{"id":"5fa30bc5-845e-49a9-a172-2aadfaeed67e","arxiv_id":"2605.27073","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"BOT-Orch recasts orchestration as a regularized bandit problem using OT distances, achieving O(sqrt(T)) regret and provably preferring agents with better distributional alignment even when mean rewards match.","lead":"The paper proposes BOT-Orch, a framework that models adaptive agent orchestration as a bandit problem regularized by optimal transport distances between agent output distributions and reference distributions to explicitly handle uncertainty. A smart generalist might read it to understand how to make better sequential delegation decisions when using multiple unreliable AI agents with different costs and behaviors.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"O(sqrt(T)) regret and preference ordering assume OT distances are known/fixed rather than estimated online from non-i.i.d. samples","rationale":"The reader's weakest_assumption directly identifies the same point. With only the abstract the derivations cannot be inspected, so the verdict remains UNVERDICTED until the full proof is checked against the concrete test above.","tokens_in":1711,"tokens_out":317,"duration_ms":46315,"concrete_test":"Extract the precise definition of the regularized reward (presumably in the methods or theory section) and re-derive the regret bound while treating each OT distance as an empirical estimate formed from the first t samples; check whether the resulting bound remains O(sqrt(T)) or acquires an extra additive term linear in the OT estimation error.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central construction adds an OT-regularization term to the per-agent reward before feeding it to a standard bandit algorithm. The O(sqrt(T)) bound then follows immediately from the usual analysis provided the effective reward remains sub-Gaussian (or bounded). The preference-ordering claim likewise follows if the OT term is treated as a known constant that strictly lowers the effective mean for mis-aligned agents. Both results therefore rest on the OT distances being exactly known a priori and independent of the online samples; nothing in the abstract indicates how the distances are obtained or whether their estimation error is folded into the regret analysis. In the stated non-i.i.d., adversarial setting this is the least-secured step.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes BOT-Orch, a framework that recasts adaptive orchestration of heterogeneous agents as a bandit problem whose per-agent rewards are regularized by optimal transport (OT) distances between each agent's output distribution and a task-specific reference distribution. It claims that the resulting regularized policy enjoys O(√T) regret under standard assumptions and that the OT term provably induces a preference ordering among agents that share identical mean rewards but differ in distributional alignment. Empirical results on synthetic adversarial task-allocation instances with non-i.i.d. agent behavior show outperformance relative to unregularized bandit baselines and simple heuristics.","tokens_in":1852,"tokens_out":519,"duration_ms":27157,"significance":"If the regret analysis remains valid once OT distances are estimated online from streaming non-i.i.d. samples, the work would supply a principled mechanism for injecting distributional alignment into sequential delegation decisions. The preference-ordering guarantee could be useful for distinguishing agents beyond first-moment performance. The paper supplies an explicit algorithmic template and reproducible synthetic experiments; these are positive features.","major_comments":[{"comment":"The O(√T) regret claim is presented as following directly from standard bandit analysis once the OT term is added to the reward. The manuscript must clarify whether the OT distances are treated as known fixed constants or are estimated from the online non-i.i.d. samples; if the latter, the estimation error must be incorporated into the regret bound, otherwise the stated guarantee does not hold in the adversarial non-i.i.d. regime described in the abstract and experiments.","section":"Theoretical analysis (regret bound)"},{"comment":"The preference-ordering result likewise treats the OT term as a deterministic, strictly positive penalty for mis-aligned agents. The manuscript should state whether this ordering survives when OT distances are replaced by finite-sample estimates and whether the ordering holds with high probability or only in expectation.","section":"Preference-ordering theorem"}],"minor_comments":[{"comment":"The abstract invokes 'standard assumptions' for the regret bound without enumerating them; an explicit list (boundedness, sub-Gaussianity, etc.) would improve readability.","section":"Abstract"},{"comment":"The empirical section would benefit from explicit parameter values for the adversarial non-i.i.d. process and from reporting standard errors or confidence intervals on the plotted performance curves.","section":"Experiments"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on the theoretical components of BOT-Orch. We address each major comment below, clarifying the assumptions in the analysis while noting where revisions will improve the manuscript.","responses":[{"response":"The O(√T) regret bound is derived under the assumption that OT distances are known fixed constants, which permits direct application of standard bandit regret analysis (e.g., via UCB or EXP3) to the regularized per-agent rewards. This matches the phrasing 'under standard assumptions' in the abstract. Online estimation of OT distances from streaming non-i.i.d. samples occurs only in the empirical section and is not folded into the theoretical bound. We agree that explicitly stating this distinction and discussing the resulting gap for the fully online adversarial regime would strengthen the paper. We will revise the relevant sections to make the assumption clear and add a brief remark on estimation error.","revision_made":"partial","referee_comment":"[Theoretical analysis (regret bound)] The O(√T) regret claim is presented as following directly from standard bandit analysis once the OT term is added to the reward. The manuscript must clarify whether the OT distances are treated as known fixed constants or are estimated from the online non-i.i.d. samples; if the latter, the estimation error must be incorporated into the regret bound, otherwise the stated guarantee does not hold in the adversarial non-i.i.d. regime described in the abstract and experiments."},{"response":"The preference-ordering theorem is proven for exact OT distances, under which the regularized reward induces a strict preference for better-aligned agents even when mean rewards coincide. With finite-sample OT estimates the ordering holds in expectation (assuming an unbiased estimator) and would hold with high probability under standard concentration results for empirical OT distances, but the manuscript does not supply such a high-probability guarantee. We will revise the theorem statement and surrounding discussion to explicitly note the exact-distance assumption and briefly indicate how finite-sample versions could be obtained via concentration inequalities.","revision_made":"partial","referee_comment":"[Preference-ordering theorem] The preference-ordering result likewise treats the OT term as a deterministic, strictly positive penalty for mis-aligned agents. The manuscript should state whether this ordering survives when OT distances are replaced by finite-sample estimates and whether the ordering holds with high probability or only in expectation."