{"id":"061e486e-2f61-47fc-a2d9-b4e00204cc29","arxiv_id":"2509.09099","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"In a critical-mass Bayesian persuasion problem, extending a communication network can strictly increase the sender's value, sometimes up to the private-signaling upper bound.","lead":"This paper studies a sender who wants to persuade at least k of n receivers connected in a network, where each receiver sees the messages sent to its neighbors. The main finding is that adding communication links can strictly help the sender, sometimes reaching the same success as fully private messages.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 4 is false as stated: the proof's gcd claim is invalid, and for n=6,k=4,p=2 the cluster network already attains V_n^k, so no strict-benefit extension exists.","rationale":"I identify the same overall weakness that the reader flagged, but I focus on a different concrete failure. The reader's weakest_assumption was Lemma 3 (the imported sufficiency condition from Babichenko et al.), which is indeed load-bearing for many constructions. However, Lemma 3's failure would be an external validity issue; Proposition 4 contains a definite internal logical error that falsifies a stated main result. This is more directly verifiable and more immediately damaging to the paper's claims about cluster networks. The proof's divisibility argument incorrectly asserts that n/2<k<n forces gcd(n,k)=1, and the counterexample n=6,k=4,p=2 shows the strict improvement cannot hold because the original network already attains V_n^k. This does not collapse the entire paper—Theorem 1, Theorem 2, and Proposition 3 may still be correct—but it requires a repair (e.g., adding gcd(n,k)=1 or excluding cases where the cluster network already achieves the upper bound). The reader's verdict of CONDITIONAL is appropriate; I see no reason to change it, hence UNCHANGED. The concrete test I propose would settle whether the n=6,k=4,p=2 case is indeed a counterexample, which would confirm the need for revision.","tokens_in":66,"tokens_out":3990,"duration_ms":52016,"concrete_test":"Check the n=6, k=4, p=2 instance. Compute or construct an experiment on the network of three disjoint edges that yields value V_6^4 = (6/4+1)λ0 = 2.5λ0. For instance, in state X send a message making all receivers choose x; in state Y, with probability e = λ0(X)/(λ0(Y))·(6/4), send posterior-1/2 messages to four receivers in two pairs and posterior-0 messages to the third pair, otherwise send posterior-0 to all. Verify each receiver's posterior after observing their partner's message equals the intended posterior, and that the resulting value equals V_6^4. If the value equals V_6^4, Proposition 4's strict-benefit claim fails for this instance, confirming the error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 4 claims that for a network of q≥2 disjoint p-clusters with n/2<k<n, the sender's value is strictly below V_n^k and there exists an extension that strictly benefits the sender and achieves V_n^k. The proof's contradiction relies on the assertion that p|n and p|(n−k) imply p|k, and then 'Since n/2<k<n, we must have that n,k are relatively prime.' This is false: p can divide both n and k. Example: n=6, k=4, p=2 satisfies n/2<k<n, and p divides both n and n−k, hence p|k. The conclusion that gcd(n,k)=1 is wrong. The error is not merely a gap: for n=6,k=4, the cluster network is three disjoint 2-cliques. Because each receiver observes its partner's message, receivers within a pair always share the same posterior, so the number of receivers with posterior 1/2 (who choose x) must be even. Since k=4 is even, an experiment achieving the private upper bound V_6^4 is feasible: in state Y, send posterior-1/2 messages to two pairs (4 receivers) and posterior-0 messages to the third pair. Thus the cluster network already achieves V_n^k, contradicting the strict improvement claim of Proposition 4. This directly undermines the cluster-family part of the paper's central claim, even though other families may be unaffected.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a multi-receiver Bayesian persuasion problem with network spillovers: each receiver observes the sender's private message to himself and to his neighbors, and the sender wants a critical mass k of receivers to choose the sender-favorable action. The main claim is that, contrary to the intuition that more communication helps receivers, extending the network can strictly increase the sender's value, and in several families (stellar, halo, constellation, galaxy, cluster) the sender can reach the empty-network upper bound V_n^k. The analysis uses a symmetry lemma, a circle construction, and a sufficient condition for attaining V_n^k based on the absence of information-dominating pairs. Section 4 also gives examples where no beneficial extension exists.","tokens_in":32091,"tokens_out":9872,"duration_ms":111343,"significance":"If