{"id":"5827f315-5ad6-4386-9d39-753a7116996c","arxiv_id":"2505.02462","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"iPFL combines graph-based personalized federated learning with a game-theoretic payment rule, proving non-negative utility and one-sided truthfulness, and showing higher economic utility in simulated markets.","lead":"This paper proposes iPFL, a personalized federated learning market where clients buy and sell access to each other's model updates through a payment rule designed to reward contributions and deter dishonest reporting. If the mechanism works, it offers a path to use decentralized private data for AI training without exposing raw data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Finite overreporting of N_i can pass Algorithm 2 and extract payments far above true marginal benefit, so Theorem 1's individual rationality fails for honest buyers; Theorem 3 only excludes N_i tending to infinity.","rationale":"The paper's central contribution is the pair of incentive properties, individual rationality and truthfulness, stated in the abstract and Section 4.3. Both guarantees rely on the server's ability to evaluate gains and payments from self-reported quantities: K_i, c_i, and N_i. The visible text proves truthfulness only for overreporting c_i (Theorem 2, Lemma 1) and robustness only for the limiting case N_i -> +infinity (Theorem 3). It does not analyze finite misreporting of N_i or K_i, even though these are private reports with no stated verification mechanism. The concrete two-client example above shows that a finite N_i overreport can pass Algorithm 2's selection rule and, through Eq. (10), transfer a payment from an honest buyer that is far larger than the buyer's true gain, making the buyer's utility negative while the liar profits. This contradicts the unconditional wording of Theorem 1 and undermines the 'inclusive market' claim, since a malicious seller can exploit the payment rule rather than being isolated. The empirical section does not address this scenario: Table 2 tests a liar who varies reported data size or cost and observes the liar's own utility drop, but that does not examine a seller who overreports N_i to extract excessive payments from buyers. Independent support is limited because the appendix is unavailable and no code is provided. For these reasons, the central claim as stated is not merely unproven; it is false under a natural manipulation of a reported type. I recommend REJECT unless the authors add an explicit verifiability assumption on N_i/K_i, prove that no finite misreport can increase utility or harm others, or substantially weaken the advertised incentive guarantees.","tokens_in":15550,"tokens_out":15514,"duration_ms":210982,"concrete_test":"Run a two-client simulation of Algorithm 2 and Eq. (10) with buyer gain G_b(n) = sqrt(1/100) - sqrt(1/(100+n)), seller true N_s = 1 reporting N_s = 1000, c_s = 0.05, and lambda * d = 0. Compute NTh_s from G_b(NTh_s) - G_b(NTh_s - 1000) = 0.05. If NTh_s > 1000, the seller is imported, and the buyer's true utility is G_b(1) - [G_b(1000) - G_b(0)] ≈ -0.0694. If this reproduces, Theorem 1's individual rationality guarantee is violated under finite N_i overreporting.","verdict_should_be":"REJECT","load_bearing_attack":"Section 4.3's Theorem 3 only excludes N_i -> +infinity, but Algorithm 2's threshold computation and Eq. (10)'s payment rule use the reported N_j values. A finite overreport of N_i can therefore be profitable and can make an honest buyer's utility negative. Concretely, consider a buyer with G_b(n) = sqrt(1/100) - sqrt(1/(100+n)) and a seller with true N_s = 1, c_s = 0.05, and identical models (d = 0, lambda = 0). If the seller reports N_s = 1000, Algorithm 2's threshold equation G_b(NTh_s) - G_b(NTh_s - 1000) = 0.05 has a solution NTh_s ≈ 1055 > 1000, so the buyer imports the seller. Eq. (10) then charges r_bs = G_b(1000) - G_b(0) = sqrt(1/100) - sqrt(1/1100) ≈ 0.0699, while the true marginal gain from one actual data point is sqrt(1/100) - sqrt(1/101) ≈ 0.0005. The buyer's per-round utility is approximately -0.0694 despite the buyer reporting honestly and uploading no poisoned model, and the seller nets about +0.0199 after paying c_s. Thus the unconditional individual rationality statement of Theorem 1 fails as soon as a participant can misreport N_i, and because N_i is a self-report, this is a strategic option, not just an adversarial extreme. Theorems 1-2 only establish safety under honest reports of N_i and K_i; they do not establish the claimed incentive properties for the full reported-type space.