Pith. sign in

REVIEW 2 major objections 4 minor 1 cited by

Complete geometric rule for when aggregating model outputs beats single-prompt elicitation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-02 20:59 UTC pith:UXTB4VDL

load-bearing objection Main characterization rests on a false equivalence; central theorems are unsupported as written, though the framework is worth engaging. the 2 major comments →

arxiv 2602.21556 v2 pith:UXTB4VDL submitted 2026-02-25 cs.AI cs.GT

Power and Limitations of Aggregation in Compound AI Systems

classification cs.AI cs.GT
keywords compound AI systemsaggregation ruleselicitability-expansionprincipal-agent modelfeasibility expansionsupport expansionbinding set contractionconic constraints
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

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.

Core claim

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

What carries the argument

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

Load-bearing premise

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.

What would settle it

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.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • 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.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

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.

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 (2)
  1. [Appendix D.4, Proposition D.6] 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.
  2. [Appendix D.6, proof of Theorem 4.3] 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.
minor comments (4)
  1. [Appendix D.4] 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.
  2. [Definition 4.2, Condition 2b] 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.
  3. [Section 5] 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.
  4. [Table 1] The × entries are supported by Proposition A.2 and Proposition A.1; adding explicit pointers in the table caption would improve readability.

Circularity Check

0 steps flagged

No significant circularity: the characterization is derived from an external (Kleinberg–Raghavan) result plus self-contained geometric/duality arguments; the flagged Appendix D.4 concern is a proof-gap issue, not a circularity.

full rationale

The paper's central claims (Theorems 4.3 and 4.4) derive elicitability-expansion from the power-characterizing condition (Definition 4.2), which is stated in terms of feasible budget-reducing directions B_{S(x),V(x)} and conic weights—objects defined independently of the elicitability-expansion definition. The main technical lemma (Lemma 4.7) explicitly extends Theorem 3 of Kleinberg and Raghavan [2020], an external result, to conic constraints and multiple agents; it is not a re-derivation of the paper's own prior work, and no argument reduces to a self-citation. The sufficiency direction constructs a feature map α from a direction d(A), and the necessity direction uses the contrapositive with translated/scaled directions u + λd(A); neither step defines the target notion in terms of itself, nor fits a parameter to the quantity being predicted. The empirical section deliberately relaxes 'elicitable' to closeness in ℓ1 distance and reports confidence intervals, so it is not a case of a fitted input being relabeled as a prediction. The skeptical counterexample about Proposition D.6—that the fixed-direction equivalence may fail when a certificate requires both nonnegativity and conic multipliers—is a substantive correctness concern about a proof step, not a circularity: even if the proof is flawed, the theorem is not being asserted as its own input. Self-citations in the paper (e.g., Jagadeesan et al. [2023]) appear in related-work context and are not load-bearing for the characterization. Author-acknowledged modeling limitations in Section 6 (deterministic argmax agents, linear conic constraints, linear feature maps) are stated as scope restrictions, not as hidden assumptions of the target result. Accordingly, the derivation chain is self-contained against external benchmarks and no circular step can be exhibited.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The theory is parameter-free: only the illustrative examples choose q by hand, and the empirical instance design is hand-picked. The main theorems rest on convex-duality tools (Motzkin, minimax, KKT sufficiency) and on the domain assumptions of the principal-agent model. No new physical or algorithmic entities are introduced; the three mechanisms are definitions over the existing state space, not invented entities.

free parameters (2)
  • q in feature matrix α(q) = [[1,0,q],[0,1,q]] = 2 (Example 3.2); 0.6 (Example 3.4); 0.2 (Example 3.6)
    Hand-chosen to make the three illustrative examples (feasibility expansion, support expansion, binding set contraction) work; not fitted to data and not used in the general theorems.
  • Topic sets and prompt specifications in Section 5 empirical instances
    The three empirical instances (output topics, prompt topics, inclusion/exclusion prompt specs) are hand-picked to exhibit each mechanism; they are not the result of a systematic search or optimization, and the empirical conclusions depend on these choices.
axioms (5)
  • standard math Motzkin's transposition theorem (infeasibility certificate for linear inequalities)
    Used in Lemma D.4 (Appendix D.1) to produce the dual certificate that yields the eliciting reward function from the empty-intersection condition.
  • standard math Minimax theorem for convex-compact bilinear saddle problems
    Used in Proposition D.6 (Appendix D.4) to reduce the multi-constraint translation-existence problem to a worst-case weighted combination of constraints.
  • standard math KKT sufficiency for concave maximization over linear constraints
    Used in Lemma D.4 to certify optimality of x in the constructed reward-maximization program; requires concavity of f_j and linearity of the conic constraints.
  • domain assumption Agent behavior is deterministic argmax (x ∈ argmax of its reward program, ties allowed)
    Core to the definition of X*(R,E) in Section 2.2. Real LLMs are stochastic; the authors list this as a future extension in Section 6.
  • domain assumption Capability limitations are exactly linear conic constraints; prompt limitations are exactly α-composed monotone concave features
    Defines B_feasible and the feature map (Assumptions 2.1-2.3, Section 2.1-2.2). If the constraints or feature maps are nonlinear, the Motzkin/duality argument in Lemma D.4 fails. The authors acknowledge the linear restriction in Section 6.

