{"id":"54795610-9656-496e-a628-58288ba395a5","arxiv_id":"2506.08292","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"ECON trains small belief networks to set each LLM's sampling parameters for coordinated reasoning, reporting better accuracy with fewer tokens than debate-based baselines.","lead":"This paper introduces ECON, a multi-LLM framework that replaces token-heavy debate with belief-based coordination: each model tunes its own sampling temperature and repetition penalty to maximize expected reward, and a coordinator produces the final answer. If the reported gains are real, it offers a cheaper way to combine several LLMs for reasoning and planning tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The BNE-convergence proof in Appendix B.2 is circular (Step 3 assumes Bellman optimality to prove it), and Appendix C.2's regret rates rely on an asserted O(1/√t) Q-learning rate even though η_t=η_0/√t violates the cited Robbins-Monro condition. The central theoretical claim is not established.","rationale":"The reader's reject verdict is correct. The paper's abstract promises a mathematical proof of a tighter regret bound and convergence to BNE; the proof in Appendix B.2 obtains convergence to a stationary point of a TD loss and then assumes the only stationary point is Bellman-optimal, which is exactly the missing result. This is not a matter of choosing different standard assumptions; it is a logical gap in the presented derivation. The regret analysis in Appendix C.2 inherits the gap and additionally uses a learning-rate schedule that contradicts the stochastic-approximation condition it cites. The comparison to debate in Appendix C.3 also models MAD as zero-sum, which does not match the cooperative debate baselines evaluated in the experiments, so the claimed linear lower bound for 'non-equilibrium schemes' is not established for the actual competitors. I would not accuse the authors of any dishonesty; the empirical framework may be useful, and the public code and detailed hyperparameters are a positive. But the central theoretical claims are load-bearing and fail on internal grounds, so the appropriate disposition is reject. A revision that removes or substantially weakens the theoretical guarantees and reframes the paper as an empirical method with ablations could be reconsidered.","tokens_in":42604,"tokens_out":8198,"duration_ms":109676,"concrete_test":"Run a minimal two-agent, two-action tabular game with a known unique BNE, using the paper's MLP belief-network and TD-loss update with simultaneous agent updates. After convergence, check whether the stationary point satisfies exact Bellman optimality and whether the joint policy is a BNE. If the converged Q-functions are only projected fixed points, or if each agent optimizes against a frozen opponent while the joint profile is not an equilibrium, Appendix B.2 Step 3's 'only when' condition is false. Also compute Σ η_t² for η_t=η_0/√t to confirm that the cited Robbins-Monro condition is violated.","verdict_should_be":"REJECT","load_bearing_attack":"The single most load-bearing concern is internal to the theory, not a disagreement with consensus. In Appendix B.2, Step 3, after showing that SGD reaches a stationary point θ*_i of the TD loss, the proof stipulates that the Q-network parameterization is such that the zero-gradient condition holds only when Q_i(s,a;θ*_i) = r_i + γ max_{a'} Q_i(s',a';θ_i^-). That is the Bellman optimality equation, i.e., the very conclusion needed. For the MLP belief network actually used, a zero expected TD gradient characterizes a projected Bellman fixed point, not global optimality, and the claim is neither proved nor measured. Step 4 then asserts Bellman optimality at the stationary point, and Step 5 concludes every agent is best-responding and hence the system is at a BNE. Even if each Q-network were optimal against frozen opponents, this does not imply the simultaneous best-response system that defines BNE, because opponent policies are non-stationary during training. This gap is load-bearing because Appendix C.2 defines V* as the BNE value and bounds regret against it; without BNE convergence, the regret decomposition has no valid benchmark. Independently, Appendix C.2.3 sets η_t=η_0/√t while invoking Robbins-Monro conditions that require Σ η_t² < ∞; this schedule has divergent Σ η_t², so the cited convergence theorem does not apply as stated. The claimed O(N√T/(1−γ)) bound is therefore unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ECON, a hierarchical multi-LLM reasoning framework that replaces explicit inter-agent debate with belief-based coordination. Each execution LLM maintains a belief network that outputs a temperature and repetition penalty, and a centralized mixing network coordinates these local components. The authors claim to formalize multi-LLM coordination as a Bayesian Nash equilibrium problem, to prove BNE