{"id":"f41d6348-6195-4055-9b77-ec350627a1df","arxiv_id":"2505.13022","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper introduces a solution concept where players' analogy categories and equilibrium strategies are jointly determined by K-means style clustering, and shows when mixed categorization or multiple categorizations must arise.","lead":"Many economic models assume players form detailed expectations about every situation. This paper lets players group situations into a small number of categories, chosen to minimize prediction errors, and studies what happens when their strategies and categories shape each other.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 5's monitoring equilibrium fails for generic parameter values: cost equality forces a c-mixing probability different from 1/2, and within the stated ν* range the employer's best response can be D rather than C.","rationale":"The paper's central existence theorem (Theorem 1) is supported by a plausible Kakutani fixed point argument; I did not find a disabling gap there. The weakest load-bearing point is the monitoring application: Proposition 5's closed-form equilibrium is not consistent with the paper's own definition on a substantial part of the stated parameter region. This is exactly the kind of specific, fixable defect that warrants a conditional verdict rather than rejection. The reader's concern is in the right place but the precise diagnosis is narrower than the real problem: even with p_a≠p_b, one can equalize clustering costs by choosing the c-worker's mixing probability appropriately; the deeper failure is that the cost-equalizing q then conflicts with the employer's best-response thresholds. Hence agreement_with_reader=partial and the verdict remains CONDITIONAL.","tokens_in":30214,"tokens_out":18791,"duration_ms":200223,"concrete_test":"Set p_a=0.01, p_b=0.98, p_c=0.01 and ν*=0.9, and check the profile described in Proposition 5 against Definition 5. Cost equality for partitions ac and bc forces q≈0.585, while the employer's C-optimality in class ac requires expected low effort (0.01+0.01(1-q))/0.02 ≥ 0.9, i.e. q≤0.2; the two conditions are incompatible, so the claim fails for parameters satisfying the proposition's hypotheses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.2, Proposition 5 claims that for any p_a,p_b,p_c with p_c no larger than p_a and p_b and p_c/(p_a+p_c) < ν* < p_b/(p_b+p_c), the unique globally clustered CD-ABEE puts ac and bc together with probabilities μ* and 1-μ*, has the c-worker mix e=0/e=1 with probability 1/2, and has the employer choose C in any class containing a and D in any class containing b. Let A=p_a, B=p_b, C=p_c and let q=Pr(e=1|c). With squared Euclidean distance, the global clustering costs of the two candidate partitions are Cost_ac = A C q^2/(A+C) and Cost_bc = B C (1-q)^2/(B+C). Both partitions can be in the support only when these costs are equal, which forces q/(1-q)=sqrt(B(A+C)/(A(B+C))), not q=1/2 unless A=B. In addition, the claimed employer strategy requires (A+C(1-q))/(A+C) ≥ ν* and C(1-q)/(B+C) < ν*. These inequalities are not implied by the hypothesis. Example: A=0.01, B=0.98, C=0.01, ν*=0.9 satisfies the stated bounds, but the cost-equalizing q≈0.585 gives expected low effort in ac ≈0.708 < 0.9, so the employer strictly prefers D in class ac. The proposed profile therefore violates Definition 5. The qualitative mixing-over-partitions insight may survive a corrected construction, but the uniqueness and characterization result in the flagship application must be revised.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces a solution concept for two-player normal-form games in which each player partitions the set of games into K analogy classes so as to minimize prediction error about the opponent's behavior, and strategies form an analogy-based expectation equilibrium given those partitions. The authors define locally and globally clustered ABEE and a distributional extension (CD-ABEE) that allows mixing over partitions. They prove (Theorem 1) that a CD-ABEE always exists in finite environments for squared Euclidean or KL divergence, and provide two learning models whose steady states correspond to these equilibria. Applications include beauty contests (self-attractive partitions, many equilibria), a monitoring game (self-repelling partitions, mixing over partitions), and linear best-response games with strategic complements/substitutes. The paper concludes