{"id":"bda099ad-1866-48f7-bb9a-24133b4af0d7","arxiv_id":"2602.21556","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"In a principal-agent model of compound AI, aggregation expands the set of outputs a designer can elicit exactly when one of three mechanisms — feasibility expansion, support expansion, or binding set contraction — holds in strengthened form.","lead":"This paper proves when combining answers from several copies of the same AI model lets a system designer obtain outputs that no single copy could produce on its own. It gives a complete mathematical rule for when aggregation adds power, and illustrates the rule with a small LLM experiment.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition D.6's equivalence proof ignores mixed support+conic certificates; fixed-d version is false, leaving Theorems 4.3/4.4 unsupported at a key step.","rationale":"The reader's weakest assumption is the fidelity of the linear-conic, deterministic-argmax model to real LLMs. My stress-test instead targets an internal correctness gap in the proof of the central characterization. Theorems 4.3 and 4.4 are the strongest claims, and their proof relies on Proposition D.6, which asserts equivalence between two power-characterizing conditions. The minimax argument in D.4 is intended to establish that a robust separation direction can be witnessed by either a strengthened support expansion or a strengthened binding-set contraction for each agent. However, the reduction to single-constraint certificates drops mixed Lagrange multipliers: a joint certificate using both a nonnegativity constraint and a conic constraint can certify emptiness even when neither individual certificate exists. The explicit counterexample above disproves the fixed-d equivalence stated in D.4. This does not by itself disprove the existential theorem, but it means the manuscript's proof of Theorems 4.3 and 4.4 is unsound at a load-bearing step. The empirical illustration is a toy demonstration and does not test the exact characterization, so it cannot compensate for this gap. I therefore keep the reader's CONDITIONAL verdict, but for a stronger reason: the mathematical core needs repair or independent verification.","tokens_in":35827,"tokens_out":31144,"duration_ms":258100,"concrete_test":"1) Verify the displayed instance: solve the small LP for v to confirm {u+λd}∩B_{S_k,V_k}=∅ while Definition 4.2 conditions 2a/2b both fail for that d. 2) Run an exhaustive search over M=3–4, C with entries in {-1,0,1}, all support/binding patterns, and d on a fine grid, comparing the existential satisfiability of Def 4.2 vs Def 4.8. If any instance satisfies Def 4.8 but not Def 4.2, Theorem 4.4 is false; if none, attempt a repaired proof of the existential equivalence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix D.4 asserts (Proposition D.6) that Definition 4.2 and Definition 4.8 are equivalent, and this equivalence is used in the contrapositive proof of Theorem 4.4. The proof of D.4 treats non-empty intersection of {u+λd} with B_{S_k,V_k} as being certified either by a single nonnegativity violation (condition 2a) or by a single conic violation (condition 2b). That separation is wrong: the feasibility LP can require a mixed certificate using both a nonnegativity multiplier and a conic multiplier.\n\nExplicit instance: M=3, C=[1,-1,0], aggregate statistics S_A={1,2,3}, V_A=∅, d=(-0.4,-0.9,0.3) (so 1^T d=-1, d≰0), and agent k with S_k={2,3}, V_k={1}. For this fixed d, {u+λd} is disjoint from B_{S_k,V_k}: otherwise λ>0 and v=u/λ gives v1≥0.4, v2≥0.5+v1, hence v1+v2≥1.3, contradicting 1^T v<1. Yet condition 2a fails (-d1-1=-0.6) and condition 2b fails (C d=0.5 < ∥(C)^-∥∞=1). Thus the fixed-d equivalence claimed in D.4 is false. Since Theorems 4.3/4.4 depend on Proposition D.6, the characterization is not established by the manuscript as written. (The existential version may still be true, but it requires a separate proof.)","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies compound AI systems that aggregate outputs from K homogeneous agents (LLMs) within a principal-agent framework. Outputs are nonnegative M-dimensional vectors; model-capability limitations are modeled by conic constraints C, and prompt-engineering limitations by a coarse nonnegative linear feature map α composed with nonlinear monotone transformations. The paper defines 'elicitability-expanding' aggregation operations, identifies three mechanisms—feasibility expansion, support expansion, and binding set contraction—proves necessity of one of them (Theorem 3.7), and claims a full characterization via a strengthened 'power-characterizing condition' (Definitions 4.2 and 4.8; Theorems 4.3 and 4.4). The technical core includes