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes iPFL, a personalized federated learning framework coupled with a model-sharing market. Clients report their data size N_i, sharing cost c_i, and data eagerness K_i; a neutral server learns a collaboration graph (Algorithm 2) from model similarities and reported quantities, and charges payments defined by Eq. (10) as marginal collaboration gain minus model-difference cost. The paper claims that iPFL satisfies individual rationality and truthfulness, is robust to malicious model/data-size reports, and achieves higher economic utility than several PFL baselines across classification and instruction-tuning tasks.","tokens_in":15922,"tokens_out":6237,"duration_ms":79935,"significance":"The idea of combining personalized federated learning with a graphical game and a payment mechanism is genuinely interesting, and the experimental scope is broad. However, the central theoretical claims are not supported as stated: the individual-rationality guarantee fails under finite overreporting of N_i, the truthfulness theorem covers only c_i, and the economic-utility comparison is partly circular because the same utility function is both the objective optimized by iPFL and the evaluation metric. The proofs of Theorems 1-3 are deferred to an appendix that is not present. If the mechanism could be repaired to handle strategic reports of N_i, the framework would be a useful contribution; with the current text, the headline incentive guarantees are falsified.","major_comments":[{"comment":"The unconditional individual-rationality statement is false when a participant overreports N_i by a finite factor. Concrete counterexample: a buyer has N_b=100, K_b=1, so G_b(x)=sqrt(1/100)-sqrt(1/(100+x)); a seller has true N_s=1, c_s=0.05, and d=0, lambda=0. If the seller reports N_s=1000, Algorithm 2's threshold equation G_b(T)-G_b(T-1000)=0.05 has a solution T around 1055, which exceeds 1000, so the buyer imports the seller's model. Eq. (10) then charges r_bs=G_b(1000)-G_b(0)=sqrt(1/100)-sqrt(1/1100)≈0.0699, while the true marginal gain from the seller's actual one data point is G_b(1)-G_b(0)≈0.0005. The honest buyer's utility is therefore about -0.0694, violating Theorem 1, and the seller gains about +0.0199. Because N_i is a self-report in Algorithm 1, finite overreporting is a strategic option, not merely an adversarial extreme; Theorem 3 only excludes N_i to infinity and does not cover this case. The theorem needs an explicit honesty assumption or, more importantly, the mechanism needs a payment rule that is robust to finite misreports of N_i.","section":"Sec. 4.3, Theorem 1 (with Algorithm 2 and Eq. (10))"},{"comment":"The economic-utility comparison is circular. The utility function in Definition 1, Eq. (1)-(4), is the social-welfare term SW(A) that iPFL explicitly optimizes in Eq. (4), and the same utility is then used in Section 2.1 as the evaluation metric for 'economic utility'. Consequently, iPFL is being evaluated on the objective it is designed to maximize, while the baselines optimize different training objectives and do not use this payment/utility model at all. The finding that iPFL 'consistently achieves the highest economic utility' is therefore partly by construction and does not by itself demonstrate a superior market outcome. A fair comparison would evaluate all methods under a common, externally specified utility or welfare metric that is not the training objective of one of the methods.","section":"Sec. 2.1 and Eq. (4)"},{"comment":"The text states that the proofs of Theorems 1-3 are in the Appendix, but the manuscript as provided contains no appendix. Since these theorems carry the paper's central incentive and robustness claims, the absence of the proofs is a blocking omission: the reader cannot verify the conditions under which individual rationality, truthfulness, and robustness actually hold. Even without the counterexample above, the theoretical contribution is unverifiable in the submitted text.","section":"Sec. 4.3, proofs of Theorems 1-3"},{"comment":"The truthfulness guarantee is narrower than the abstract's claim. Theorem 2 only addresses overreporting of c_i, while the paper itself acknowledges in Section 4.3 that clients can exaggerate N_i to extract extra payment. The counterexample in my first comment shows that the mechanism is not truthful in N_i for finite overreports, and Section 2.2's Table 2 only tests lying about data size and cost at a few ratios in one setting, not a strategic equilibrium analysis over the full report space. The abstract's statement that iPFL adheres to 'truthfulness' therefore overstates the proven and demonstrated property.","section":"Sec. 4.3, Theorem 2 and the abstract's truthfulness claim"}],"minor_comments":[{"comment":"The payment r_ij can be negative when the model-difference penalty lambda*(N_j/N_i)*d(theta_i,theta_j) exceeds the marginal collaboration gain, which would mean the seller pays the buyer for importing its model. The paper does not discuss negative transfers, whether they are allowed, or how they affect the individual-rationality argument.","section":"Eq. (10)"},{"comment":"The pseudocode does not specify how to break ties when multiple clients have the same maximal threshold NTh_j, and it does not define what happens if the threshold equation has no solution for every j; the statement 'Set NTh_j = 0' covers the