{"id":"cc02e7e8-b821-4e1b-848c-c67b28de6140","arxiv_id":"2607.09765","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Heterogeneous cost-coupled corrector placement in a consensus swarm stays submodular, so cost-benefit greedy is near-optimal, and cost-quality curvature decides few-strong versus many-medium oracles in a task-dependent way.","lead":"A swarm of cheap, unreliable AI agents can be steered to a correct answer by a few expensive oracle correctors. This paper shows how much budget is needed, where to place the oracles, and whether to buy a few strong models or many medium ones—and finds the answer is task-dependent.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"The curvature dichotomy and deployment reading rest on an unvalidated accuracy-to-pin-strength proxy that the Stage-B checks only partially close.","rationale":"The reader's weakest assumption is exactly the load-bearing bridge: the proxy from measured accuracy to abstract wi, plus the adequacy of linear consensus for multi-agent LLM correctness. The formal package (submodularity under heterogeneous pins, (1-1/e) greedy, closed-form complete-graph frontier, curvature dichotomy) is internally consistent and carefully proved; the appendices extend it to sparse, modular and directed graphs. The empirical ladders are more careful than typical LLM-ops work. What remains soft is precisely the transfer of the measured curvature into the model that produces the deployment verdict. The Stage-B linearization remark and the small real-swarm checks reduce but do not eliminate that gap. Hence the verdict stays CONDITIONAL; no stronger rejection is warranted, and no upgrade to ACCEPT is justified until the continuous-pin calibration is performed.","tokens_in":25317,"tokens_out":599,"duration_ms":15635,"concrete_test":"On the same Qwen3 ladder and 2-AFC items, fit an effective continuous pin strength wi by matching the observed steady-state error (or magnetization) of a small grounded-Laplacian swarm (N≈20-40) whose free agents are the weakest model and whose single corrector is each ladder size in turn; recompute the curvature sign of the resulting empirical w(c). If the sign flips relative to the raw accuracy curve on any of the three tasks, the deployment reading of Theorem 3 is unsupported by the linear model.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Theorem 3 decides spread vs concentrate from the sign of (N+κ+w)w''-3(w')^{2} applied to the measured cost-quality law w(c). Section 7 explicitly states that the 2-AFC log-prob accuracy is only a proxy for abstract pin strength wi and is \"not otherwise validated.\" The later Stage-B checks (N=25, Fig. 7a) show that measured reliability enters the nonlinear cascade roughly as the weight 2q-1 (r=0.98) and that ladder strengths predict truth-win rates (corr 0.94), but these are small-scale, content-free or single-band experiments that never re-estimate the continuous diagonal pin wi inside the linear grounded-Laplacian M(R) whose coherence H is the object of Theorems 1-3. If the mapping from accuracy to wi is nonlinear, saturating, or task-dependent in a way that flips the sign of the curvature expression, the empirical concave/convex classification (and therefore the \"spread for factual/math, concentrate for code\") does not transfer to the model that actually justifies the deployment claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper studies budgeted placement of strong, cost-coupled oracle correctors in a weak multi-agent consensus swarm. Agents evolve under a grounded Laplacian; each oracle at node i adds a diagonal pin of strength wi = w(ci). Quality is the coherence H(R) = tr M(R)^{-1}. Theorem 1 shows that the error reduction ρ(R) = H(∅) − H(R) remains monotone submodular for heterogeneous wi > 0 via an M-matrix / entrywise-nonnegative-inverse argument, so cost-benefit greedy attains a (1−1/e) guarantee (Theorem 2) and yields a budget–correctness frontier B⋆(ε). On the complete graph, Proposition 1 gives a closed form for H, and Theorem 3 shows that the fixed-budget optimum is decided by the sign of (N+κ+w)w'' − 3(w')²: concave w spreads, sufficiently convex w concentrates. Empirically, Qwen3 (and Gemma-4) cost–quality laws are concave for factual and math 2-AFC verification but convex for emergent code tracing. Stage B treats a nonlinear majority cascade, giving influence-balance thresholds (exact on KN, rigorous on dense ER) and small real-LLM swarm checks.","tokens_in":25584,"tokens_out":1599,"duration_ms":40436,"significance":"If the results hold, the paper supplies a missing dynamical foundation for budgeted multi-agent LLM assembly: a consensus-correctness object with a cost-coupled quality law, a submodular knapsack guarantee under heterogeneous pins, a closed-form frontier on KN, and a single scalar (curvature of w) that decides few-strong versus many-medium oracles. Strengths include carefully scoped proofs (M-matrix argument; pairwise-balancing for equal spread), recovery of the homogeneous companion as Corollary 1, released code and per-item scores, bootstrap CIs, a second-family replication, and an honest Limitations section. The Stage-B influence balance and fallible-corrector extension (Proposition 2, Corollary 2) usefully connect average-error and high-probability views. The task-dependent empirical curvature finding is practically relevant if the accuracy-to-pin proxy transfers.","major_comments":[{"comment":"Section 7 explicitly states that 2-AFC log-probability accuracy is only a proxy for abstract pin strength wi and is “not otherwise validated,” yet the abstract and Remark 3 present