a single-agent elicitability characterization (Prop D.1), a fixed-feature characterization of elicitability-expansion (Lemma 4.7), and an equivalence between two forms of the power-characterizing condition (Prop D.6). Section 5 gives an empirical LLM illustration in a toy reference-generation task.","tokens_in":36160,"tokens_out":21736,"duration_ms":183916,"significance":"If the characterization is correct, the paper provides a clean geometric rule for when aggregating multiple homogeneous agents strictly expands the set of elicitable outputs, extending Kleinberg-Raghavan (2020) to conic capability constraints and multi-agent aggregation. The paper has real strengths: Lemma 4.7 reduces the problem to direction-set intersections; Prop D.1 is proved via Motzkin transposition and KKT certificates; the examples are concrete and the empirical code is provided; and the authors are explicit about the model's simplifying assumptions. However, the proof of Proposition D.6—the bridge used to prove Theorems 4.3 and 4.4—contains a load-bearing error, and there is a second sign error in the proof of Theorem 4.3. The central result is therefore not established as written, although the issues appear localized and likely repairable.","major_comments":[{"comment":"The proof asserts a pointwise equivalence between non-intersection of {u+λd} with B_{S_k,V_k} and Conditions 2a/2b. This pointwise claim is false. Take M=3, C=(1,-1,0), S_A={1,2,3}, V_A=∅, d=(-0.4,-0.9,0.3), so 1^T d=-1 and d≰0, and agent k with S_k={2,3}, V_k={1}. Then {u+λd} is disjoint from B_{S_k,V_k}: otherwise v=u/λ satisfies v1≥0.4 and v2≥0.5+v1, so v1+v2≥1.3, contradicting 1^T(d+v)<0 ⇔ 1^T v<1. Yet Condition 2a fails (-d1-1=-0.6) and Condition 2b fails (γ^T C_{V_k}d=0.5γ but ||(γ^T C_{V_k})^-||∞=γ). The Farkas certificate for disjointness requires a mixed conic-plus-nonnegativity multiplier. Since the proofs of Theorems 4.4 (D.5) and 4.3 (D.6) invoke Proposition D.6, the main characterization is not established by the manuscript as written.","section":"Appendix D.4, Proposition D.6"},{"comment":"There is a sign error in the construction of λ. After defining the features F_{p,q}, the text says 'choosing λ less than d_p^A d_q / d_p d_q^A' yields d_q - λ d_q^A ≥ 0. Since d_q^A < 0 for q ∈ N0, the required condition is λ ≥ d_q/d_q^A, not λ ≤ ...; for d_q ≥ 0 the displayed bound is negative, so the stated choice is impossible. The preceding inequality d_q ≥ d_p d_q^A/d_p^A can be used to repair the argument, but as written the claimed inclusion D_α ⊆ {u+λd^(A)} is not proved. This is a second load-bearing gap in the sufficiency direction.","section":"Appendix D.6, proof of Theorem 4.3"}],"minor_comments":[{"comment":"The proof contains several typos: 'C_{Vi}' should be 'C_{Vk}' and 'S_i' should be 'S_k'. Please also define the negative-part norm notation explicitly.","section":"Appendix D.4"},{"comment":"The expression with nested min(0,·) is hard to read; writing ||(γ^T C_V)^-||∞ would clarify that it is the ℓ∞ norm of the negative part of the vector.","section":"Definition 4.2, Condition 2b"},{"comment":"The empirical study replaces exact elicitability with closeness in ℓ1 distance to the best prompt output. The text says 'relax' but should make explicit that this tests an approximate analogue of Definition 2.4, not the exact theoretical notion.","section":"Section 5"},{"comment":"The × entries are supported by Proposition A.2 and Proposition A.1; adding explicit pointers in the table caption would improve readability.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is promising and the geometric characterization is interesting, but the proof of Proposition D.6 is invalid in a way that directly affects Theorems 4.3 and 4.4. The issue appears fixable: the authors need either a correct proof of the existential equivalence between Definitions 4.2 and 4.8, or they should restate the main theorems in terms of the alternate condition (Definition 4.8) and prove sufficiency/necessity directly. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core question is good: when does aggregating homogeneous agents expand what a principal can elicit? The paper's three mechanisms (feasibility expansion, support expansion, binding set contraction) are a useful conceptual lens, and the single-agent characterization via Motzkin/KKT is carefully built. The empirical section is a toy illustration but honestly labeled, with confidence intervals and explicit acknowledgment that the feasibility-expansion gap is small. So the paper deserves credit for a clean setup and a serious attempt at a hard problem.