{"id":"d90a55ab-ea19-4dbb-8d44-1dde20c1fae3","arxiv_id":"2507.02710","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A delegation-based voting mechanism, viscous-retained democracy, is proposed to select federated learning clients for weight transfer, with theoretical claims that it never performs worse than FedAvg in an idealized competence model.","lead":"This paper proposes using delegation-based voting, called fluid democracy, to pick which clients in federated learning transmit their model updates, reducing communication cost. It introduces a new variant, viscous-retained democracy, claimed to match or beat the standard FedAvg baseline under an idealized model of client competence.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2 is false as stated: a chain construction with n=m^2 voters, max guru weight sqrt(n)=o(n), and alpha=1-1/sqrt(n) makes viscous-retained accuracy tend to 1/2 while direct voting accuracy tends to Phi(1)≈0.84, violating Do No Harm.","rationale":"The reader's weakest_assumption concerned the mapping between the FL setting and the paper's binary epistemic social-choice model. That is a valid applicability concern, but the more load-bearing issue is that Theorem 2 is internally false under the paper's own model. The reader did flag the proof's 'unjustified assertion about path lengths' and rated the central claim salvageable; the construction above shows the assertion is not merely unjustified but materially false, and that the claimed Do No Harm property fails asymptotically with a constant gap in accuracy. This is decisive because Theorem 2 is the theoretical basis for the abstract's 'always does better than 1p1v' claim. The only experiment is a hand-built 40-node example, so no empirical evidence repairs the counterexample. The result could likely be repaired by adding a stronger assumption, such as bounded path length or maximum guru weight o(sqrt(n)), but that is not the theorem as stated and would weaken the advertised guarantee. Therefore the current submission should be rejected, or at minimum substantially revised, rather than conditionally accepted on the strength of the stated theorem.","tokens_in":9162,"tokens_out":29456,"duration_ms":356690,"concrete_test":"Fix n=m^2 for m in {64,128,256}. Build m chains of length m with p_j=0.5+j/m^2 and alpha=1-1/m. For each m, compute exactly or by Monte Carlo with at least 10^6 samples the probability that the final weighted sum of independent Bernoulli votes exceeds n/2, for both direct voting and viscous-retained voting. If direct accuracy stays near 0.84 while viscous-retained accuracy approaches 0.5 as m grows, Theorem 2's sufficient condition is falsified. As an analytic cross-check, compute the quantity in Lemma 4.1 condition (2): if sum_i w_i p_i - sum_i p_i is O(sqrt(n)) rather than Omega(n), the lemma cannot be applied, confirming that the proof's linear-improvement step fails.","verdict_should_be":"REJECT","load_bearing_attack":"The proof of Theorem 2 asserts that since no guru has O(n) delegators, the longest path length m is a constant. That implication is false: a forest of components of size sqrt(n) has longest path sqrt(n), and this is not just a gap. Let n=m^2 and partition voters into m chains of length m. Along each chain set p_j=0.5+j/m^2, so delegation is strictly upward, and set alpha=1-1/m. Each guru receives m-1 delegators, which is o(n), so the theorem's stated hypotheses hold. In the model of Section 3.4, each final voter casts one Bernoulli draw for its whole weight. Direct voting has mean n/2 + m/2 and standard deviation m/2, so accuracy tends to Phi(1)≈0.84. Under viscous-retained, each guru's weight is about (1-1/e)m, there are m gurus, and non-guru weights are O(1); hence the variance of the weighted sum is Theta(m^3)=Theta(n^{3/2}), so the standard deviation is Theta(n^{3/4}), while the mean excess over n/2 is only Theta(m)=Theta(sqrt(n)). The z-score therefore tends to 0 and the delegated accuracy tends to 1/2. For every epsilon<0.34, gain(P_VRD,P_Direct)<-epsilon for all large n, violating the Do No Harm property. This also refutes the abstract's claim that viscous-retained 'always does better than 1p1v.'","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper frames federated learning aggregation as an epistemic two-alternative social-choice problem in which each client has a binary competence qi and can delegate weight to more competent neighbors. It