{"id":"54bcb5b5-6696-480d-82ab-6ad4fdd378ae","arxiv_id":"2602.00934","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"In an active-learning network model, homophily slows learning when networks are sparse but accelerates it when they are dense, provided the group being studied has enough experimenters.","lead":"This paper builds a model in which people learn about a risky opportunity from friends who may be similar or different, and shows that homophily can either help or hurt learning depending on how connected people are. The result matters because it offers a mechanism for why social connections across groups can matter for economic mobility, but the broad claim is only proven under special conditions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unproven b(1)=1 threshold limits the sparse/dense homophily result; abstract overstates scope.","rationale":"The reader identified the weakest assumption as the unproven b(1)=1 threshold, which is exactly the most load-bearing concern. The central claim's quantitative prediction about sparse versus dense networks is derived under this assumption, and the paper provides no proof of its domain of validity. The general Proposition 2 gives a condition in terms of equilibrium objects, but the translation to primitives in Proposition 3 requires b(1)=1. Without knowing the cb threshold, one cannot tell whether the abstract's broad statement about sparse/dense networks applies to the cases it claims. This is not a rejection of the paper—the mechanism is plausible and the b(1)=1 case is worked out correctly—but it is a genuine gap that should be addressed before full acceptance. The wrong appendix reference in Proposition 2's proof is a minor issue; the substantive gap is the b(1)=1 assumption and the dropped π_g > π_b condition in the abstract. The proposed check—deriving the threshold and examining the b(1)<1 region—would settle whether the sparse/dense dichotomy is robust. Therefore, conditional acceptance is the appropriate verdict, consistent with the reader's assessment.","tokens_in":17562,"tokens_out":15112,"duration_ms":146365,"concrete_test":"Derive the threshold on c_b guaranteeing b(1)=1 by evaluating the blue-agent posterior at the steady state under the most pessimistic feasible signal profile (e.g., all high-cost observed friends are greens who choose safe) and imposing the posterior remains at least c_b. Then, for parameters where b(1)<1, solve the full two-dimensional fixed-point system for (g(1), b(1)) and compute the sign of ∂g(1)/∂h_g via the implicit function theorem; check whether it is positive for some d_g < log_{1−π_b}((π_g−π_b)/π_g). If such a region exists, Proposition 3's threshold is not robust to relaxing the b(1)=1 assumption.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim—that homophily hampers learning in sparse networks and enhances it in dense ones—rests on Proposition 3, which assumes b(1)=1 (blue high-cost agents always invest in the good state). The paper asserts this holds for c_b below an unspecified 'certain threshold' but never derives it. If b(1)<1, the general condition from Proposition 2 is π_g g(1) > π_b b(1), which is weaker than π_g g(1) > π_b used in the sparse/dense derivation. There may exist parameter regions with d_g below the paper's threshold where g(1) is nonetheless increasing in h_g, so the 'hampers in sparse networks' result could fail. Additionally, the abstract drops the required π_g > π_b condition, making the headline broader than proven. Without a characterization of when b(1)=1 (or a proof that the comparative static survives its failure), the scope of the central claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops an overlapping-generations model of social learning in which members of two groups (blues and greens) observe the actions and payoff outcomes of friends from the previous cohort. The risky action has an unknown common value, while agents have group-specific probabilities of a high cost. Homophily affects both the quality of information (similar others are more informative) and the quantity of information (homophily changes the probability that informative actions are taken). The main theoretical results are: (i) in the partial-homophily case, the steady-state green investment rate is increasing in green homophily iff π_g g(1) > π_b b(1) (Proposition 2); (ii) under an assumed corner condition b(1)=1, this comparative static reduces to the primitive condition π_g > π_b and d_g above a log threshold (Proposition 3); and (iii) in a multi-cost extension, complete learning is the unique stable steady state exactly under perfect cost homophily (Proposition 4), and color-blind cost homophily generates incidental blue/green homophily (Proposition 5). The paper argues that homophily hampers learning in sparse networks but enhances it in dense networks.","tokens_in":17834,"tokens_out":8835,"duration_ms":102239,"significance":"If the claimed results hold in the stated generality, the paper makes a useful contribution to the social-learning literature by modeling active information generation in a homophilous network and