former but the loop's arg max over zero-valued thresholds is still underspecified.","section":"Algorithm 2"},{"comment":"The rows for lying on cost show utility exactly 0.00 for lying ratios 2, 5, and 10. It would be helpful to clarify whether this is because the liar is no longer selected as a collaborator, a boundary effect of the payment rule, or an artifact of the experimental setting; the current table does not explain why utility is exactly zero rather than merely lower.","section":"Sec. 2.2, Table 2"},{"comment":"The robustness experiments appear to use a single run per attack type, with no reported variance. Given the claim that only iPFL reduces the attacker's utility while preserving benign clients' accuracy, some measure of variability across multiple runs or seeds would strengthen the claim.","section":"Sec. 2.2, Fig. 4"},{"comment":"The horizontal-axis labels such as '50 0 50 100' appear to be misformatted tick labels; the reader cannot tell whether utilities are negative or positive without reading the figure more carefully. Rewriting the axis labels and adding a legend would improve clarity.","section":"Sec. 2.1, Fig. 2 labels"}],"recommendation":"reject","confidential_remarks":"The manuscript's central individual-rationality guarantee is falsified by a concrete finite-overreport counterexample within the model's own report space, and the truthfulness and robustness theorems do not cover the strategy that the counterexample exploits. The economic-utility evaluation is also circular because the utility metric is the method's training objective. Unless the mechanism is redesigned so that payments are robust to finite misreports of N_i, or the claims are substantially narrowed, I do not see a way to repair the contribution within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this paper is a creative attempt to build a market for model sharing in personalized federated learning, and the authors deserve credit for tackling a real problem. But the main theoretical guarantee—individual rationality—is false as stated. A seller can over-report its data size by a finite amount, still get selected by Algorithm 2, and extract payments far above any real benefit to the buyer. So the abstract's claim that 'every participant benefits' holds only under the hidden assumption that everyone reports N_i honestly, and Theorem 3 only handles the infinite case.\n\nWhat is actually new: the per-edge payment rule r_ij = G_i(a_i) - G_i(a_i - e_j) - λ(N_j/N_i)d(θ_i, θ_j) is a sensible way to charge for marginal contribution minus model dissimilarity, and integrating social welfare into the training objective is a neat idea. The experiments span eleven tasks, including LLM instruction tuning, which is impressive scope. The market simulation with buyers, sellers, traders, and attackers is a nice illustration of the intended ecosystem.\n\nWhere it gets soft. First, the proofs for Theorems 1–3 are in an appendix that is not included in the reviewed text, so they cannot be checked. Worse, a concrete counterexample shows the problem is not just missing proof. Take a buyer with G(n) = sqrt(1/100) - sqrt(1/(100+n)), a seller with true N=1, c=0.05, and identical models (λ=0, d=0). If the seller reports N=1000, Algorithm 2's threshold equation has a solution NTh ≈ 1055 > 1000, so the buyer imports the seller. The payment is G(1000) - G(0) ≈ 0.07, while the true marginal gain from one data point is about 0.0005. The buyer's per-round utility is approximately -0.069, and the seller nets +0.02. Since N_i is self-reported, this is a strategic option, not just an adversarial extreme. Theorem 1, as written, is therefore false over the full reported-type space.\n\nSecond, truthfulness is only one-sided: the paper proves something about overstating c_i, and explicitly says understating c_i can lead to selling at a low price. The abstract's unqualified 'truthfulness' is overclaiming. Third, the evaluation is partly circular: the economic utility used as the evaluation metric is exactly the social welfare term in the training objective. Of course iPFL, which optimizes that quantity, beats baselines that do not. A fairer comparison would include other incentive mechanisms. Also, many accuracy comparisons lack error bars.\n\nCredit where due: the paper is clearly written, the problem is well-motivated, and the empirical effort is real. The framework is worth discussing, and the flaw is fixable: clarify that the theorems assume honest reports of N_i, or redesign the payment rule to be robust to finite over-reporting.