existence and convergence, to derive a sublinear regret bound O(N sqrt(T)/(1-gamma)) for ECON versus linear regret for multi-agent debate, and to show consistent empirical gains across six benchmarks with reduced token usage. The appendix contains the theoretical proofs, assumptions, experimental details, and tutorial-style examples.","tokens_in":43001,"tokens_out":4390,"duration_ms":53278,"significance":"If the theoretical and empirical claims were both established, this would be a valuable contribution: a communication-free coordination mechanism for LLM ensembles with convergence guarantees and measured efficiency gains. The paper has notable strengths: the code is public, the experiments cover several models and datasets, the scalability study is informative, and the ablation studies test some design choices. However, the central theoretical results are not supported as stated. The BNE convergence proof is circular, the regret analysis relies on an invalid learning-rate schedule and unverified regularity assumptions, and the BNE existence proof imposes conditions that do not hold for the implemented discrete-text LLM system. The headline empirical improvement is also not backed by the reported aggregate numbers. The significance of the paper therefore rests on claims that the manuscript does not justify.","major_comments":[{"comment":"The convergence-to-BNE proof assumes that a zero gradient of the TD loss holds only when Q_i(s,a;theta*_i) = r_i + gamma max_{a'} Q_i(s',a';theta^-_i), and then uses this Bellman optimality condition to conclude that agents are best-responding. This assumption is essentially the conclusion being proved. For the neural Q-networks and belief networks actually used, a stationary point of the TD objective characterizes a projected Bellman fixed point, not global optimality, and the paper provides neither proof nor measurement that the stronger condition holds. Consequently the claimed convergence to BNE is not established, and the regret decomposition in Appendix C.2, which benchmarks against V* under the BNE, lacks a valid target.","section":"Appendix B.2, Steps 3-5"},{"comment":"The proof of the O(N sqrt(T)/(1-gamma)) bound invokes Robbins-Monro conditions (sum eta_t = infinity and sum eta_t^2 < infinity) while choosing eta_t = eta_0/sqrt(t); this schedule has divergent sum eta_t^2, so the cited stochastic approximation theorem does not apply as stated. Moreover, the O(1/sqrt(t)) rates for Q-function estimation and policy suboptimality are asserted for belief networks that output only temperature and repetition penalty while the LLM produces discrete text; no argument connects neural-network TD learning or online convex optimization in the action-embedding space to the stochastic LLM generation process. The central regret bound is therefore unsupported.","section":"Appendix C.2.3, Lemmas C.1-C.2"},{"comment":"The BNE existence proof assumes payoff functions are continuous and quasi-concave in actions, but in the implemented system actions are prompt embeddings mapped through an LLM to discrete text, and rewards are computed on text outputs; continuity and quasi-concavity are not established and are implausible for this mapping. The proof's statement that 'strategies are linear in the space of mixed strategies' does not imply quasi-concavity of the payoff. Thus the existence theorem is not proven for the stated setting.","section":"Appendix B.1, Theorem 2.1"},{"comment":"The regret and convergence claims depend on assumptions B.4, B.6, and B.8 (approximate posterior alignment, game regularity, and concentrability), none of which are measured, derived from the ECON architecture, or otherwise justified. Similarly, the comparison with multi-agent debate in Appendix C.3 models debate as zero-sum with persistent suboptimality, but this model is not shown to describe the MAD baseline used in the experiments. The claimed linear-versus-sublinear gap is therefore not established.","section":"Appendix B.3 and Appendix C.3"},{"comment":"The headline claim that ECON outperforms existing multi-LLM approaches by 11.2% on average across six benchmarks is not supported by the numbers reported in Section 4.2: the listed improvements of 25.6%, 6.3%, 10.9%, 11.2%, and 6.4% are against different baselines (zero-shot CoT, few-shot CoT, SC, ToT, rStar), not an average over existing multi-LLM approaches, and no aggregate comparison with error bars is provided. The benchmark count is also inconsistent: the paper says six benchmarks but reports five reasoning datasets plus TravelPlanner.","section":"Abstract and Section 4.2"}],"minor_comments":[{"comment":"The symbol theta_i is used both for agent types in the BNE proof and for network parameters elsewhere; this collision makes several formulas ambiguous and should be fixed.","section":"Appendix A and Section 2.2"},{"comment":"The method label 'EcoNash' in Figure 3 differs from the