delivering a primitive-based comparative static. The contrast with passive-learning results such as Lobel and Sadler (2016) is interesting and potentially important for policy discussions of cross-group mentorship and network density. The paper does not provide code or data, and the proofs are analytical. Propositions 1, 4, and 5 have reasonably complete proof sketches, and Proposition 3's fixed-point computation is internally correct conditional on its assumptions. However, the central sparse/dense homophily claim depends on an assumption—b(1)=1—whose primitive domain is not characterized, and the abstract states the result more broadly than the proven hypotheses. The contribution is credible, but the scope of the headline result needs to be corrected or the missing threshold condition supplied.","major_comments":[{"comment":"The central comparative static is proven only under the condition b(1)=1. The text says: 'This holds true for any c_b below a certain threshold that guarantees the cost is low enough to prevent any indirect inference from convincing the blue group to take the safe course of action.' No such threshold is ever derived or even bounded. If b(1)<1, then in state v=1 a positive-cost blue agent choosing the safe action sends an indirect negative signal, which couples the v=1 and v=0 dynamics and invalidates the decoupled fixed-point equation g=1−(1−[h_gπ_g g+(1−h_g)π_b])^{d_g}. The threshold d_g>log_{1−π_b}((π_g−π_b)/π_g) is derived from that equation. Without a characterization of when b(1)=1 is guaranteed, the paper does not establish the stated domain of the sparse/dense result.","section":"§3.2, Proposition 3"},{"comment":"The abstract's claim that 'homophily enhances learning in sufficiently dense networks' is not true as stated. In the framework of Proposition 3, if π_g ≤ π_b, the steady-state g(1) is decreasing in h_g regardless of how large d_g is; if π_g > π_b, the result requires d_g to exceed a specific threshold that depends on π_g, π_b, and h_g through the fixed point. The abstract also drops the maintained conditions c_g > p ≥ c_b and b(1)=1. Either the headline claim should be qualified to state the full parameter conditions, or the model needs a separate result showing dense networks are sufficient in a broader region. As written, the headline overstates the proven comparative static.","section":"Abstract and §3.2"},{"comment":"Proposition 2 is announced with no explicit parameter restrictions: 'Let (g(·),b(·)) be a regular steady state. Then g(1) is increasing in h_g if and only if π_g g(1) > π_b b(1).' The proof, however, invokes the dynamics in Appendix C, which are derived under the special assumption c_g > p ≥ c_b, and relies on the equality g_{t+1}(0)=0 'for any α_t'—a property that holds for green agents only when c_g > p. If Proposition 2 is intended as a statement about the special case c_g > p ≥ c_b, that restriction must be written into the proposition. As stated, the proof does not cover regular steady states with other cost-prior configurations (e.g., where green agents can be induced to take the risky action from the prior, or where both groups have high default costs). Additionally, the proof cites 'equations (6)–(9) in appendix B,' but these equations appear in Appendix C; the cross-reference sh","section":"§3.2, Proposition 2"}],"minor_comments":[{"comment":"The cross-references inside the proof of Proposition 2 point to 'appendix B' when the detailed dynamics and equations (6)–(9) are in Appendix C. This should be fixed throughout.","section":"Appendix A / Appendix C"},{"comment":"The French and Poterba reference contains a typo: 'Stock Pries' should read 'Stock Prices.' Also, the accents in 'Hélène' appear as 'H´ el` ene' in the reference list.","section":"References"},{"comment":"The figure plots steady-state g(1) for real-valued d_g while the model restricts d_g to positive integers. The caption notes this, but it would be clearer to mark which integer values are actually admissible in the model and how the comparative static is intended when d_g is not on the plotted grid.","section":"Figure 1"},{"comment":"The main text says the existence proof 'follows from a standard fixed-point argument' and is omitted; Appendix C later supplies a Kakutani fixed-point argument. Since the appendix already contains the argument, the main text could refer to it instead of saying the proof is omitted.","section":"§2"},{"comment":"The statement of Proposition 4 assumes d_{θ,c}>1 for all (θ,c), but the proof of the 'only if' direction appears to use only the existence of a value v between two costs. It may be useful to state explicitly why the degree condition is needed there, in particular for the stability of complete learning in the 'if' direction.","section":"Proposition 4"}],"recommendation":"major_revision","confidential_remarks":"The paper's main idea is worthwhile and the special-case results are mostly solid, but the central comparative static is not established for the claimed scope. The missing b(1)=1 threshold is a gap that the authors can plausibly fix either by deriving explicit sufficient conditions