with a discussion of belief heterogeneity and equilibrium multiplicity.","tokens_in":30511,"tokens_out":18924,"duration_ms":171359,"significance":"The conceptual contribution is valuable: endogenizing analogy partitions via clustering connects behavioral game theory with prototype theory and K-means, and the distinction between self-attractive and self-repelling partitions offers a new channel for multiplicity and heterogeneity. The existence proof for CD-ABEE is a genuine technical contribution and appears sound, and the learning foundations give the solution concept behavioral content. However, the monitoring application's main result (Proposition 5) is false for generic parameter values, and the proof of Proposition 1 in the appendix contains an error. These problems are fixable but currently preclude acceptance.","major_comments":[{"comment":"The profile described in Proposition 5 cannot be a globally clustered CD-ABEE when p_a differs from p_b. With d the squared Euclidean distance and q defined as Pr(e=1|c), the global clustering costs of the two candidate partitions are V(ac)=p_a p_c q^2/(p_a+p_c) and V(bc)=p_b p_c (1-q)^2/(p_b+p_c). Both partitions can belong to the support of the equilibrium distribution only if V(ac)=V(bc), which forces q/(1-q) = sqrt(p_b(p_a+p_c)/(p_a(p_b+p_c))); this ratio equals 1 (so q=1/2) only when p_a=p_b. The proposition assumes only p_c no larger than p_a and p_b, so for generic type frequencies the proposed q=1/2 makes the two partitions unequal, violating Definition 5. In addition, the employer's prescribed actions require (p_a+p_c(1-q))/(p_a+p_c) at least nu* in class ac and p_c(1-q)/(p_b+p_c) below nu* in class bc, which are not implied by the stated bounds on nu*. For example, with p_a=0.01, p_b=0.98, p_c=0.01 and nu*=0.9, the cost-equalizing q about 0.585 gives expected low effort in ac about 0.708, below 0.9, so the employer strictly prefers D in ac. Thus the uniqueness claim and the characterization in the monitoring application fail for generic parameter values; the statement and proof need to be revised.","section":"Section 3.2, Proposition 5"},{"comment":"The case analysis appears to contain an error. The proof states that 'if sigma1(x1)=D, then sigma2(x1)=R', but in the matching pennies game of Example 1, if Row plays D, Column's payoff is 1 from L and 0 from R, so Column's best response is L, not R. This invalidates the subsequent inference that beta_L is at least 1/2 and the conclusion that any ABEE requires beta_L = 1/(2+x1). The non-existence claim may well be true, but the proof as written does not establish it.","section":"Appendix, proof of Proposition 1"},{"comment":"The proof's first case asserts that from beta_{e=1} at most p_c/(p_a+p_c) and p_c/(p_a+p_c) < nu* it follows that the employer chooses C. Since the employer chooses C only when the expected probability of low effort is at least nu*, i.e., beta_{e=1} at most 1-nu*, this inference requires p_c/(p_a+p_c) at most 1-nu*, which is not among the assumptions. For nu* greater than 1/2 the stated bound does not suffice. The contradiction may be recoverable by a subcase analysis, but the proof as given is incomplete.","section":"Section 3.2, Proposition 4"}],"minor_comments":[{"comment":"The name 'Clustered Distributional Analogy-Based Expected Equilibrium' should be 'Expectation Equilibrium' for consistency with the ABEE terminology used elsewhere in the paper.","section":"Definition 5"},{"comment":"The domain of the GC correspondence is written as 'Sigma-bar times ((Delta A_i)^{K_i|K_i|} times (Delta A_j)^{K_j|K_j|})', which appears to be a typographical error; it should presumably be the product of the relevant simplex spaces for each of the |K_i| and |K_j| partitions.","section":"Appendix, proof of Theorem 1"},{"comment":"The expression 'probability 2/2+x1' should read '2/(2+x1)'.","section":"Appendix, proof of Proposition 1"},{"comment":"The claim that steady states correspond to locally clustered CD-ABEE is asserted rather than proved; the text says an explicit study 'would require further work.' Since this is one of the two learning foundations for the solution concept, a formal statement or a clearer caveat would be desirable.","section":"Section 2.6, Learning model 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's central theoretical result (Theorem 1) is sound, and the conceptual framework is interesting. However, the monitoring application contains a false proposition, and the proof of Proposition 1 has an error; both appear in the current manuscript and need correction. I recommend major revision because the problems are local and fixable, and the paper's core contribution (the existence theorem plus the distinction between self-attractive and self-repelling partitions) does not depend on the particular characterization in Proposition 5."