\n\nBut there is a load-bearing flaw. Theorem 4.3 and Theorem 4.4 depend on Proposition D.6, which claims that two versions of the power-characterizing condition (Definition 4.2 and Definition 4.8) are equivalent. The proof of that proposition is wrong. It treats the feasibility LP for \"{u + lambda d} intersects B_{S_k,V_k}\" as certified either by a single nonnegativity violation or by a single conic violation, but the LP can require a mixed certificate using both. The stress-test gives a concrete counterexample: M=3, C=[1,-1,0], aggregate with full support and no binding constraints, d=(-0.4,-0.9,0.3), and an agent with S_k={2,3}, V_k={1}. For this d, the alternative condition (Definition 4.8) holds, but both strengthened conditions 2a and 2b in Definition 4.2 fail. So the fixed-d equivalence is false, and the proof of the main theorems collapses at that step. The existential version may still be salvageable, but it requires a separate, correct argument; the current manuscript does not contain one.\n\nMinor concerns: the empirical claims are weaker than the exact theoretical notion (they use l1 closeness), and the model's linear-conic/deterministic-argmax assumptions are stylized, which the authors themselves flag. Those are proportionate and not fatal.\n\nBottom line: this is a serious paper with a genuine gap. It should not be cited as establishing the characterization, but it deserves peer review because the framework is timely and the flaw is localized. A referee should be asked to re-prove or replace Proposition D.6 and then re-check Theorems 4.3/4.4. I would not put it in a reading group until the proof is fixed, but I would not desk-reject it either.","headline":"Main characterization rests on a false equivalence; central theorems are unsupported as written, though the framework is worth engaging.","tokens_in":36760,"tokens_out":3967,"would_cite":false,"duration_ms":33834,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Complete geometric rule for when aggregating model outputs beats single-prompt elicitation.","keywords":["compound AI systems","aggregation rules","elicitability-expansion","principal-agent model","feasibility expansion","support expansion","binding set contraction","conic constraints"],"falsifier":"Falsify Theorem 4.4 by finding conic constraints C and an aggregation operation whose aggregate is feasible and fails the power-characterizing condition, yet for some feature map α each agent's feasible, budget-reducing directions miss the feature-improving cone while the aggregate's intersect it (the two conditions of Lemma 4.7). A weaker empirical falsifier: an LLM aggregation that yields an exact output (not merely ℓ1-close) that no single prompt can elicit while implementing none of the three mechanisms.","tokens_in":35612,"feed_emoji":"🧩","tokens_out":10316,"duration_ms":89661,"temperature":0.7,"pith_summary":"This paper asks whether querying several copies of the same model and aggregating their outputs can elicit outputs that no single reward-and-prompt specification could produce. In a principal-agent model where each agent maximizes a designer-supplied monotone reward over a feasible set, the authors prove that aggregation expands the elicitable set exactly when it implements one of three mechanisms: producing an output outside the agents' feasible region (feasibility expansion), combining low-support outputs into a richer-support output (support expansion), or combining boundary outputs into an interior point (binding set contraction). The basic versions of these mechanisms are necessary but not sufficient; strengthened versions, packaged as a single power-characterizing condition, are both necessary and sufficient. The upshot is a concrete test for when ensemble-style aggregation adds power to a compound AI system, along with an empirical illustration on a reference-generation task with LLMs.","feed_headline":"Pooling model outputs beats one prompt under exactly three mechanisms","feed_subtitle":"A geometric rule tells when combining several agents' answers yields outputs no single reward can elicit.","key_machinery":"The central object is the set B_{S(x),V(x)} of feasible, budget-reducing directions at an output x: directions d such that moving a small distance keeps all binding conic and nonnegativity constraints satisfied and strictly reduces the ℓ1 norm. Lemma 4.7 shows that, for a fixed feature map α, an output is elicitable iff B_{S(x),V(x)} ∩ {d : αd ≥ 0} is empty, reducing the entire problem to a geometric intersection test. The power-characterizing condition (Definition 4.2) strengthens the three natural mechanisms by demanding one direction d from the aggregate's feasible-direction set that, for every agent, violates a binding constraint by a margin that survives adding any nonnegative vector u","core_discovery":"The paper's central