on c_b, π_b, p, and the homophily parameters, or by weakening the abstract and Proposition 3 to the proven subcase. I would not recommend rejection, since the model and the local results are credible and the required repair is within the scope of a revision. The referee report should emphasize that the abstract's 'dense networks suffice' wording is misleading unless the conditions π_g>π_b and d_g above threshold are incorporated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the paper: it's a genuine new result on when homophily helps social learning. The model has agents choosing risky vs safe actions with unknown common value, observing friends' payoffs. The twist is that information is endogenous—actions generate the signals, so homophily changes both quality and quantity of information. The notable finding is a sign reversal: homophily hurts in sparse networks, helps in dense ones. That contrasts with Lobel-Sadler's passive-learning result, and the sample herding mechanism is new. The core mechanism is right.\n\nWhat's good: the setup is clean, Lemma 1's monotonicity is solid, and Proposition 1's full-homophily benchmark is fully worked. The extension in Appendix B—cost homophily generating incidental blue/green homophily—is a nice free-standing contribution. The paper is honest about many of its own limitations.\n\nThe soft spots are real but fixable. Proposition 2 is stated for every regular steady state, but the proof is sketchy: it says equations are in Appendix B when they're actually in Appendix C, and the local decoupling of the v=0 and v=1 dynamics is asserted rather than fully proven. More importantly, the headline comparative static in Proposition 3 assumes b(1)=1, i.e., the blue high-cost agents always invest in the good state. The paper says this holds for c_b below a 'certain threshold' but never derives it. If b(1)<1, the condition becomes π_g g(1) > π_b b(1), which is weaker, and the closed-form dynamics in Appendix C no longer apply. The abstract drops these conditions entirely, so the 'homophily hampers in sparse, helps in dense' claim is stated more broadly than proven.\n\nI wouldn't call any of this a dealbreaker. The special-case results are coherent, and the mechanism is plausible. But a referee should ask the authors to give the b(1)=1 threshold or show the comparative static survives without it, and to state Proposition 3 with its hypotheses plainly in the abstract or introduction.\n\nBottom line: send it out. This deserves a careful referee. It's a serious paper with a real idea, and the issues are the sort that revision can fix.","headline":"A real new mechanism—homophily changes the endogenous supply of information—but the headline comparative static is proven only under an uncharacterized boundary condition.","tokens_in":18294,"tokens_out":3731,"would_cite":true,"duration_ms":35692,"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":"The paper proves homophily helps social learning when networks are dense and hurts it when they are sparse, and gives an exact threshold condition for when green homophily raises green investment.","keywords":["social learning","homophily","endogenous information","herding","overlapping generations","networks","risky investment","group inequality"],"falsifier":"Simulate or solve the model with blue cost c_b above the unstated threshold so that b(1)<1; if g(1) then fails to increase in h_g even when π_g>π_b and d_g exceeds the log threshold, Proposition 3's comparative static is falsified. Alternatively, calibrate the model to a network dataset and check whether the predicted crossover in homophily's effect by degree is observed.","tokens_in":17420,"feed_emoji":"🕸️","tokens_out":5483,"duration_ms":59937,"temperature":0.7,"pith_summary":"The paper asks whether homophily—people's tendency to befriend similar others—helps or hurts social learning when the information people exchange is itself produced by their choices. It models two groups, blues and greens, who choose between a safe action and a risky action whose payoff is unknown; each generation observes its friends' choices and outcomes. Homophily improves the quality of information, because similar friends reveal more about one's own payoff, but it can dry up the quantity of information, because similar friends may all avoid the risky action. The paper's central finding is that the net effect depends on network density: homophily harms learning in sparse networks and helps it in sufficiently dense networks. It also proves a sharp condition for when green homophily raises green investment in the risky action, and shows that cost-based assortativity can generate 'incidental' homophily between groups even without color-based preferences.","feed_headline":"Dense networks turn homophily from barrier into boost","feed_subtitle":"Same similarity bias that blocks learning in sparse networks speeds it up in dense ones.","key_machinery":"The core object is the pair of steady-state participation rates (g(v), b(v)) for high-cost agents of each group in each value state v∈{0,1}, along with the degrees d_g, d_b and frequencies π_g, π_b of