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Philippe, two things you should know. First, the central concept is real: CABEE and CD-ABEE endogenize analogy partitions by having players cluster opponent behavior, and the existence theorem (Theorem 1) via Kakutani is standard and looks sound. The self-attractive vs self-repelling taxonomy is a useful way to organize when mixing over partitions is needed, and the learning foundations in the online appendix are a genuine effort, not a gesture. Second, the monitoring application in Section 3.2 is not correct as stated. The stress-test checks out: Proposition 5 claims that with p_c ≤ p_a, p_b and p_c/(p_a+p_c) < ν* < p_b/(p_b+p_c), the unique globally clustered CD-ABEE has the c-worker mixing 1/2 and the employer choosing C in any class containing a. But global clustering costs of the two candidate partitions are A C q^2/(A+C) and B C (1-q)^2/(B+C) (with q = Pr(e=1|c)). Both partitions can be in the support only if these are equal, which forces q/(1-q) = sqrt(B(A+C)/(A(B+C))), so q=1/2 only when p_a = p_b. Moreover, the employer's C/D threshold is not implied by the stated bound: take p_a=0.01, p_b=0.98, p_c=0.01, ν*=0.9; the cost-equalizing q≈0.585 gives expected low effort in ac ≈0.708 < 0.9, so the employer strictly prefers D in that class. So the characterization and uniqueness claims in the paper's main applied section fail for generic parameters. The qualitative point—that self-repelling partitions force mixing—may survive with a corrected construction, but the proposition and its proof must be rewritten, and the parameter ranges re-derived. The paper's other applications look more solid: the beauty contest result that high coordination makes almost any partition self-attractive is intuitive and the proof is short, and the linear best-response section has a nice complement/substitute contrast. The citation pattern is fine; the self-citations to the working paper are appropriate since the full characterization lives there. Overall: this is a paper with a valuable new idea and a serious flaw in one application. It deserves peer review, but a referee should not accept it until the monitoring section is fixed. I'd cite the framework if I were working on coarse expectations, but I would not cite Proposition 5 in its current form.","headline":"A genuinely new idea—endogenous analogy partitions via clustering—with a clean existence theorem, but the flagship monitoring application is built on a false proposition and needs major revision.","tokens_in":31044,"tokens_out":5399,"would_cite":true,"duration_ms":49832,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A10","91A26","62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that when players group games into analogy classes to minimize prediction error, an equilibrium over strategies and partitions always exists if players can randomize over partitions.","keywords":["Analogy-Based Expectation Equilibrium","Prototype theory","K-means clustering","Endogenous categorization","Coarse beliefs","Equilibrium multiplicity","Self-attractive vs self-repelling partitions","Bounded rationality"],"falsifier":"Run the monitoring game with unequal majority probabilities—say $p_a=0.4$, $p_b=0.3$, $p_c=0.3$—and pick $\\nu^*=0.45$, which lies inside $(p_c/(p_a+p_c),\\, p_b/(p_b+p_c))$. Compute the global clustering costs of the two two-element partitions $\\{a,c\\}$ and $\\{b,c\\}$ under the strategies described in Proposition 5; if the costs are not equal, no CD-ABEE can put positive probability on both, which would contradict the paper's monitoring conclusion.","tokens_in":29987,"feed_emoji":"🧩","tokens_out":18006,"duration_ms":168538,"temperature":0.7,"pith_summary":"This paper asks what happens when players cannot keep separate expectations for every possible game and instead group games into a small number of analogy classes, chosen to minimize how wrong their opponent predictions are. The authors define clustered analogy-based expectation equilibria and prove that, in any finite two-player normal-form environment, an equilibrium always