correct, the paper would make a valuable and counterintuitive contribution: it would show that the sender's persuasion value is non-monotone in network density, with concrete structural families where denser communication helps the sender and hurts receivers. The constructive experiments, especially the two-message circle construction, are a useful strength, and the transparent reliance on the information-domination lemma makes the proof strategy easy to follow. However, the cluster-family result (Proposition 4) is false as stated, which materially weakens the paper's claim to cover 'general families' of networks. The stellar and galaxy results may still be salvageable, but the current manuscript overstates the scope of its conclusions.","major_comments":[{"comment":"Proposition 4 is false as stated. The proof derives p|n and p|n-k, hence p|k, and then asserts that n/2<k<n forces gcd(n,k)=1. This is wrong: n=6, k=4, p=2 satisfies the hypotheses and gcd(6,4)=2. In that case, the network is three disjoint 2-cliques, and the sender can already achieve V_6^4. Send x to all receivers in state X; in state Y, with probability r = 6λ0/(4(1-λ0)), choose two of the three clusters uniformly and send x to those four receivers and y to the third cluster, and with probability 1-r send y to all. Each receiver in a targeted cluster has posterior 1/2, exactly four receivers choose x, and the value is (6/4+1)λ0 = V_6^4. Since V_n^k is the upper bound, no extension can strictly benefit. The proposition needs a divisibility condition such as p∤k, and the divisible case must be handled separately.","section":"Appendix B, Proof of Proposition 4; Proposition 4"},{"comment":"The paper's central positive results all conclude by invoking Lemma 3, imported from Babichenko et al. (2021), which asserts that a network with no information-dominating pairs attains V_n^k. Lemma 3 is therefore load-bearing. The manuscript should either prove it or provide a precise reference with all hypotheses verified. The Proposition 4 counterexample shows the importance of checking the lemma's precondition: if the original network already attains V_n^k, the lemma cannot be used to establish strict improvement. The current proofs do not independently verify this precondition in a way that rules out the counterexample.","section":"Lemma 3 and its uses in Theorem 1, Theorem 2, Proposition 3, Proposition 4"},{"comment":"Corollary 2 is stated without proof. It is not an immediate corollary of the displayed results, and it is used to extend the paper's nonmonotonicity claim to halo networks. The authors should provide a proof or a precise reduction to Theorem 1; otherwise the claim should be labeled as a conjecture or removed.","section":"Corollary 2 (Halo)"}],"minor_comments":[{"comment":"The set T_i is defined as {i, i+1, ..., i+n-1-k} (mod n), which has n-k elements, not k. To obtain the claimed experiment, T_i should have exactly k elements, e.g., {i, ..., i+k-1}. As written, the construction does not reach the critical mass in state Y.","section":"Lemma 2 proof"},{"comment":"The notation 'p|pq=n' is awkward; since n=pq, the authors mean p divides n. More importantly, the subsequent claim about relative primality is false, as noted above.","section":"Proposition 4 proof, line 'p|pq=n'"},{"comment":"The example sets λ0(X)=1/3, which equals k/(n+k) rather than satisfying the strict inequality assumed at the end of Section 2. The text acknowledges this, but it would be cleaner to choose a generic prior and adjust the probabilities.","section":"Example 2"}],"recommendation":"major_revision","confidential_remarks":"The Proposition 4 error is serious and should be fixed before publication. The paper also leans heavily on Lemma 3 from a preprint coauthored by one of the current authors; this is not circular, but it is a single point of failure, and the editorial process should ensure the cited result is correct and that all preconditions are checked. If the cluster claim is restricted appropriately, the remaining stellar/galaxy results may still justify publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this paper has a genuinely interesting core—adding links can strictly help the sender, so the sender's value is non-monotonic in network density—and it gives constructive experiments that hit the private-signaling upper bound on circle, stellar, and galaxy families. But one of the four main constructions, Proposition 4 (clusters), is false as stated, and the cluster part needs a real repair before the paper is publishable.\n\nWhat's new and good: the circle construction uses two messages and a simple uniform experiment; Theorem 1 and Theorem 2 cover hierarchical and decentralized topologies; the 9-client example in the introduction is effective. The authors are appropriately careful about the relation to Babichenko et al.: Lemma 3 is imported, not re-proved, and it does a lot of lifting, but it's cited explicitly and there is no circularity—the paper's own contribution is not assumed. Citation patterns look fine, including self-citations to prior work that actually proves the quoted results.