pith-pipeline@v1.3.0-alltime-deepseek · 35273 in / 15360 out tokens · 138302 ms · 2026-08-02T20:59:50.858795+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Power and Limitations of Aggregation in Compound AI Systems." pith.science (2026). https://pith.science/paper/UXTB4VDL

@misc{pith2026260221556,
  author       = {Pith},
  title        = {Pith review of: Power and Limitations of Aggregation in Compound AI Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UXTB4VDL}},
  note         = {Machine review of arXiv:2602.21556}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

When designing compound AI systems, a common approach is to query multiple copies of the same model and aggregate the responses to produce a synthesized output. Given the homogeneity of these models, this raises the question of whether aggregation unlocks access to a greater set of outputs than querying a single model. In this work, we investigate the power and limitations of aggregation within a stylized principal-agent framework. This framework models how the system designer can partially steer each agent's output through its reward function specification, but still faces limitations due to prompt engineering ability and model capabilities. Our analysis uncovers three natural mechanisms -- feasibility expansion, support expansion, and binding set contraction -- through which aggregation expands the set of outputs that are elicitable by the system designer. We prove that any aggregation operation must implement one of these mechanisms in order to be elicitability-expanding, and that strengthened versions of these mechanisms provide necessary and sufficient conditions that fully characterize elicitability-expansion. Finally, we provide an empirical illustration of our findings for LLMs deployed in a toy reference-generation task. Altogether, our results take a step towards characterizing when compound AI systems can overcome limitations in model capabilities and in prompt engineering.

Figures

Figures reproduced from arXiv: 2602.21556 by Meena Jagadeesan, Nivasini Ananthakrishnan.

Figure 1
Figure 1. Figure 1: Three mechanisms by which the aggregation operation [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of output vectors for a reference-generation task (Section [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Empirical illustration of the three mechanisms in a toy reference-generation task with GPT-4o-mini [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. The Cost of Consensus: Isolated Self-Correction Prevails Over Unguided Homogeneous Multi-Agent Debate

    cs.MA 2026-04 unverdicted novelty 7.0

    Homogeneous multi-agent debate introduces sycophantic conformity, contextual fragility, and consensus collapse, leading to equal or lower accuracy than isolated self-correction at 2.1-3.4x higher token cost on GSM-Har...

Reference graph

Works this paper leans on

58 extracted references · 17 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Multiagent evaluation mechanisms

    Tal Alon, Magdalen Dobson, Ariel Procaccia, Inbal Talgam-Cohen, and Jamie Tucker-Foltz. Multiagent evaluation mechanisms. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 1774--1781, 2020

  2. [2]

    How we built our multi-agent research system

    Anthropic . How we built our multi-agent research system. https://www.anthropic.com/engineering/multi-agent-research-system, 2025. Blog post

  3. [3]

    Ask me anything: A simple strategy for prompting language models

    Simran Arora, Avanika Narayan, Mayee F Chen, Laurel Orr, Neel Guha, Kush Bhatia, Ines Chami, Frederic Sala, and Christopher R \'e . Ask me anything: A simple strategy for prompting language models. arXiv preprint arXiv:2210.02441, 2022

  4. [4]

    Compound ai systems

    BAIR Research Blog . Compound ai systems. https://bair.berkeley.edu/blog/2024/02/18/compound-ai-systems/, 2024. Blog post

  5. [5]

    Does the whole exceed its parts? the effect of ai explanations on complementary team performance

    Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel Weld. Does the whole exceed its parts? the effect of ai explanations on complementary team performance. In Proceedings of the 2021 CHI conference on human factors in computing systems, pages 1--16, 2021

  6. [6]