exists once players are allowed to randomize over which analogy partition they use. They also show that pure clustered equilibria can fail, and that environments can be self-attractive (many partitions are self-sustaining) or self-repelling (players must mix over partitions). Applications to beauty contests, employer-worker monitoring, and Bertrand/Cournot duopolies suggest that this endogeneity of categories is a new channel for equilibrium multiplicity and for heterogeneous beliefs within a society.","feed_headline":"Mixed analogy categories guarantee equilibrium in finite games","feed_subtitle":"Clustering expectations into categories has a fixed point; some games force mixed beliefs.","key_machinery":"The load-bearing object is the clustered distributional analogy-based expectation equilibrium, built from two fixed-point requirements. On the strategic side, each player best-responds to coarse beliefs $\\beta_i(\\alpha)$ about the opponent's play in each analogy class $\\alpha$, and $\\beta_i(\\alpha)$ is the probability-weighted mean of the opponent's actual play in $\\alpha$. On the clustering side, an analogy partition is locally clustered if every game is assigned to the class whose mean is nearest, and globally clustered if it minimizes total prediction error; the mean is the optimal prototype because both distances are Bregman divergences, for which the conditional mean is the best predictor. Theorem 1 composes the best-response correspondence and the global-clustering correspondence into a single upper-hemicontinuous, convex-valued mapping and applies Kakutani's fixed point theorem.","core_discovery":"The paper's central claim is Theorem 1: in finite environments, with prediction error measured by squared Euclidean distance or Kullback-Leibler divergence, there always exists a locally and a globally clustered distributional analogy-based expectation equilibrium—a profile of mixed strategies and a distribution over analogy partitions such that each strategy is a best response to coarse beliefs given its partition, and every partition in the support minimizes the opponent-prediction error given the aggregate play. The same endogeneity that makes pure clustered equilibria fail in some three-matching-pennies environments is what the distributional extension absorbs: when the data being clustered are themselves produced by the clustering, a single partition may be self-defeating, and only mixing over partitions can be self-consistent. The paper then classifies environments by whether analogy partitions are self-attractive, sustaining many equilibria (beauty contests with strong coordination, strategic complements), or self-repelling, forcing mixing and heterogeneous beliefs (the monitoring game, strategic substitutes).","pith_inferences":["A testable extension: in a laboratory beauty contest with the coordination weight $r$ near 1, different groups facing identical fundamentals should settle on different category boundaries and different action distributions, whereas with $r$ near 0 they should converge to the variance-minimizing equal-split partition.","The self-attractive/self-repelling dichotomy suggests a policy lever: interventions that change the interaction parameter could shift a market between multiple stable categorization regimes and regimes where categorization heterogeneity is unavoidable.","The paper's continuum applications (beauty contest and linear best-response families) fall outside the finite-environment existence theorem; they are handled by direct construction, so a general existence theorem for continuum games remains an open extension, not established here.","If clustering were allowed to use additional attributes such as the player's own payoff structure or the state, the equilibrium set and the self-attractive/self-repelling classification would in general change; the paper notes this but does not develop it."],"forward_implications":["In any finite two-player game environment with finitely many actions, a clustered distributional ABEE exists for both squared Euclidean and KL divergence, so endogenous analogy classes are consistent with equilibrium existence.","Some finite environments (three matching pennies games with two classes for one player and three for the other) admit no pure clustered ABEE; equilibrium then requires players in the same role to use different analogy partitions despite seeing the same data.","Environments with self-attractive partitions—beauty