compares liquid democracy, viscous democracy, and a new 'viscous-retained democracy' against direct majority voting (identified with FedAvg), defines the Do No Harm (DNH) property, and claims that viscous-retained democracy satisfies DNH under the same assumptions as liquid democracy while preventing influence accumulation. It also proposes a threshold-based algorithm, FedVRD, and gives a small star/chain example to illustrate a cost-robustness trade-off. The central theoretical result, Theorem 2, is false as stated and the proof contains a false inference; the abstract's stronger claim that the protocol 'always does better than 1p1v' is not supported by the formal definitions.","tokens_in":9426,"tokens_out":18591,"duration_ms":221192,"significance":"The paper's high-level idea of using delegation rules as a client-selection mechanism for FL is interesting, and the viscous-retained rule is a natural variant that conserves total voting mass while bounding guru weight. The adversarial cost-budget discussion in Section 5 is also a useful concrete direction. However, the paper's main contribution is a DNH theorem, and that theorem is invalid under the stated hypotheses. With the theorem removed, the manuscript offers an algorithm with no convergence or robustness guarantees and a small illustrative experiment whose figure is absent from the text. I therefore do not see a publishable core in the current submission.","major_comments":[{"comment":"The theorem is refuted by the following construction, which satisfies the stated hypotheses. Let n=m^2 and partition the voters into m disjoint chains of length m. In each chain, with voters indexed from the tail to the guru, set competence p_j = 1/2 + j/m^2 and let every non-guru delegate to the next voter; each guru then receives exactly m-1 delegators, which is o(n). Take α=1-1/m; the statement does not forbid α from depending on the graph size. For direct voting, the total mean excess over n/2 is Θ(m) and the standard deviation is Θ(m), so direct accuracy tends to Φ(1)≈0.841. Under viscous-retained democracy, the guru in each chain has weight (1-α^m)/(1-α)=Θ(m), each non-guru has weight at most 1, and total vote mass is n; the weighted sum therefore has mean excess Θ(m) and standard deviation Θ(m^{3/2}). The z-score tends to 0, so the delegated accuracy tends to 1/2. Hence gain(P_VRD,P_Direct) tends to -(Φ(1)-1/2)<0, violating DNH. The proof's claim that 'Since no guru has O(n) delegators, m is a constant' is false: this construction has max guru indegree m-1=o(n) and longest path length m=√n. If the authors intended α to be fixed independent of n, they must state that restriction explicitly and supply a different argument; as written, the theorem and proof are incorrect.","section":"Section 4.4 (Theorem 2)"},{"comment":"The abstract's claim that viscous-retained democracy 'always does better than 1p1v' is not a consequence of the Do No Harm property defined in Section 4.1. DNH only requires gain(M,Gn) ≥ -ε for all sufficiently large n, i.e., asymptotic no-harm, not strict improvement. Section 4.1 itself states this and then immediately draws the converse conclusion ('always performing better than direct democracy would imply performing better than FedAvg'); the implication is not established. Moreover, the formal results live entirely in the two-alternative epistemic model of Section 3.1, where 'correct' means drawing an update from data subset a+. No theorem connects binary vote correctness to the actual FedAvg objective or to the cosine-similarity heuristic in Algorithm 1, line 12, so the title's claim about improving FL aggregation is unsupported even if the social-choice theorem were correct.","section":"Abstract and Section 4.1"},{"comment":"Both proofs rely on a nonexistent uniform competence gap. The assumption that no two voters have equal competencies (Assumption 4.2) does not imply the existence of a constant cmin>0 that is independent of n; the minimum gap can tend to zero along a sequence of graphs. Similarly, in Theorem 2, from 'no guru has O(n) delegators' the proof concludes that the longest path length m is constant; even in a version where α is fixed, m can grow and α^m does not provide a uniform lower bound. Without a uniform positive lower bound on the per-delegation competence improvement, the claimed total competence increase of order n in Equation (2) is not established.","section":"Section 4.2 and