\n\nRecommendation: this deserves peer review, but a serious referee should insist on seeing the appendix, a corrected theorem with explicit assumptions, and a less circular evaluation. I would not cite the theorems as they stand, but I would engage with the framework and the question it raises.","headline":"Interesting mechanism, but the central individual-rationality theorem fails as stated under finite over-reporting of data size, which is a self-reported variable.","tokens_in":16445,"tokens_out":4478,"would_cite":false,"duration_ms":51234,"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":"A market for AI models where every participant gains and liars lose","keywords":["personalized federated learning","incentive mechanism","model sharing market","individual rationality","truthfulness","collaboration graph","game theory","federated learning"],"falsifier":"A direct test is to simulate a client whose true gain from importing a collaborator's data is much larger than the concave volume-based $G_i$ predicts, for instance a small high-quality dataset that exactly answers the buyer's task; if the buyer's true gain exceeds $G_i$, the payment $r_{ij}$ undercharges the seller and the seller's utility can become negative, breaking individual rationality. The paper's own Table 2 already varies reported data size and cost; running the same table with a lying report of $K_i$ would settle whether the truthfulness guarantee extends beyond $c_i$, since the theorem does not cover $K_i$.","tokens_in":15309,"feed_emoji":"🤝","tokens_out":13661,"duration_ms":128208,"temperature":0.7,"pith_summary":"The paper tries to establish that the privacy and incentive barriers to pooling decentralized private data can be overcome by turning federated learning into a model-sharing market: participants trade model parameters rather than raw data, and a neutral server prices every collaboration. It claims that its iPFL mechanism guarantees individual rationality, meaning no client ends a round with negative utility, and truthfulness, meaning over-reporting private cost cannot improve a client's utility, while matching or beating standard personalized federated learning on model accuracy. The payoff, if the claims hold, is that institutions with different purposes—pure buyers, pure sellers, traders, and even attackers—can coexist in one training loop, with prices set from the buyer's marginal gain and the seller's model difference. The guarantees rest on a specific concave gain function and a threshold-based graph-learning rule, so they hold within that model of how clients value data.","feed_headline":"A market for AI models where every participant gains and liars lose","feed_subtitle":"By pricing each model exchange from its marginal gain, iPFL reports non-negative utility for every client across 11 tasks.","key_machinery":"The load-bearing object is the pair of a directed collaboration graph $A$ and a concave collaboration gain $G_i(a_i)=\\sqrt{K_i/N_i}-\\sqrt{K_i/(N_i+\\sum_j a_{ij}N_j)}$, where $K_i$ is client $i$'s eagerness for data and $N_i$ its data volume. The mechanism's working part is a greedy threshold rule: for each potential collaborator $j$, the server computes a threshold $N_j^{\\mathrm{Th}}$ by solving $g_i(N_j^{\\mathrm{Th}})-g_i(N_j^{\\mathrm{Th}}-N_j)=c_j+\\lambda (N_j/N_i)d(\\theta_i,\\theta_j)$, and client $i$ imports the collaborator with the largest threshold until the accumulated data reaches it. The payment for an import is the marginal gain minus the model-difference penalty, so the same graph serves as both the training topology and the billing ledger; that coupling is what makes the individual-rationality and truthfulness proofs go through.","core_discovery":"The central claim is that economic incentives and personalized model training can be driven by the same object: a directed collaboration graph $A$ in which $a_{ij}=1$ means client $i$ imports client $j$'s model, with each row chosen to balance model similarity, sharing cost, and a concave data-gain $G_i(a_i)=\\sqrt{K_i/N_i}-\\sqrt{K_i/(N_i+\\sum_j a_{ij}N_j)}$. The server learns that graph with a greedy threshold rule and then bills each client the difference between the marginal benefit of each imported model and the model-distance penalty. Under that rule the paper proves that every client's utility is non-negative in every round (Theorem 1), that over-reporting one's cost $c_i$ cannot raise utility (Theorem 2), and that a client reporting an infinite data amount is never imported (Theorem 3). Empirically, across nine classification settings and two instruction-tuning scenarios, iPFL reports the highest economic utility with accuracy comparable to or better than FedAvg, FedProx, Ditto, FedAMP, CFL, FedFomo, and pFedGraph, including a 49% higher economic utility and 9% higher model utility than the best baseline in the LLM instruction-tuning case.","pith_inferences":["The paper's truthfulness theorem is stated for over-reporting the cost $c_i$; its experiments also test lying on data size, but a formal truthfulness guarantee for $K_i$ and $N_i$ is not derived, so the honesty result is cost-specific as proved.","Because the gain function counts only total imported data volume, the mechanism implicitly rewards quantity over content quality; a testable extension is to weight each collaborator's data by quality and check whether the greedy threshold rule still runs in $O(m)$ time.","The Discussion