}],"tokens_in":1371,"tokens_out":505,"duration_ms":25879,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's core move is to turn agent selection into a bandit where each agent's effective reward is its observed performance minus an OT distance to a task reference distribution. This produces the usual sqrt(T) regret under standard sub-Gaussian assumptions and, more interestingly, a strict preference ordering among agents that share the same mean but differ in how their output distributions align with the reference.\n\nThe formulation is straightforward and the preference-ordering claim follows directly once the OT term is added as a constant shift. The synthetic adversarial experiments show it beating plain UCB and simple heuristics, which is consistent with the theory.\n\nThe soft spot is the one the stress-test note flags. The regret bound and the ordering both treat the OT distances as known a priori and independent of the online samples. The abstract describes a non-i.i.d., adversarial setting, yet gives no account of how those distances are obtained or whether their estimation error enters the regret. If the distances must be estimated from finite or drifting samples, the analysis does not cover the additional error, so the stated guarantees may not hold in the regime the paper claims to address.\n\nCitation coverage of related distributional or robust bandit work is thin, but that is secondary. The empirical section is limited to synthetic data, which is acceptable for an initial theoretical paper but leaves open how the method behaves with real agent outputs.\n\nThis is aimed at researchers working on multi-agent routing or LLM orchestration where distributional match matters more than raw mean reward. A reader who wants a clean way to encode distributional preferences inside a bandit will find the preference result useful. The work is coherent on its own terms and the central construction is distinct enough that it deserves a serious referee to check the missing estimation step and any additional assumptions in the full proofs.","headline":"BOT-Orch adds OT regularization to a bandit for agent orchestration and gets the standard O(sqrt(T)) regret plus a distributional preference result, but both rest on treating the OT distances as known fixed constants.","tokens_in":2338,"tokens_out":440,"would_cite":false,"duration_ms":27085,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A bandit formulation regularized by optimal transport distances between agent outputs and task references achieves O(sqrt(T)) regret and orders agents by distributional alignment.","keywords":["agent orchestration","bandit algorithms","optimal transport","regret analysis","uncertainty in AI agents","multi-agent systems"],"falsifier":"Observing that the learned orchestration fails to prefer the better-aligned agent when two agents have the same mean reward but different OT distances to the reference, or that the regret grows faster than sqrt(T) under the stated assumptions.","tokens_in":2615,"feed_emoji":"🤖","tokens_out":607,"duration_ms":25376,"temperature":0.7,"pith_summary":"The paper develops a method for deciding which agent to delegate tasks to when their behaviors are uncertain and changing. It models the choice as a bandit problem where the effective reward includes a penalty based on how well each agent's output distribution matches a task-specific reference, measured by optimal transport distance. This regularization ensures that even agents with the same average performance can be ranked by how closely their output patterns fit the task. A sympathetic reader would care because coordinating multiple AI systems or models often involves hidden differences in their reliability and output styles that simple average-reward methods miss. The approach provides both a theoretical performance guarantee and a way to prefer better-aligned agents.","feed_headline":"Regularized bandit orders agents by output alignment","feed_subtitle":"Penalizing distributional mismatch with optimal transport yields O(sqrt(T)) regret and distinguishes identical-mean agents.","key_machinery":"The OT-regularized reward in the bandit formulation, which augments the observed reward with a term that penalizes distributional mismatch.","core_discovery":"BOT-Orch recasts the orchestration problem as a regularized multi-armed bandit where each arm corresponds to an agent and the regularization term uses optimal transport distances to reference distributions. Under standard assumptions the algorithm attains O(sqrt(T)) regret. It also provably induces a strict preference ordering over agents that share the same mean reward but differ in their output distributions' alignment with the task reference.","pith_inferences":["This approach could be tested in settings where task references evolve over time to see if the regret bound still holds.","Similar regularization might help in domains like recommendation systems where item distributions matter beyond averages."],"forward_implications":["The orchestration policy achieves sublinear cumulative regret over time.","Agents with identical mean rewards receive different selection probabilities based on their distributional alignment.","The framework outperforms standard bandit algorithms and heuristics in non-i.i.d. adversarial settings.","Delegation decisions explicitly account for uncertainty in agent output distributions."],"fun_headline_variants":["OT-regularized bandits order agents by distributional match","Orchestration attains sqrtT regret with OT regularization","Regularization induces strict preference among equal-reward agents","Agent delegation prefers distributionally aligned outputs"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The assumption that optimal transport distances to task references are well-defined and computable, together with the standard sub-Gaussian or bounded reward conditions needed for the regret analysis.","fun_headline_variants_meta":{"raw":{"variants":["OT-regularized bandits order agents by distributional match","Orchestration attains sqrtT regret with OT regularization","Regularization induces strict preference among equal-reward agents","Agent delegation prefers distributionally aligned outputs"]},"model":"grok-4.3","cost_usd":0.00609,"raw_usage":{"total_tokens":2854,"prompt_tokens":621,"num_sources_used":0,"completion_tokens":57,"cost_in_usd_ticks":60899500,"prompt_tokens_details":{"text_tokens":621,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2176,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":621,"tokens_out":57,"duration_ms":27699,"temperature":1.0,"reasoning_tokens":2176,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T18:54:57.867433+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Observing that the learned orchestration fails to prefer the better-aligned agent when two agents have the same mean reward but different OT distances to the reference, or that the regret grows faster than sqrt(T) under the stated assumptions.","supporting_citations":[],"review_version":1}