contests with a high coordination motive and strategic complements—can support many partitions and correspondingly many equilibrium behaviors, a new channel for cross-society differences.","Environments with self-repelling partitions—the monitoring game and strategic substitutes—force mixing over partitions; in the monitoring game this produces polarized beliefs about a minority type, with some employers treating it like the always-shirking type and others like the always-working type."],"supporting_citations":[{"why":"Defines analogy-based expectation equilibrium, the strategic backbone of the solution concept.","marker":"Jehiel (2005)"},{"why":"Extends ABEE to private-information environments, the normal-form foundation used here.","marker":"Jehiel and Koessler (2008)"},{"why":"Presents ABEE with distributions over analogy partitions, motivating the distributional extension.","marker":"Jehiel (2022)"},{"why":"Shows the conditional mean is the optimal predictor for Bregman divergences, justifying the mean prototype for both distances.","marker":"Banerjee et al. (2005)"},{"why":"Supplies the Kakutani fixed-point theorems used in the proof of Theorem 1.","marker":"Border (1985)"},{"why":"Provides the Berk-Nash equilibrium with mixtures over misspecified models, the comparison point for mixing over partitions.","marker":"Esponda and Pouzo (2016)"}],"fun_headline_variants":["Mixing analogy partitions always yields equilibrium","Distribution over analogy classes fixes equilibrium existence","Coarse beliefs with mixed categories always stabilize","Even self-defeating analogies admit mixed equilibrium","Endogenous clustering needs distributional equilibrium"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The monitoring-game claim that heterogeneous partitions must coexist relies on the two majority worker types being exactly equally likely; if $p_a \\neq p_b$, the two candidate partitions cannot both minimize the employer's prediction error, and the described mixed-partition equilibrium disappears.","fun_headline_variants_meta":{"raw":{"variants":["Mixing analogy partitions always yields equilibrium","Distribution over analogy classes fixes equilibrium existence","Coarse beliefs with mixed categories always stabilize","Even self-defeating analogies admit mixed equilibrium","Endogenous clustering needs distributional equilibrium"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1232,"prompt_tokens":812,"completion_tokens":420,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":428,"completion_tokens_details":{"reasoning_tokens":356}},"tokens_in":428,"tokens_out":420,"duration_ms":4640,"temperature":1.0,"reasoning_tokens":356,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:22:44.048658+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the monitoring game with unequal majority probabilities—say $p_a=0.4$, $p_b=0.3$, $p_c=0.3$—and pick $\\nu^*=0.45$, which lies inside $(p_c/(p_a+p_c),\\, p_b/(p_b+p_c))$. Compute the global clustering costs of the two two-element partitions $\\{a,c\\}$ and $\\{b,c\\}$ under the strategies described in Proposition 5; if the costs are not equal, no CD-ABEE can put positive probability on both, which would contradict the paper's monitoring conclusion.","supporting_citations":[{"cited_title":"(2005): ``Analogy-based expectation equilibrium,'' Journal of Economic Theory 123, 81--104","cited_arxiv_id":null,"evidence_quote":"Defines analogy-based expectation equilibrium, the strategic backbone of the solution concept."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Extends ABEE to private-information environments, the normal-form foundation used here."},{"cited_title":"(2022): ``Analogy-based expectation equilibrium and related concepts: Theory, applications, and beyond,'' Prepared for the twelfth World Congress of the Econometric Society","cited_arxiv_id":null,"evidence_quote":"Presents ABEE with distributions over analogy partitions, motivating the distributional extension."},{"cited_title":"Guo and H","cited_arxiv_id":null,"evidence_quote":"Shows the conditional mean is the optimal predictor for Bregman divergences, justifying the mean prototype for both distances."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Kakutani fixed-point theorems used in the proof of Theorem 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Berk-Nash equilibrium with mixtures over misspecified models, the comparison point for mixing over partitions."}],"review_version":1}