even when client resources are unknown.","key_machinery":"The central object is the DaringFed mechanism, a 2-tuple $(S,P)$: a Bayesian persuasion signal rule $S$ that maps a communication-resource value $\\tau$ to a distribution over posterior means $\\mu$, and a dynamic pricing rule $P$ that sets the reward $\\gamma\\leftarrow \\rho(\\sigma|\\tau)\\theta$. The signal distribution must satisfy three constraints (Bayesian consistency, Bayesian plausibility, and Bayesian benefit), and the pricing side estimates the client survival function $s(\\theta)$ with an upper-confidence-bound estimate whose confidence term is $\\sqrt{\\ln N/(2N_t(\\hat{\\theta}))}$. Algorithm 1 searches the discretized reward space and resource-threshold space, uses Theorem 2's four-signal formulas to build the signal distribution, and Theorem 3's bound $c_s(\\gamma^+,\\rho^+)-c_s(\\gamma^*,\\rho^*)\\le 2\\xi$ certifies the gap between the approximate and optimal server cost. The optimal signal rule itself is the ratio formula $\\rho(\\mu|\\tau)=\\rho(\\mu)(\\tau-\\mu)/(\\lambda(\\tau)(\\tau-\\underline{\\tau}))$ and its mirror at the upper endpoint, which is what lets the server influence client beliefs without revealing the actual bandwidth.","core_discovery":"The central claim is that DaringFed, a two-part mechanism built from a Bayesian persuasion signal rule and a dynamic pricing rule, approximately solves the server's cost-minimization problem in online federated learning under two-sided incomplete information. The signal rule lets the server choose a posterior distribution over communication resources for each arriving client, subject to Bayesian consistency, plausibility, and benefit constraints, so that clients who lack bandwidth information update their beliefs in a way favorable to the server. The pricing rule estimates the unknown client computation-resource distribution via an upper-confidence-bound estimate of the survival function and then picks the reward and threshold on a discrete grid. The paper proves a unique Bayesian persuasion Nash equilibrium exists, gives the optimal signal formula for the one-sided case, and proves that the approximate solution found over the grid differs from the true optimum by at most twice the grid step. Empirically, the mechanism filters out low-resource clients and improves accuracy and convergence speed on four datasets, with the estimated reward and computation threshold converging in simulations.","pith_inferences":["Beyond the paper, a natural fix for the endpoint-only gap is to discretize the prior or to use the general convex-hull signal characterization from Bayesian persuasion; that would extend Theorem 1 to arbitrary priors.","Beyond the paper, replacing the Hoeffding-style confidence term with a variance-aware bandit bound could shrink the practical gap below $2\\xi$ or reduce the number of rounds needed to converge, which the paper does not test.","Beyond the paper, the reported 16.99% accuracy gain depends on the four chosen datasets and the fixed cost model; stress tests with heavy-tailed resource distributions or non-stationary client populations would show whether the $2\\xi$ bound remains the limiting factor."],"forward_implications":["The server can set rewards online without knowing the client computation-resource distribution; the upper-confidence-bound estimator supplies the missing information and the total loss is at most $2\\xi$.","Clients are only shown a posterior signal about bandwidth and a reward, so the server's actual communication-resource allocation stays private.","Because the game has a unique Bayesian persuasion Nash equilibrium, neither side can improve its own payoff by unilateral changes to signal, reward, or participation rule.","The $\\xi$ discretization directly trades optimality against computation: a smaller grid step gives closer-to-optimal cost at a larger search cost.","Filtering out low-resource clients through the reward threshold is what improves model accuracy and convergence speed in the real-data experiments, with a reported 16.99% improvement."],"supporting_citations":[{"why":"Supplies the Bayesian persuasion model and the Bayesian plausibility and benefit conditions that define feasible signals.","marker":"[Kamenica and Gentzkow, 2011]"},{"why":"Supplies the dynamic pricing with Bayesian persuasion setup and the confidence-bound technique for learning an unknown distribution online.","marker":"[Agrawal et al., 2024]"},{"why":"Supplies the client cost model and the Bayesian-game treatment of incomplete information in federated edge learning, plus the survival-function assumption.","marker":"[Hu et al., 2022]"},{"why":"Defines the online federated learning setup of sequential client arrival and immediate aggregation that the mechanism is built on.","marker":"[Chen et al., 2020]"},{"why":"Provides the convexity facts used in the equilibrium proof and in deriving the form of the optimal signal.","marker":"[Boyd, 2004]"},{"why":"Grounds the concavity assumption on the survival function that the existence proof relies on.","marker":"[Bergstrom et al., 1986]"}],"fun_headline_variants":["Persuasion pricing for federated learning with unknown resources","Bayesian persuasion meets bandits in online federated learning","Near-optimal pricing for federated learning under uncertainty","DaringFed: cutting federated learning costs with dynamic pricing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of the optimal signal drops all interior values of the communication-resource distribution and uses only its two endpoints, so the claimed general formula for the optimal signal depends on an unstated assumption that the prior has no mass in between.","fun_headline_variants_meta":{"raw":{"variants":["Persuasion pricing for federated learning with unknown resources","Bayesian persuasion meets bandits in online federated learning","Near-optimal pricing for federated learning under uncertainty","DaringFed: cutting federated learning costs with dynamic pricing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1351,"prompt_tokens":1005,"completion_tokens":346,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":279}},"tokens_in":621,"tokens_out":346,"duration_ms":4200,"temperature":1.0,"reasoning_tokens":279,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:54:23.373326+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a prior distribution over communication resources that has substantial mass strictly inside the interval, compute the server's cost under the signaling rule of Theorem 1, and compare it with the true minimum over all feasible signal distributions obtained by brute-force search or convex optimization; if the two differ beyond the discretization gap, the theorem's optimality claim for general priors is false.","supporting_citations":[{"cited_title":"Bayesian Persuasion.American Economic Review, 101(6):2590–2615,","cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian persuasion model and the Bayesian plausibility and benefit conditions that define feasible signals."},{"cited_title":"Dynamic Pricing and Learning with Bayesian Persuasion.Proceedings of the Advances in Neural In- formation Processing Systems, pages 1-9, New Orleans, USA, December","cited_arxiv_id":null,"evidence_quote":"Supplies the dynamic pricing with Bayesian persuasion setup and the confidence-bound technique for learning an unknown distribution online."},{"cited_title":"Asynchronous Online Federated Learning for Edge Devices with Non-IID Data.Proceed- ings of the IEEE International Conference on Big Data, pages 15–24, Atlanta, USA, December","cited_arxiv_id":null,"evidence_quote":"Defines the online federated learning setup of sequential client arrival and immediate aggregation that the mechanism is built on."},{"cited_title":"Convex Optimization","cited_arxiv_id":null,"evidence_quote":"Provides the convexity facts used in the equilibrium proof and in deriving the form of the optimal signal."}],"review_version":1}