name ECON used throughout the text; the axis labels are also visually corrupted and should be regenerated.","section":"Figure 3"},{"comment":"The accuracy comparison figures in the appendix render as hexadecimal/unicode strings rather than readable values, so the per-dataset results cannot be verified from the appendix.","section":"Figures 9-13"},{"comment":"There are several small typos and inconsistencies: 'SV AMP' should be 'SVAMP', 'GPT4 turbo)' has a stray parenthesis, and Appendix B.4 refers to 'Appendix 2' instead of a proper section.","section":"Various"},{"comment":"The proof of Lemma 2.2 states that the difference in state distributions is 'negligible (justified under Assumption B.8)', but Assumption B.8 is a concentrability condition, not an equality of state distributions; this step should be made explicit or removed.","section":"Appendix C.1"}],"recommendation":"reject","confidential_remarks":"The paper has useful empirical components, but the theoretical claims are central to the framing and are not established; the BNE convergence proof is circular and the regret analysis contains an invalid learning-rate schedule. The empirical headline is also not accurately supported by the reported comparisons. In my view these are load-bearing errors that cannot be fixed within the manuscript's current scope; a resubmission would need to either substantially revise or remove the theoretical claims and report the empirical results in a more measured way."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nQuick take: the empirical framework — small belief networks that set temperature and repetition penalty, plus a QMIX-style mixing network — is genuinely new in the multi-LLM debate literature, and the results are suggestive. But the headline theoretical claims are not supported. The BNE convergence proof in Appendix B.2 is circular: Step 3 assumes the Q-network satisfies Bellman optimality at the stationary point, which is exactly what it needs to conclude agents are best-responding. And Appendix C.2 sets eta_t = eta_0/sqrt(t) while invoking Robbins-Monro conditions that require sum eta_t^2 < infinity; that sum diverges for this schedule. So the O(N sqrt(T)/(1-gamma)) regret bound rests on a theorem that does not apply.\n\nWhat is genuinely good: the belief-network mechanism is a real departure from message-passing debate, the coordinator-executor split is sensible, and the paper is honest enough to note that communication is reduced but not eliminated. The code is public and the hyperparameters are detailed, so the empirical claims are re-implementable. That is worth something.\n\nThe soft spots are in the theory and the statistics. BNE existence is asserted via Glicksberg, but the continuity and quasi-concavity of the payoff over the LLM output distribution are not established, and the reward design does not obviously provide them. The regret analysis leans on unmeasured concentrability and posterior-alignment assumptions (B.4, B.6, B.8) plus generic O(1/sqrt(t)) rates that have not been shown to hold for belief networks controlling a discrete text generator. Empirically, no error bars, seeds, or significance tests are reported, so the 11.2% average gain could be noise.\n\nStill, I would not desk-reject. The method is concrete and the empirical phenomenon is worth a careful look; a referee could push for a revised theory and a more honest treatment of the empirical claims. For my own work, I would not cite the theoretical results as given.\n\nBest,\n\n[Name]","headline":"A plausibly useful empirical recipe for communication-light multi-LLM coordination, but the claimed BNE convergence and regret bounds do not survive contact with the appendix.","tokens_in":43514,"tokens_out":2677,"would_cite":false,"duration_ms":31548,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that replacing debate-style message passing among LLMs with belief-driven coordination toward a Bayesian Nash equilibrium produces more accurate answers at lower token cost, supported by a sublinear regret bound and a…","keywords":["multi-agent LLM reasoning","Bayesian Nash equilibrium","belief-driven coordination","regret bound","decentralized POMDP","multi-agent debate","reinforcement learning","LLM ensembles"],"falsifier":"A direct test would freeze the belief networks at random initialization — constant temperature and repetition penalty — while keeping ECON's coordinator prompts and rewards identical; if accuracy on MATH or GSM8K is unchanged, the learned equilibrium mechanism is not the source of the reported gains. A second test would log the TD error and the gap between learned and optimal Q-values during training on one benchmark and check whether either actually decays as $O(1/\\sqrt{t})$, since the regret proof's $\\epsilon_t$ and $\\delta_t$ rates are asserted, not measured. A third check: the appendix's convergence proof assumes the Q-network is