    Regression equilibrium

    Omer Ben-Porat and Moshe Tennenholtz. Regression equilibrium. In Proceedings of the 2019 ACM Conference on Economics and Computation, pages 173--191, 2019

  7. [7]

    Multitask principal--agent problems: Optimal contracts, fragility, and effort misallocation

    Philip Bond and Armando Gomes. Multitask principal--agent problems: Optimal contracts, fragility, and effort misallocation. Journal of Economic Theory, 144 0 (1): 0 175--211, 2009

  8. [8]

    Solutions Manual to Accompany Contract Theory, volume 1

    Arthur Campbell, Moshe Cohen, Florian Ederer, and Johannes Spinnewijn. Solutions Manual to Accompany Contract Theory, volume 1. The MIT Press, 2007

  9. [9]

    Frugalgpt: How to use large language models while reducing cost and improving performance

    Lingjiao Chen, Matei Zaharia, and James Zou. Frugalgpt: How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176, 2023

  10. [10]

    Direct preference optimization with unobserved preference heterogeneity: The necessity of ternary preferences

    Keertana Chidambaram, Karthik Vinary Seetharaman, and Vasilis Syrgkanis. Direct preference optimization with unobserved preference heterogeneity: The necessity of ternary preferences. arXiv preprint arXiv:2510.15716, 2025

  11. [11]

    Deep reinforcement learning from human preferences

    Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In Advances in Neural Information Processing Systems 30 (NeurIPS), volume 30, 2017

  12. [12]

    Emergent alignment via competition

    Natalie Collina, Surbhi Goel, Aaron Roth, Emily Ryu, and Mirah Shi. Emergent alignment via competition. arXiv preprint arXiv:2509.15090, 2025

  13. [13]

    Social choice should guide ai alignment in dealing with diverse human feedback

    Vincent Conitzer, Rachel Freedman, Jobst Heitzig, Wesley H Holliday, Bob M Jacobs, Nathan Lambert, Milan Moss \'e , Eric Pacuit, Stuart Russell, Hailey Schoelkopf, et al. Social choice should guide ai alignment in dealing with diverse human feedback. arXiv preprint arXiv:2404.10271, 2024

  14. [14]

    Or-bench: An over-refusal benchmark for large language models

    Justin Cui, Wei-Lin Chiang, Ion Stoica, and Cho-Jui Hsieh. Or-bench: An over-refusal benchmark for large language models. arXiv preprint arXiv:2405.20947, 2024

  15. [15]

    Mapping social choice theory to rlhf

    Jessica Dai and Eve Fleisig. Mapping social choice theory to rlhf. arXiv preprint arXiv:2404.13038, 2024

  16. [16]

    Incentive design with spillovers

    Krishna Dasaratha, Benjamin Golub, and Anant Shah. Incentive design with spillovers. arXiv preprint arXiv:2411.08026, 2024

  17. [17]

    Ensemble methods in machine learning

    Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1--15. Springer, 2000

  18. [18]

    Impacts of aggregation on model diversity and consumer utility

    Kate Donahue and Manish Raghavan. Impacts of aggregation on model diversity and consumer utility. Preprint, 2026

  19. [19]

    Human-algorithm collaboration: Achieving complementarity and avoiding unfairness

    Kate Donahue, Alexandra Chouldechova, and Krishnaram Kenthapadi. Human-algorithm collaboration: Achieving complementarity and avoiding unfairness. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1639--1656, 2022

  20. [20]

    Improving factuality and reasoning in language models through multiagent debate

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. In International Conference on Machine Learning, pages 11733--11763. PMLR, 2024

  21. [21]

    A market for accuracy: Classification under competition

    Ohad Einav and Nir Rosenfeld. A market for accuracy: Classification under competition. In Forty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025 , volume 267 of Proceedings of Machine Learning Research. PMLR / OpenReview.net, 2025

  22. [22]

    Bayesian persuasion with multiple senders and rich signal spaces

    Matthew Gentzkow and Emir Kamenica. Bayesian persuasion with multiple senders and rich signal spaces. Games and Economic Behavior, 104: 0 411--429, 2017

  23. [23]

    Distortion of ai alignment: Does preference optimization optimize for preferences? arXiv preprint arXiv:2505.23749, 2025

    Paul G \"o lz, Nika Haghtalab, and Kunhe Yang. Distortion of ai alignment: Does preference optimization optimize for preferences? arXiv preprint arXiv:2505.23749, 2025

  24. [24]