high-cost types. The key identity is that the derivative of g(1) with respect to green homophily, ∂g(1)/∂h_g, has the same sign as π_g g(1) − π_b b(1): homophily raises green investment exactly when a random green friend is more likely than a random blue friend to be a high-cost agent taking the risky action. When b(1)=1 the dynamics collapse to a one-dimensional concave fixed-point equation, from which the explicit threshold d_g > log_{1−π_b}((π_g−π_b)/π_g) is derived.","core_discovery":"The central discovery is that the comparative static of homophily flips with connectivity. In the benchmark case where blues always take the risky action in the good state and greens take it only after observing positive evidence, the steady-state fraction of greens investing, g(1), is increasing in green homophily h_g if and only if π_g > π_b and d_g > log_{1−π_b}((π_g−π_b)/π_g). That is, homophily helps greens precisely when green high-cost agents are more common than blue ones and each green has enough friends to be likely to see the risky action being tried. The same model shows that in sparse networks homophily can produce 'sample herding,' where a group rationally avoids the risky acti","pith_inferences":["The threshold condition suggests an empirical test: in data on adoption of a risky technology or educational investment, the effect of ethnic or income homophily on investment should be negative among low-degree individuals and positive among high-degree individuals, with the crossover located near the log threshold.","The appendix's 'incidental homophily' result implies that measuring homophily on one trait (say race) may confound sorting on costs: even if friendship formation is color-blind, different cost distributions across groups generate apparent race homophily, so interventions that equalize cost distributions may be more effective than direct attempts to mix groups.","A dynamic policy implication of the model's logic: subsidize cross-group mentoring early in a diffusion process, then switch to same-group mentoring once the disadvantaged group's participation rate exceeds π_b/π_g; the paper does not discuss the optimal timing, but it follows from the fixed-point equation."],"forward_implications":["In sparse networks, homophily induces sample herding: a group can settle at a steady state where no high-cost member invests, because nobody observes the risky action being tried; this can persist even when the action is profitable.","In dense networks, the same homophily raises the steady-state investment rate and accelerates learning, so policies that reduce homophily may backfire once connectivity is high.","Cross-group links are most valuable early, when one group has little experience with the risky action; once both groups act similarly, same-group links become more informative.","When greens have more high-cost types than blues and each green has enough friends, increasing green homophily strictly increases the green investment rate; the threshold degree is given explicitly.","Higher correlation of costs across groups can increase the positive impact of homophily, because groups act more alike and the informativeness of similar others grows."],"fun_headline_variants":["Homophily flips from hindrance to help as networks densify","When homophily boosts learning: the density threshold","Sparse networks punish similarity; dense ones reward it","Homophily's learning payoff flips with network connectivity","The network density that turns homophily into an asset"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proposition that homophily helps in dense networks assumes that blue high-cost agents always invest in the risky action in the good state (b(1)=1); the paper asserts this holds for blue costs below an unspecified threshold but never derives it, and if it fails the closed-form dynamics and the sign condition π_g g(1)>π_b b(1) can break down.","fun_headline_variants_meta":{"raw":{"variants":["Homophily flips from hindrance to help as networks densify","When homophily boosts learning: the density threshold","Sparse networks punish similarity; dense ones reward it","Homophily's learning payoff flips with network connectivity","The network density that turns homophily into an asset"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000681,"raw_usage":{"total_tokens":2878,"prompt_tokens":638,"completion_tokens":2240,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":382,"completion_tokens_details":{"reasoning_tokens":2172}},"tokens_in":382,"tokens_out":2240,"duration_ms":15280,"temperature":1.0,"reasoning_tokens":2172,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:49:59.620346+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate or solve the model with blue cost c_b above the unstated threshold so that b(1)<1; if g(1) then fails to increase in h_g even when π_g>π_b and d_g exceeds the log threshold, Proposition 3's comparative static is falsified. Alternatively, calibrate the model to a network dataset and check whether the predicted crossover in homophily's effect by degree is observed.","supporting_citations":[],"review_version":1}