concedes the federation is assumed static; a natural extension is to add refund or escrow rules so buyers who exit before round $T$ are not charged for model value they never received."],"forward_implications":["If the claims hold, institutions with purely financial motives and institutions with purely accuracy motives can share one federation without one subsidizing the other.","A server that does not see any client's raw data can still set prices that leave every client with non-negative utility each round.","Over-reporting private cost becomes self-defeating, so the mechanism gives clients a reason to reveal their true unwillingness to share.","Attackers who upload poisoned or infinite-data models are structurally isolated through the model-difference and threshold rules rather than by a separate verification step.","The same mechanism extends to instruction-following language-model training, where the reported gains are 49% in economic utility and 9% in model utility over the best baseline."],"supporting_citations":[{"why":"prior work on incentivizing data sharing in federated learning that the market formulation extends.","marker":"[25]"},{"why":"supplies the graph-based personalized federated learning approach whose collaboration-graph idea iPFL adapts and which serves as a baseline.","marker":"[29]"},{"why":"FedAvg baseline that sets the standard comparison for federated model performance.","marker":"[44]"},{"why":"FedProx baseline for heterogeneous federated optimization against which iPFL reports comparable accuracy.","marker":"[45]"},{"why":"Ditto baseline for personalized and robust federated learning used in accuracy, utility, and attack comparisons.","marker":"[30]"},{"why":"CFL baseline for clustered federated learning used in robustness and classification comparisons.","marker":"[32]"},{"why":"FedFomo baseline for personalized federated learning used in the accuracy-utility trade-off.","marker":"[46]"},{"why":"source of the individual-rationality requirement that Theorem 1 is designed to satisfy.","marker":"[52]"}],"fun_headline_variants":["Truthful model sharing market for AI","Everyone gains in this AI model market","Incentivizing honest AI model contributions","A fair market for personalized federated learning","Guaranteed payoffs for model sharing clients"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a client's true value for imported data is fully captured by the concave function $G_i(a_i)=\\sqrt{K_i/N_i}-\\sqrt{K_i/(N_i+\\sum_j a_{ij}N_j)}$ with a single reported scalar $K_i$, and that the only strategic degrees of freedom are the reports of cost $c_i$ and data size $N_i$; if institutions value content rather than just data volume, or can profitably misreport $K_i$ or $N_i$, the threshold selection and payment rule can lose their individual-rationality and truthfulness guarantees.","fun_headline_variants_meta":{"raw":{"variants":["Truthful model sharing market for AI","Everyone gains in this AI model market","Incentivizing honest AI model contributions","A fair market for personalized federated learning","Guaranteed payoffs for model sharing clients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000607,"raw_usage":{"total_tokens":2852,"prompt_tokens":989,"completion_tokens":1863,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":1798}},"tokens_in":605,"tokens_out":1863,"duration_ms":16319,"temperature":1.0,"reasoning_tokens":1798,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:53:25.534409+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to simulate a client whose true gain from importing a collaborator's data is much larger than the concave volume-based $G_i$ predicts, for instance a small high-quality dataset that exactly answers the buyer's task; if the buyer's true gain exceeds $G_i$, the payment $r_{ij}$ undercharges the seller and the seller's utility can become negative, breaking individual rationality. The paper's own Table 2 already varies reported data size and cost; running the same table with a lying report of $K_i$ would settle whether the truthfulness guarantee extends beyond $c_i$, since the theorem does not cover $K_i$.","supporting_citations":[{"cited_title":"Personalized federated learning with inferred collaboration graphs","cited_arxiv_id":null,"evidence_quote":"supplies the graph-based personalized federated learning approach whose collaboration-graph idea iPFL adapts and which serves as a baseline."},{"cited_title":"Clustered federated learning: Model-agnostic distributed multitask optimization under privacy constraints","cited_arxiv_id":null,"evidence_quote":"CFL baseline for clustered federated learning used in robustness and classification comparisons."},{"cited_title":"Incentive mechanism for reliable federated learning: A joint optimization approach to combining reputation and contract theory","cited_arxiv_id":null,"evidence_quote":"source of the individual-rationality requirement that Theorem 1 is designed to satisfy."}],"review_version":1}