claim is an if-and-only-if characterization: for any conic capability constraints C, an aggregation operation x(1),…,x(K) → x(A) is elicitability-expanding if and only if the power-characterizing condition holds. The condition requires either that x(A) is infeasible under C (feasibility expansion), or that there exists a feasible, budget-reducing direction d from x(A), not all coordinates nonpositive, such that for every agent k, either some coordinate outside the support of x(k) is decreased by at least |1⊤d| (strengthened support expansion), or some nonnegative weighted combination of x(k)'s binding conic constraints is violated by a margin of at least |1⊤d| times the m","pith_inferences":["An immediate practical diagnostic follows: before deploying an ensemble or debate protocol, check whether the aggregate output escapes every agent's feasible and binding geometry; if not, the protocol cannot outperform the best single prompt under the model's assumptions.","The empirical demonstrations approximate elicitability by ℓ1-closeness, so they illustrate, rather than strictly instantiate, the theoretical notion; an exact-elicitation experiment would be stronger.","A testable extension suggested by the paper is to measure on real LLM tasks whether debate, prompt ensembling, or multi-agent research systems actually implement feasibility expansion, support expansion, or binding set contraction—the authors explicitly leave this as future work.","The characterization transfers to any domain where agents optimize a monotone objective under linear constraints and a coarse feature map, such as ensembles of classifiers or crowdsourcing workflows."],"forward_implications":["Intersection aggregation can never implement support expansion, and addition (convex-combination) aggregation can never implement feasibility expansion, so the three mechanisms are not all available to every natural rule.","Feasibility expansion alone guarantees elicitability-expansion; support expansion or binding set contraction alone do not.","If an aggregation operation implements none of the three mechanisms, it offers no power to the designer for any feature map—that is, no degree of prompt-engineering limitation makes it useful.","Even as model capability constraints relax, support expansion and binding set contraction remain sources of aggregation power against prompt-engineering limitations."],"fun_headline_variants":["Aggregation expands outputs iff one of three mechanisms holds","Three mechanisms determine when pooled AI responses beat a single prompt","Geometric rule tells when combining model outputs yields more than one prompt","Principal-agent model characterizes when multi-agent aggregation expands elicitable sets","Why pooling identical models can expand output sets: exactly three conditions"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The full characterization rests on the model's idealizations: nonnegative vector outputs, capability limits that are linear conic constraints, prompt limits that are nonnegative linear feature maps composed with strictly increasing smooth weakly concave functions, monotone rewards, and deterministic reward-maximizing agents—assumptions the authors explicitly flag in Section 6. If agents are stochastic or limitations are nonlinear or non-conic, the theorems may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Aggregation expands outputs iff one of three mechanisms holds","Three mechanisms determine when pooled AI responses beat a single prompt","Geometric rule tells when combining model outputs yields more than one prompt","Principal-agent model characterizes when multi-agent aggregation expands elicitable sets","Why pooling identical models can expand output sets: exactly three conditions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000291,"raw_usage":{"total_tokens":1537,"prompt_tokens":746,"completion_tokens":791,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":707}},"tokens_in":490,"tokens_out":791,"duration_ms":7729,"temperature":1.0,"reasoning_tokens":707,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T20:59:50.858795+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Falsify Theorem 4.4 by finding conic constraints C and an aggregation operation whose aggregate is feasible and fails the power-characterizing condition, yet for some feature map α each agent's feasible, budget-reducing directions miss the feature-improving cone while the aggregate's intersect it (the two conditions of Lemma 4.7). A weaker empirical falsifier: an LLM aggregation that yields an exact output (not merely ℓ1-close) that no single prompt can elicit while implementing none of the three mechanisms.","supporting_citations":[],"review_version":1}