    An analysis of the principal-agent problem

    Sanford J Grossman and Oliver D Hart. An analysis of the principal-agent problem. Econometrica, 51 0 (1): 0 7--46, 1983

  25. [25]

    Moral hazard and observability

    Bengt Holmstr \"o m. Moral hazard and observability. The Bell journal of economics, pages 74--91, 1979

  26. [26]

    Moral hazard in teams

    Bengt Holmstrom. Moral hazard in teams. The Bell journal of economics, pages 324--340, 1982

  27. [27]

    Multitask principal--agent analyses: Incentive contracts, asset ownership, and job design

    Bengt Holmstrom and Paul Milgrom. Multitask principal--agent analyses: Incentive contracts, asset ownership, and job design. The Journal of Law, Economics, and Organization, 7 0 (special\_issue): 0 24--52, 1991

  28. [28]

    Strategic classification with externalities

    Safwan Hossain, Evi Micha, Yiling Chen, and Ariel Procaccia. Strategic classification with externalities. arXiv preprint arXiv:2410.08032, 2024

  29. [29]

    Is best-of-n the best of them? coverage, scaling, and optimality in inference-time alignment

    Audrey Huang, Adam Block, Qinghua Liu, Nan Jiang, Akshay Krishnamurthy, and Dylan J Foster. Is best-of-n the best of them? coverage, scaling, and optimality in inference-time alignment. In Forty-second International Conference on Machine Learning, 2025

  30. [30]

    The consensus game: Language model generation via equilibrium search

    Athul Paul Jacob, Yikang Shen, Gabriele Farina, and Jacob Andreas. The consensus game: Language model generation via equilibrium search. arXiv preprint arXiv:2310.09139, 2023

  31. [31]

    Improved bayes risk can yield reduced social welfare under competition

    Meena Jagadeesan, Michael Jordan, Jacob Steinhardt, and Nika Haghtalab. Improved bayes risk can yield reduced social welfare under competition. Advances in Neural Information Processing Systems, 36: 0 66940--66952, 2023

  32. [32]

    Adversaries can misuse combinations of safe models

    Erik Jones, Anca Dragan, and Jacob Steinhardt. Adversaries can misuse combinations of safe models. In Forty-second International Conference on Machine Learning, 2025 a

  33. [33]

    Uncovering gaps in how humans and llms interpret subjective language

    Erik Jones, Arjun Patrawala, and Jacob Steinhardt. Uncovering gaps in how humans and llms interpret subjective language. In The Thirteenth International Conference on Learning Representations, 2025 b

  34. [34]

    Debating with more persuasive llms leads to more truthful answers

    Akbir Khan, John Hughes, Dan Valentine, Laura Ruis, Kshitij Sachan, Ansh Radhakrishnan, Edward Grefenstette, Samuel R Bowman, Tim Rockt \"a schel, and Ethan Perez. Debating with more persuasive llms leads to more truthful answers. In International Conference on Machine Learning, pages 23662--23733. PMLR, 2024

  35. [35]

    How do classifiers induce agents to invest effort strategically? ACM Transactions on Economics and Computation (TEAC), 8 0 (4): 0 1--23, 2020

    Jon Kleinberg and Manish Raghavan. How do classifiers induce agents to invest effort strategically? ACM Transactions on Economics and Computation (TEAC), 8 0 (4): 0 1--23, 2020

  36. [36]

    The condorcet jury theorem, free speech, and correlated votes

    Krishna K Ladha. The condorcet jury theorem, free speech, and correlated votes. American Journal of Political Science, pages 617--634, 1992

  37. [37]

    The Theory of Incentives: The Principal-Agent Model

    Jean-Jacques Laffont and David Martimort. The Theory of Incentives: The Principal-Agent Model. Princeton University Press, Princeton, NJ, 2002

  38. [38]

    Rank-order tournaments as optimum labor contracts

    Edward P Lazear and Sherwin Rosen. Rank-order tournaments as optimum labor contracts. Journal of political Economy, 89 0 (5): 0 841--864, 1981

  39. [39]

    Alpacaeval: An automatic evaluator of instruction-following models, 2023

    Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Alpacaeval: An automatic evaluator of instruction-following models, 2023

  40. [40]

    Strategic ranking

    Lydia T Liu, Nikhil Garg, and Christian Borgs. Strategic ranking. In International Conference on Artificial Intelligence and Statistics, pages 2489--2518. PMLR, 2022

  41. [41]

    Recast: Strengthening llms' complex instruction following with constraint-verifiable data