\n\nSoft spots: Proposition 4's proof says p|n and p|(n−k) imply p|k, and then 'since n/2<k<n, n,k are relatively prime.' That is wrong. For n=6, k=4, p=2, gcd(6,4)=2. And the issue is not just a proof gap: three disjoint 2-cliques already achieve V_6^4. In state Y, put two pairs at posterior 1/2 and one pair at posterior 0; four receivers choose the sender's action, exactly the upper bound. So no extension can strictly benefit, contradicting Proposition 4's statement. The proposition needs a restriction (e.g., gcd(n,k)=1) or a corrected version. Corollary 2 (halo) is also stated without proof; maybe acceptable as a corollary, but it is a substantive result, so it should be justified or downgraded to a remark. Minor: Remark 2 claims the value cannot increase when no information-dominating pair is removed, but the paper doesn't use it; that's fine.\n\nI don't think the stress-test concern about Lemma 3 is fatal. If Babichenko et al.'s no-information-domination theorem failed, most of Section 4 would collapse, but that is true of any paper relying on a known theorem, and I have no reason to doubt it. The paper's central claim survives even without Proposition 4, since the circle and stellar results stand. But as written, the cluster family—advertised as a main decentralized family—is broken.\n\nWho this is for: people working on Bayesian persuasion with networks, information design for voting, and anyone interested in when transparency backfires. It deserves a serious referee, but the referee should require a corrected Proposition 4 and a proof of Corollary 2 before acceptance. I wouldn't cite the cluster result as it stands.","headline":"Novel and mostly convincing non-monotonicity results, but Proposition 4 (clusters) is false as stated and needs a real repair before the paper is publishable.","tokens_in":32571,"tokens_out":3360,"would_cite":false,"duration_ms":36099,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A28","91B12","91D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"In critical-mass Bayesian persuasion on networks, adding communication links among receivers can strictly increase a sender's payoff, and in several natural network families it reaches the maximum value achievable with private signaling.","keywords":["Bayesian persuasion","communication networks","critical mass","peer communication","information spillovers","information domination","non-monotonic value","voting"],"falsifier":"For a small instance of Theorem 1 (e.g., n=8, k=4, lambda0(X)=0.3, with a stellar component of six nodes of depth 2 and two external nodes), explicitly build the extension described in the proof and numerically optimize the sender's experiment over finite message sets; if the optimum is strictly below V_8^4 = 0.9, the central claim fails. More directly, any network with no information-dominating pairs whose optimal value is below V_n^k would refute the imported Lemma 3 and with it the paper's upper-bound constructions.","tokens_in":31622,"feed_emoji":"🗳️","tokens_out":7270,"duration_ms":76109,"temperature":0.7,"pith_summary":"The paper studies a strategic sender—a campaign, firm, or platform—who wants to push a critical mass of receivers to take one action, and each receiver sees his own private message plus those of his immediate neighbors. The common-sense view is that more communication among receivers should make them harder to manipulate. The authors show this can fail: for hierarchical stellar networks, multi-center constellations, disconnected galaxies, and clustered cliques, there are link additions that strictly raise the sender's expected payoff, often to V_n^k, the maximum value the sender could get in an empty network. Thus the sender's value is not monotone in network density, and who observes whom matters more than how many links exist. The paper also exhibits networks in which no extension helps and notes that these sender-beneficial additions worsen receivers' ability to learn the true state.","feed_headline":"Denser networks can fully restore a persuader's power","feed_subtitle":"When receivers share local messages, strategic link additions can hit the empty-network upper bound.","key_machinery":"The central tool is the notion of information domination: receiver i information-dominates j if j's neighborhood is a subset of i's. The paper relies on Lemma 3, a cited sufficient condition, which says that if a network has no information-dominating pairs, the sender can achieve the empty-network upper bound V_n^k. The constructive half of the paper builds link additions that break every information-dominating pair, then invokes Lemma 3 to conclude the upper bound is reached.","core_discovery":"The paper's central claim