Section 4.4 (proofs of Lemma 4.3 and Theorem 2)"}],"minor_comments":[{"comment":"The text refers to Figure 1 for the cost-robustness trade-off, but no figure appears in the manuscript, so the claimed plot cannot be checked.","section":"Section 5"},{"comment":"The threshold τ is introduced with τ ∈ (0,1), but the example uses τ=1 and calls it standard viscous democracy; the relationship between τ and the vote-share threshold is not formalized.","section":"Section 5"},{"comment":"The letter C(n) is used both for an o(n) bound on maximum weight in Lemma 4.1 and for an O(n) bound on the number of edges in Assumption 4.2, which is confusing.","section":"Section 4.1 and Assumption 4.2"},{"comment":"The proof of Theorem 2 writes α^m > 0 without treating the case α=0; the case of no delegation should be handled separately.","section":"Section 4.4"}],"recommendation":"reject","confidential_remarks":"Given that the central theorem is false under a natural reading and the proof contains a false implication, I recommend rejection. If the authors can prove a corrected theorem with α fixed independently of n and a uniform competence-gap assumption, and if they remove the 'always better' overclaim, a resubmission could be considered. The workshop acceptance note in the header is not relevant to my assessment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: there is a real idea here — a viscous-retained rule that keeps total vote mass constant — but the central theorem is false as stated, and the abstract's 'always does better than 1p1v' is unsupported. I would send it to a serious referee, not because it is near acceptance, but because the variant is worth a close look and the main error is clean enough for a referee to catch.\n\nWhat is genuinely new: the retention rule, beta=1 with alpha damping, is a natural modification of viscous democracy. Keeping the total vote mass constant is exactly what removes the mechanism by which viscous democracy loses competence along delegation paths, so the intuition is good. The paper also correctly identifies that failure mode in Theorem 1. The application to FL as a client-selection/aggregation step is a reasonable extension, and the fluid democracy literature is cited accurately. I credit the paper for stating the do-no-harm property and its assumptions explicitly.\n\nThe soft spot is load-bearing. Theorem 2's proof asserts that because no guru has O(n) delegators, the longest delegation path m is constant. That implication is false: a forest of sqrt(n) chains of length sqrt(n) has max guru weight sqrt(n)=o(n) but path length sqrt(n). The stress-test construction realizes this. With alpha=1-1/sqrt(n) and competencies increasing along each chain, viscous-retained accuracy tends to 1/2 while direct voting tends to Phi(1), so the mechanism violates do no harm. Small perturbations handle the 'no equal competencies' assumption without changing the asymptotics. So this is not just a gap in the proof; the theorem is false under the stated assumptions.\n\nThe FL side is also thinner than the theory. The mapping from a two-alternative social choice model to 'correct' model updates is a stretch, no real dataset is used, the experiment is a hand-built 40-node example, and FedVRD's threshold selection is left unspecified. That part is early-stage.\n\nWho is this for? People working on fluid democracy variants may get a useful idea from the viscous-retained rule; people looking for a working FL protocol should wait. I would not cite it in current form, but the counterexample is worth discussing. If I were an editor, I would send it to review with a clear request to check the path-length step; as submitted, it should not be accepted.","headline":"The viscous-retained variant is a genuinely sensible idea, but Theorem 2 is false as stated and the abstract's 'always better than 1p1v' claim does not survive the paper's own model.","tokens_in":10043,"tokens_out":7330,"would_cite":false,"duration_ms":87759,"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 introduces viscous-retained democracy, a delegation protocol for federated learning that provably never performs asymptotically worse than direct one-person-one-vote averaging (FedAvg) under the same assumptions that make…","keywords":["federated learning","fluid democracy","viscous democracy","liquid democracy","do no harm property","client selection","adversarial robustness","FedVRD"],"falsifier":"Run