    Wenhao Liu, Zhengkang Guo, Mingchen Xie, Jingwen Xu, Zisu Huang, Muzhao Tian, Jianhan Xu, Muling Wu, Xiaohua Wang, Changze Lv, et al. Recast: Strengthening llms' complex instruction following with constraint-verifiable data. arXiv preprint arXiv:2505.19030, 2025

  42. [42]

    Distributed algorithms

    Nancy A Lynch. Distributed algorithms. Elsevier, 1996

  43. [43]

    Ai alignment and social choice: Fundamental limitations and policy implications

    Abhilash Mishra. Ai alignment and social choice: Fundamental limitations and policy implications. arXiv preprint arXiv:2310.16048, 2023

  44. [44]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 27730--27744, 2022

  45. [45]

    Citeme: Can language models accurately cite scientific claims? Advances in Neural Information Processing Systems, 37: 0 7847--7877, 2024

    Ori Press, Andreas Hochlehnert, Ameya Prabhu, Vishaal Udandarao, Ofir Press, and Matthias Bethge. Citeme: Can language models accurately cite scientific claims? Advances in Neural Information Processing Systems, 37: 0 7847--7877, 2024

  46. [46]

    Competition and diversity in generative ai

    Manish Raghavan. Competition and diversity in generative ai. arXiv preprint arXiv:2412.08610, 2024

  47. [47]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019

  48. [48]

    Direct alignment with heterogeneous preferences

    Ali Shirali, Arash Nasr-Esfahany, Abdullah Alomar, Parsa Mirtaheri, Rediet Abebe, and Ariel Procaccia. Direct alignment with heterogeneous preferences. arXiv preprint arXiv:2502.16320, 2025

  49. [49]

    A mathematical abstraction for balancing the trade-off between creativity and reality in large language models

    Ritwik Sinha, Zhao Song, and Tianyi Zhou. A mathematical abstraction for balancing the trade-off between creativity and reality in large language models. arXiv preprint arXiv:2306.02295, 2023

  50. [50]

    Multitask agency and contract choice: An empirical exploration

    Margaret E Slade. Multitask agency and contract choice: An empirical exploration. International Economic Review, pages 465--486, 1996

  51. [51]

    Multi-agent reinforcement learning: Independent vs

    Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning, pages 330--337, 1993

  52. [52]

    Scibench: Evaluating college-level scientific problem-solving abilities of large language models

    Xiaoxuan Wang, Ziniu Hu, Pan Lu, Yanqiao Zhu, Jieyu Zhang, Satyen Subramaniam, Arjun R Loomba, Shichang Zhang, Yizhou Sun, and Wei Wang. Scibench: Evaluating college-level scientific problem-solving abilities of large language models. arXiv preprint arXiv:2307.10635, 2023 a

  53. [53]

    Self-consistency improves chain of thought reasoning in language models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations, 2023 b

  54. [54]

    Benchmarking complex instruction-following with multiple constraints composition

    Bosi Wen, Pei Ke, Xiaotao Gu, Lindong Wu, Hao Huang, Jinfeng Zhou, Wenchuang Li, Binxin Hu, Wendy Gao, Jiaxing Xu, et al. Benchmarking complex instruction-following with multiple constraints composition. Advances in Neural Information Processing Systems, 37: 0 137610--137645, 2024

  55. [55]

    Heterogeneous data game: Characterizing the model competition across multiple data sources

    Renzhe Xu, Kang Wang, and Bo Li. Heterogeneous data game: Characterizing the model competition across multiple data sources. arXiv preprint arXiv:2505.07688, 2025

  56. [56]

    What prompts don't say: Understanding and managing underspecification in llm prompts

    Chenyang Yang, Yike Shi, Qianou Ma, Michael Xieyang Liu, Christian K \"a stner, and Tongshuang Wu. What prompts don't say: Understanding and managing underspecification in llm prompts. arXiv preprint arXiv:2505.13360, 2025

  57. [57]

    Cot-based synthesizer: Enhancing llm performance through answer synthesis, 2025

    Bohan Zhang, Xiaokang Zhang, Jing Zhang, Jifan Yu, Sijia Luo, and Jie Tang. Cot-based synthesizer: Enhancing llm performance through answer synthesis, 2025

  58. [58]

    Consequences of misaligned ai

    Simon Zhuang and Dylan Hadfield-Menell. Consequences of misaligned ai. Advances in Neural Information Processing Systems, 33: 0 15763--15773, 2020