is that in a Bayesian persuasion problem where a sender needs a critical mass of receivers to choose a desired action, and where each receiver observes the messages sent to himself and to his direct neighbors, extending the communication network can strictly increase the sender's expected utility. For several natural families—stellar networks with a hierarchical center, constellations with multiple centers, galaxies built from disconnected star components, and cluster networks of equal-sized cliques—the paper constructs an explicit extension on which the sender achieves the upper bound V_n^k, the payoff of optimal private signaling on an empty network. The mechanism","pith_inferences":["The information-domination condition suggests a general design principle: any network extension that makes all neighborhoods incomparable should restore private-signaling power; this may apply to random graph models or community detection beyond the specific families studied.","A testable behavioral prediction: in a lab experiment with a threshold voting task, adding a small number of cross-hierarchy links (e.g., connecting a peripheral member of one clique to an outsider) should increase the sender's success rate relative to the base network, while adding all links should lower it back to public-signaling levels.","For social media policy, this implies that recommender systems that create 'boundary spanner' connections—links between users of different hierarchical depths—may inadvertently amplify the platform's or advertisers' persuasive power; density alone will not reveal the risk.","The paper's non-monotonicity could also inform design of deliberation procedures: if an institution controls the communication graph (e.g., committee meetings, town halls), choosing which participants can talk to whom is itself a design instrument that can be used for manipulation."],"forward_implications":["The sender's value is not monotone in network density: adding links can strictly increase the probability of reaching critical mass, even when starting from a network where the sender is constrained.","In stellar, constellation, galaxy, and cluster networks satisfying the stated conditions, an optimal experiment on the extension reaches V_n^k, the same value as in the empty network—so the sender loses nothing from communication if links are chosen well.","Such beneficial extensions come at the receivers' expense: the probability that the collective outcome matches the true state decreases.","Platform designers and policymakers should not equate more communication with greater consumer or voter protection; link recommendations can actually increase manipulability.","For marketing and voting, the results imply semi-connected sets of consumers or voters are not necessarily harder to persuade; the structure of the added links determines whether persuasion power rises or falls."],"fun_headline_variants":["Strategic network links can restore full persuasion power","Peer chatter can boost a persuader's reach to the max","Denser peer networks can hit the persuasion upper bound","Link additions can fully revive a persuader's influence","More local sharing can restore the sender's optimal payoff"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The upper-bound results depend on Lemma 3, a theorem imported from a cited paper: if a network has no information-dominating pairs, the sender can achieve the private-signaling value V_n^k; if that theorem is false or carries unstated conditions, the main constructions in Theorem 1, Theorem 2, Proposition 3, and Proposition 4 collapse.","fun_headline_variants_meta":{"raw":{"variants":["Strategic network links can restore full persuasion power","Peer chatter can boost a persuader's reach to the max","Denser peer networks can hit the persuasion upper bound","Link additions can fully revive a persuader's influence","More local sharing can restore the sender's optimal payoff"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1009,"prompt_tokens":686,"completion_tokens":323,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":430,"completion_tokens_details":{"reasoning_tokens":247}},"tokens_in":430,"tokens_out":323,"duration_ms":4497,"temperature":1.0,"reasoning_tokens":247,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T19:41:20.344683+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a small instance of Theorem 1 (e.g., n=8, k=4, lambda0(X)=0.3, with a stellar component of six nodes of depth 2 and two external nodes), explicitly build the extension described in the proof and numerically optimize the sender's experiment over finite message sets; if the optimum is strictly below V_8^4 = 0.9, the central claim fails. More directly, any network with no information-dominating pairs whose optimal value is below V_n^k would refute the imported Lemma 3 and with it the paper's upper-bound constructions.","supporting_citations":[],"review_version":1}