the Theorem 1 construction—n/2 chains of length 10 with competences in (0.65,0.7) and n/2 stars of size 10 with center competence 0.4 and leaves 0.39—using viscous-retained weights with α=0.5, and compute the weighted-mean competence after delegation; if it falls below the direct-voting mean of 0.5 for some finite n, the do-no-harm claim is false.","tokens_in":8871,"feed_emoji":"🗳️","tokens_out":7767,"duration_ms":82116,"temperature":0.7,"pith_summary":"Federated learning typically asks every client to send its model weights to a central server every round, which wastes communication on clients whose data contributes little. This paper proposes to instead let clients delegate their 'vote' to a more competent neighbor for a few rounds, so the server only aggregates a subset of clients' weights. The central theoretical claim is that a new rule, viscous-retained democracy, never asymptotically underperforms direct one-person-one-vote voting (equated with FedAvg) under the same mild conditions that make liquid democracy safe, while still damping influence accumulation. The key is that each delegator keeps a fraction of their vote instead of losing it to decay, which prevents high-competence voters in long delegation chains from being drowned out by low-competence voters in star-shaped clusters. On top of this, the paper's FedVRD algorithm tunes a threshold to trade communication cost against resistance to adversarial clients.","feed_headline":"Delegation rule never underperforms direct voting in federated learning","feed_subtitle":"Viscous-retained democracy keeps vote mass constant, so competence can't leak away through delegation chains.","key_machinery":"The central object is the vote-weight rule of viscous-retained democracy: a voter who delegates to a more competent neighbor passes a fraction α of their weight along each hop but retains the complementary fraction 1−α, so the total vote mass in the system stays constant even though guru weights are dampened by α per hop as in viscous democracy. The retained fraction is what carries the argument: it guarantees that every delegation strictly increases the competence of the weight that travels, with a minimum per-delegation competence increase of c_min · α^m (where m is the bounded longest path), which is exactly what Lemma 4.1's condition (2) needs to prove the Do No Harm property.","core_discovery":"The paper's central claim is Theorem 2: in an upward-delegation process where each delegator passes a fraction α of their weight to a more competent neighbor and retains the complementary fraction 1−α, the resulting 'viscous-retained democracy' satisfies the do-no-harm property—its accuracy is asymptotically within ε of direct one-person-one-vote voting—provided no guru accumulates O(n) delegators and Assumption 4.2 holds (the delegation graph has at least C(n) ∈ O(n) edges and no two voters share a competence). This is contrasted with Theorem 1, which shows standard viscous democracy fails do-no-harm: in a network of length-10 chains of high-competence voters and size-10 stars of low-competence voters, the chain gurus' vote mass decays to at most 1/(1−α) while star gurus keep 1+9α, so with α=0.5 the weighted-mean competence falls from above 0.5 to at most 0.48 and the Condorcet Jury Theorem drives the collective decision to the wrong alternative. Viscous-retained democracy preserves the constant total vote mass and therefore a strict competence increase per delegation, which is exactly what Lemma 4.1 requires for do-no-harm.","pith_inferences":["Because the retained-vote rule keeps total vote mass invariant, the do-no-harm proof could likely be recast as a martingale argument, where each delegation is a mean-increasing contraction; that reformulation might relax the bounded-delegator assumption.","The paper's adversarial analysis assumes the adversary can place agents optimally but only in star/chain topologies; a testable extension is to derive the minimal adversary budget for arbitrary graphs using the Laplacian spectrum of the delegation graph.","If the binary-competence assumption fails in practice, viscous-retained democracy may still be a better heuristic than viscous democracy, because it cannot suffer the pure weight-loss failure; comparing FedVRD against FedAvg and trimmed-mean on non-IID benchmarks would test this."],"forward_implications":["Federated clients can delegate to a more competent neighbor via update