the task-dependent spread/concentrate verdict as a deployment conclusion of Theorem 3. Limitations claim the concave sign survives identity, logit, and tested monotone reparametrizations because saturation is per-item, but this is not elevated to a formal statement with conditions under which sign((N+κ+w)w''−3(w')²) is preserved. Either (i) add a short proposition quantifying curvature-sign preservation under monotone maps of measured accuracy to wi, with the adversarial maps tested, or (ii) soften the abstract/conclusion so the empirical half is clearly a measurement of a proxy scalar, not a validated transfer into M(R). Without one of these, the headline “verdict is genuinely task-dependent” overreaches the","section":"Section 7, Theorem 3, Abstract"},{"comment":"The optimization object throughout Theorems 1–3 is H = tr M^{-1}, while the motivating tracking error under isotropic fault covariance is σ² tr M^{-2} (Section 3). Submodularity and the Sherman–Morrison gain (3) are proved for H; the complete-graph curvature dichotomy (Theorem 3, Appendix D) uses the closed form (6) specific to tr M^{-1}. Remark 6 notes both functionals are minimized by leverage placement near consensus, but that does not automatically transfer the equal-spread / single-oracle dichotomy or the factor 3 in (7) to tr M^{-2}. Please either (a) state that the paper’s object is deliberately the standard coherence criterion and that the MSE dichotomy is left open, or (b) sketch whether the pairwise-balancing argument extends to tr M^{-2} on KN (even a numerical check on the equal-split vs single-oracle comparison under concave/convex w would help).","section":"Section 3, Eq. (2); Theorem 3; Appendix D"},{"comment":"The real-swarm confirmations that partially close the proxy gap (Section 8, Figs. 6–7) are small and narrow: N=25, one BA graph, one hardness band, 30–40 trials, and for real reasoning correctors the leverage curve reaches parity rather than certainty. Proposition 2 / Corollary 2 give a clean composition condition (reliability-weighted degree balance plus no single fallible hub dominating), but the experiments do not check the dominance condition d_free_i ≤ μ_{-i} under the measured medium-reliability correctors on hubs. Given that Corollary 2 is what makes “many medium on high-degree nodes” composition-safe, a brief check or explicit caveat that the N=25 runs do not verify the safety margin would keep Stage A/B composition claims proportionate.","section":"Section 8, Proposition 2, Corollary 2, Figs. 6–7"}],"minor_comments":[{"comment":"The abstract has several garbled phrases (“manye curvature”, “sharply increasion”, “math verificatio convex foremergent”, truncated GitHub URL). These should be cleaned to match the body abstract.","section":"Abstract"},{"comment":"Assumption 1 forces w concave, while Theorem 3 and the code-tracing experiment treat sharply convex w. A one-sentence note that Assumption 1 is the diminishing-returns case and that Theorem 3 also covers the opposite curvature regime would avoid a framing inconsistency.","section":"Assumption 1; Theorem 3"},{"comment":"Figure captions are dense; labeling panels (a)/(b) more explicitly in the text when first cited (Figs. 1–2, 6–8) would help navigation.","section":"Figures 1–2, 6–8"},{"comment":"Related work on model cascades and FrugalGPT [18] is cited; a brief contrast with cascade escalation (confidence-triggered single strong model) versus simultaneous multi-oracle placement would clarify the operational difference.","section":"Section 2"},{"comment":"Appendix H correctly notes that submodularity does not transfer to directed M; flagging this earlier (e.g., end of Section 4 or Limitations) would help readers who jump to directed swarms.","section":"Appendix H; Section 4"}],"recommendation":"minor_revision","confidential_remarks":"The mathematical core (Theorems 1–3, complete-graph closed form, appendices) is solid and appropriately scoped; the main risk is over-claiming the empirical deployment reading relative to the admitted proxy. I would not reject on that basis—the authors already caveat it—but the abstract should not outrun the model. Fit for a theory-leaning AI / multi-agent venue is good if the proxy language is tightened. Same-author companion [1] is cleanly recovered as a special case rather than double-counted novelty."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core is clean: coherence H = tr M(R)^{-1} stays monotone submodular under heterogeneous positive pin strengths (Theorem 1), so cost-benefit greedy gets the usual 1-1/e knapsack guarantee (Theorem 2). On the complete graph they give an exact closed form and a sharp curvature test—(N+κ+w)w'' − 3(w')²—that decides equal spread versus concentration (Theorem 3). That scalar is the actual novelty relative to classical identical-leader selection. The appendices then push the same object to sparse regular, SBM, and directed graphs via cavity methods; the directed case correctly restricts feedback to reciprocated edges. Code and data are released.\n\nThe empirical half is more careful than most LLM-ops papers: single-family Qwen3 ladder, difficulty bands, bootstrap CIs, a second family (Gemma-4), and a separate code-tracing task that actually flips the curvature. So the claim “concave for factual/math, convex for emergent code” is not hand-waved.