optimal at the stationary point, which is close to assuming the Bayesian Nash equilibrium it aims to establish, so testing the equilibrium condition directly — whether any single agent can improve its reward by changing its belief-derived prompting strategy — would settle whether ECON actually reaches a BNE.","tokens_in":42341,"feed_emoji":"⚖️","tokens_out":10630,"duration_ms":115587,"temperature":0.7,"pith_summary":"ECON's project is to show that a team of language models can get better answers by reasoning about each other rather than talking to each other. The paper models a multi-LLM ensemble as an incomplete-information game and argues that if each execution model maintains a learned belief about its teammates' strategies and acts on it, while a coordinator merges those beliefs, the whole system converges to a Bayesian Nash equilibrium — a stable point where no model can improve its answer by changing strategy alone. It backs this with a regret proof, $O(N\\sqrt{T}/(1-\\gamma))$ for ECON versus $\\Omega(NT/(1-\\gamma))$ for debate-style schemes, and with experiments across six benchmarks where ECON beats existing multi-LLM methods by 11.2% on average and uses 21.4% fewer tokens than a three-round debate. If both claims hold, explicit multi-agent debate is not only wasteful but unnecessary: coordination-by-belief delivers the accuracy gain with a fraction of the communication.","feed_headline":"Beat debate by 11.2%: LLM teams coordinate via shared beliefs","feed_subtitle":"Each model reasons about its peers' strategies, reaching equilibrium without token-hungry debate.","key_machinery":"The load-bearing object is the Bayesian Nash equilibrium of an incomplete-information game, implemented as a coordinator–executor hierarchy: each execution LLM is steered by a belief network that outputs a two-dimensional prompt embedding (sigmoid-rescaled temperature and repetition penalty) and a local Q-value; a shared belief encoder aggregates the belief states with multi-head attention; and a centralized mixing network combines local Q-values and the group representation into a global Q-value whose monotonicity in each local Q-value is proven by induction over layers with non-negative weights and non-decreasing activations. The regret argument runs through a performance-difference lemma: the value gap between two policy profiles equals a discounted expectation of the Q-value gap, which splits the total regret into a Q-estimation error $\\epsilon_t$ and a policy suboptimality $\\delta_t$; the paper asserts both decay as $O(1/\\sqrt{t})$ and sums them via a harmonic bound to reach $O(N\\sqrt{T}/(1-\\gamma))$. Against this stand two lemmas for debate-style settings: no-regret learners in zero-sum games without pure-strategy equilibria must keep a minimum level of randomization, leaving a persistent suboptimality $\\delta_{\\min}$ that accumulates linearly.","core_discovery":"The central claim, stated on the paper's own terms, is that multi-LLM coordination can be recast as the pursuit of a Bayesian Nash equilibrium in an incomplete-information game, and that doing so yields both a convergence guarantee and better answers at lower cost than message-passing debate. Each execution LLM's belief network maps its local history to a prompt embedding — the temperature and repetition penalty of its sampler — plus a local Q-value; a belief encoder fuses all agents' belief states, and a centralized mixing network trains the local networks so that the global objective improves whenever any local policy improves. The paper proves BNE existence by Glicksberg's fixed point theorem, derives the sublinear Bayesian regret bound $O(N\\sqrt{T}/(1-\\gamma))$ through a performance-difference lemma with Q-estimation and policy-suboptimality error terms, and contrasts it with a linear $\\Omega(NT/(1-\\gamma))$ lower bound for debate settings where agents must randomize to avoid exploitation. Empirically it reports an average improvement of 10.9% over single-agent baselines and 11.2% over existing multi-LLM methods across six benchmarks, a 21.4% token reduction versus three-round debate, and an 18.1% further gain when scaling to nine execution LLMs under a hierarchy of coordinators.","pith_inferences":["Because the belief network learns only a temperature and a repetition penalty while the LLM stays frozen, the mechanism is essentially adaptive prompt-shaping; the same coordinator–belief architecture could plausibly steer any stochastic text generator toward consensus, not just LLMs.","The paper reports that adding direct inter-agent communication to ECON raised accuracy by only 1.1% while increasing token use by 42.4%, which suggests the practical value of belief coordination is mostly cost reduction; a natural comparison is a non-learned baseline that simply averages answers sampled under a fixed temperature schedule with the same coordinator prompts.","If the asserted $O(1/\\sqrt{t})$ error decay were verified empirically, the regret framework would transfer to other hierarchical coordinator–executor designs, implying that partial observability does not block equilibrium-based coordination as long as some central belief aggregator exists."],"forward_implications":["Multi-LLM ensembles can be scaled to at least nine agents with only moderate added cost: adding two local coordinators and a central coordinator yields an 18.1% gain over the three-execution-LLM baseline.","Token budgets fall because execution models never see one another's outputs; the paper reports 21.4% fewer tokens than a three-round debate at equal or higher accuracy.","Cooperative belief coordination, unlike competitive debate, is claimed to carry a sublinear $O(N\\sqrt{T}/(1-\\gamma))$ regret bound, meaning repeated application converges toward a stable strategy profile instead of cycling.","The framework is model-agnostic across open-weight models from 7B to 405B and GPT-4-Turbo, and heterogeneous execution teams still beat the few-shot CoT baseline, though they reach equilibrium less easily than homogeneous teams."],"supporting_citations":[{"why":"Supplies the multi-agent debate baseline that ECON compares against on both accuracy and token cost, and whose three-round variant is the main efficiency target.","marker":"(Du et al., 2024)"},{"why":"Introduced the multi-agent debate (MAD) message-passing design whose token overhead and lack of coordination protocol motivate the belief-based replacement.","marker":"(Liang et al., 2023)"},{"why":"Evidence that consensus thresholds in multi-agent debate significantly influence performance, motivating a principled coordination mechanism.","marker":"(Smit et al., 2024)"},{"why":"The learning-in-games reference used to argue that debate agents must maintain minimum-entropy randomization, producing the linear-regret lower bound.","marker":"(Fudenberg & Levine, 1998)"},{"why":"The stochastic approximation framework used to derive the $O(1/\\sqrt{t})$ Q-function estimation error in the regret proof.","marker":"(Borkar, 2009)"},{"why":"The online convex optimization machinery used to bound policy suboptimality at $O(1/\\sqrt{t})$.","marker":"(Hazan, 2016)"},{"why":"The provably efficient RL result whose regret decomposition the paper adapts for its own Bayesian regret bound.","marker":"(Jin et al., 2020)"}],"fun_headline_variants":["LLM teams find equilibrium: 11.2% better than debate","From debate to equilibrium: LLMs coordinate via Nash beliefs","Bayesian Nash equilibrium makes LLM teams cheaper and smarter","ECON: LLM reasoning via equilibrium cuts cost, boosts accuracy","No more debate: LLM agents reach Nash equilibrium, 11.2% gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the unmeasured premise that the small belief networks learn the way textbook Q-learners do — their Q-value estimates and policy choices improving as $O(1/\\sqrt{t})$ — even though these networks only set a temperature and a repetition penalty while a frozen LLM writes the actual answer, and that each model's guesses about its teammates' behavior stay close to the coordinator's own assessment.","fun_headline_variants_meta":{"raw":{"variants":["LLM teams find equilibrium: 11.2% better than debate","From debate to equilibrium: LLMs coordinate via Nash beliefs","Bayesian Nash equilibrium makes LLM teams cheaper and smarter","ECON: LLM reasoning via equilibrium cuts cost, boosts accuracy","No more debate: LLM agents reach Nash equilibrium, 11.2% gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000563,"raw_usage":{"total_tokens":2711,"prompt_tokens":1027,"completion_tokens":1684,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":1574}},"tokens_in":643,"tokens_out":1684,"duration_ms":12600,"temperature":1.0,"reasoning_tokens":1574,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:14:43.914932+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would freeze the belief networks at random initialization — constant temperature and repetition penalty — while keeping ECON's coordinator prompts and rewards identical; if accuracy on MATH or GSM8K is unchanged, the learned equilibrium mechanism is not the source of the reported gains. A second test would log the TD error and the gap between learned and optimal Q-values during training on one benchmark and check whether either actually decays as $O(1/\\sqrt{t})$, since the regret proof's $\\epsilon_t$ and $\\delta_t$ rates are asserted, not measured. A third check: the appendix's convergence proof assumes the Q-network is optimal at the stationary point, which is close to assuming the Bayesian Nash equilibrium it aims to establish, so testing the equilibrium condition directly — whether any single agent can improve its reward by changing its belief-derived prompting strategy — would settle whether ECON actually reaches a BNE.","supporting_citations":[],"review_version":1}