similarity and the server can aggregate only the elected clients' weights, cutting per-round transfer cost without asymptotically sacrificing accuracy relative to FedAvg.","Viscous-retained democracy matches liquid democracy's do-no-harm guarantee while keeping guru weights identical to viscous democracy, so influence cannot accumulate in a few misaligned clients.","The threshold τ in FedVRD lets the server trade communication cost against adversarial robustness at each round; the paper's star-chain experiments show an optimal τ exists for a fixed budget.","Standard viscous democracy is unsuitable as an FL aggregation rule because its weight decay makes high-competence chains lose influence to low-competence stars, which is exactly the failure mode Theorem 1 constructs."],"supporting_citations":[{"why":"Defines the do-no-harm property and shows liquid democracy satisfies it under upward delegation, the baseline this paper extends.","marker":"Kahng et al., 2018"},{"why":"Provides Lemma 4.1, the sufficient conditions (bounded max guru weight and linear competence gain) that Theorem 2 verifies for viscous-retained democracy.","marker":"Berinsky et al., 2024"},{"why":"Introduces viscous democracy and the per-hop dampening factor α that viscous-retained democracy modifies by retaining 1−α per delegator.","marker":"Boldi et al., 2011"},{"why":"Supplies the comparison of viscous vs liquid democracy in upward delegation settings that motivates the topological-dependence diagnosis.","marker":"Armstrong et al., 2024"},{"why":"Defines FedAvg, the one-person-one-vote aggregation baseline against which do-no-harm is measured in the FL setting.","marker":"McMahan et al., 2017"}],"fun_headline_variants":["Viscous-retained democracy guarantees do-no-harm in federated voting","Delegation rule in federated learning never loses to direct voting","New fluid democracy protocol keeps federated accuracy intact","Federated clients can delegate without sacrificing model quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole guarantee rests on treating federated learning as a two-alternative binary-choice problem in which each client has a fixed probability of producing a 'correct' update, and clients can reliably identify a more competent neighbor from update-similarity scores; if that mapping does not hold, the Do No Harm theorem says nothing about whether the selected updates improve the global model versus FedAvg.","fun_headline_variants_meta":{"raw":{"variants":["Viscous-retained democracy guarantees do-no-harm in federated voting","Delegation rule in federated learning never loses to direct voting","New fluid democracy protocol keeps federated accuracy intact","Federated clients can delegate without sacrificing model quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000811,"raw_usage":{"total_tokens":3576,"prompt_tokens":986,"completion_tokens":2590,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":2521}},"tokens_in":602,"tokens_out":2590,"duration_ms":21554,"temperature":1.0,"reasoning_tokens":2521,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:23:51.785571+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the Theorem 1 construction—n/2 chains of length 10 with competences in (0.65,0.7) and n/2 stars of size 10 with center competence 0.4 and leaves 0.39—using viscous-retained weights with α=0.5, and compute the weighted-mean competence after delegation; if it falls below the direct-voting mean of 0.5 for some finite n, the do-no-harm claim is false.","supporting_citations":[{"cited_title":"Liquid democracy: An algorithmic perspective","cited_arxiv_id":null,"evidence_quote":"Defines the do-no-harm property and shows liquid democracy satisfies it under upward delegation, the baseline this paper extends."},{"cited_title":"Tracking Truth with Liquid Democracy","cited_arxiv_id":"2107.11868","evidence_quote":"Provides Lemma 4.1, the sufficient conditions (bounded max guru weight and linear competence gain) that Theorem 2 verifies for viscous-retained democracy."},{"cited_title":"Optimizing viscous democracy","cited_arxiv_id":null,"evidence_quote":"Supplies the comparison of viscous vs liquid democracy in upward delegation settings that motivates the topological-dependence diagnosis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines FedAvg, the one-person-one-vote aggregation baseline against which do-no-harm is measured in the FL setting."}],"review_version":1}