\n\nThe soft spot is real but scoped. Section 7 itself says 2-AFC log-prob accuracy is only a proxy for abstract pin strength wi and is “not otherwise validated.” The Stage-B N=25 checks show measured reliability enters the nonlinear cascade roughly as 2q−1 and that ladder strengths predict truth-win rates, but they never re-estimate continuous diagonal pins inside the linear M(R) whose H is the object of Theorems 1–3. If that map is nonlinear enough to flip the curvature sign, the deployment reading does not transfer. That is a modeling-bridge gap, not a hole in the math. Companion recovery as the unit-cost case is legitimate, not circular.\n\nWho it is for: people who already think about multi-agent LLM budgeting or resilient consensus and want a formal handle plus a measurable scalar. It deserves a serious referee. I would bring it to reading group and would cite the submodularity + curvature results; I would not yet treat the spread-vs-concentrate deployment advice as settled for production swarms.","headline":"Clean theory package on budgeted corrector placement with a real curvature dichotomy; the deployment reading is only partially bridged to actual LLM swarms.","tokens_in":26255,"tokens_out":531,"would_cite":true,"duration_ms":6089,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"How much correctness costs in a weak multi-agent swarm is decided by where oracles sit and by the curvature of the cost-quality law.","keywords":["multi-agent systems","oracle placement","submodular maximization","grounded Laplacian","budget-correctness frontier","cost-quality law","consensus dynamics","LLM swarms"],"falsifier":"Measure cost-quality curves and run real multi-agent swarms on the same tasks: if a clearly concave law still produces better average correctness by concentrating on one large model, or if a convex law still prefers many medium correctors, the curvature dichotomy fails to transfer.","tokens_in":26120,"feed_emoji":"⚖️","tokens_out":675,"duration_ms":6545,"temperature":0.7,"pith_summary":"A swarm of cheap unreliable agents can still reach a correct consensus if a few expensive oracles pin selected nodes toward the truth. The paper models that loop as a grounded consensus on a graph, measures tracking quality by the coherence H(R) = tr M(R)^{-1}, and couples each oracle's pin strength to its cost through a quality law w(c). It proves that the error reduction stays monotone submodular even when the oracles have different strengths, so a simple cost-benefit greedy placement is within a 1-1/e factor of optimal at any budget and yields the least spend B*(eps) that certifies an eps-correct consensus. On the complete graph that frontier is closed-form, and a single curvature test on w decides whether the same budget is better spent on many medium oracles or on a few strong ones. Measurements on model ladders show the law is concave for factual and math verification (spread wins) but convex for emergent code tracing (concentrate wins), so the practical verdict is task-dependent.","feed_headline":"Correctness cost is set by oracle placement and quality curvature","feed_subtitle":"Greedy placement is near-optimal; concave laws prefer many medium correctors, convex ones prefer few strong ones.","key_machinery":"The grounded operator M(R) = L + kappa I + sum of diagonal pin strengths, whose coherence H(R) = tr M(R)^{-1} is the quality measure; its M-matrix structure keeps the resolvent entrywise nonnegative and thereby preserves submodularity of the error reduction even for unequal wi.","core_discovery":"The coherence objective remains monotone submodular under heterogeneous positive pin strengths, so cost-benefit greedy is a (1-1/e) approximation for budgeted oracle placement; on the complete graph the same budget optimally spreads under a concave cost-quality law and concentrates only under a sharply convex law whose curvature exceeds a precise threshold, and real LLM ladders realize both regimes depending on the task.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Submodular coherence keeps greedy oracle placement near-optimal","Budgeted oracles: few strong or many medium by cost curvature","Swarm correctness frontier set by oracle placement and curvature","Concave laws spread oracles; convex ones concentrate them","Task-dependent curvature chooses few strong or many medium correctors"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That a model's measured accuracy on a two-choice verification task is a valid stand-in for the abstract pin strength that appears in the linear consensus equations, and that those linear dynamics adequately describe how multi-agent LLM swarms actually correct each other.","fun_headline_variants_meta":{"raw":{"variants":["Submodular coherence keeps greedy oracle placement near-optimal","Budgeted oracles: few strong or many medium by cost curvature","Swarm correctness frontier set by oracle placement and curvature","Concave laws spread oracles; convex ones concentrate them","Task-dependent curvature chooses few strong or many medium correctors"]},"model":"grok-4.5","effort":"low","cost_usd":0.003832,"raw_usage":{"total_tokens":1242,"prompt_tokens":809,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":38320000,"prompt_tokens_details":{"text_tokens":809,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":349,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":809,"tokens_out":84,"duration_ms":4025,"temperature":1.0,"reasoning_tokens":349,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T16:09:51.879586+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Measure cost-quality curves and run real multi-agent swarms on the same tasks: if a clearly concave law still produces better average correctness by concentrating on one large model, or if a convex law still prefers many medium correctors, the curvature dichotomy fails to transfer.","supporting_citations":[],"review_version":1}