acceptodds
Under review as a conference paper at ICLR 2027

Upon the Certified Regulation of Deliberation in Multi-Agent Systems Possessed of Unshared Evidence

Abstract

Where the evidence pertinent to a decision lies dispersed among several agents, deliberation constitutes the sole instrument by which it may be consolidated; yet deliberation is customarily left unregulated. Fixed schedules, round-robin exchange and unconstrained debate afford no assurance whatsoever that a given conversational act is the one warranted by the circumstances, and assemblies of language models, in the manner of human committees, are disposed to rehearse what is already held in common and to acquiesce prematurely in a plausible error. We advance We advance QUORUM (Quantified Uncertainty and Oversight of Regret for Unshared-evidence Multi-agent deliberation), an inference-time framework which regulates deliberation rather than merely countenancing it. At each round the framework reduces the proceedings to a compact state comprising uncertainty, dissent, redundancy, evidential gain and premature concurrence; thereupon it determines both the intervention warranted and the agent most fittingly charged with its execution. Each proposal must thereafter withstand a round-wise conformal examination, which circumscribes its one-step regret under a distribution-free, finite-sample assurance and substitutes a certified alternative wherever the proposal is adjudged inadmissible. Regulated deliberation surpasses its unregulated counterpart throughout twelve models and two benchmarks, and not infrequently surpasses the condition wherein every agent is furnished with the entirety of the evidence.

Primary area:
foundation or frontier models, including LLMs
Keywords:
certified deliberation, conformal calibration, one-step regret, hidden-profile reasoning, regulation of communication, allocation of speakership, collective adjudication, multi-agent language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8bYKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose a closed-loop discussion control framework QUORUM to address the issues of insufficient information sharing and premature consensus in hidden-profile multi-agent reasoning. The system first uses six metrics to represent the current discussion status, and then uses a controller to dynamically select discussion actions. The authors computed the One-Step Regret of candidate actions using Exhaustive Counterfactual Rollouts, filtered and replaced controller proposals using Round-Wise Conformal Calibration, and selected the speaking agent using Action-Conditioned Routing. The authors conducted experiments on HiddenBench and a synthetic travel decision dataset, and the results showed that QUORUM significantly outperforms fixed Round-Robin Discussion and surpasses Full-Information Baseline in some settings.

Critical strengths
  1. This paper proposes a systematic closed-loop framework to solve the problem of information sharing failure in multi-agent reasoning tasks. By combining discussion status monitoring, adaptive intervention selection, and a speaker guidance mechanism based on action conditions, this paper provides a structured approach that can effectively encourage parties to show private evidence and mitigate the tendency to reach consensus prematurely.
  2. This paper evaluates several different controllers and provides several ablation studies. The results show that these designs are a big improvement over traditional polling-based discussions. And the ablation study also provides useful insights into the contributions of evidence-seeking behavior, speaker scheduling, and controller state characteristics.
Critical weaknesses
  1. The conformal certification mechanism proposed in this paper lacks a sufficient theoretical and empirical support. First, the mechanism concerns only a single-round surrogate regret rather than the correctness of the final decision or the optimality of the global policy. Furthermore, the validity of the mechanism relies on an assumption of round-specific exchangeability. However, if early conformal replacements change the state distribution in subsequent rounds, this exchangeability assumption may be no longer valid. More importantly, the actual contribution of conformal acceptance remains unclear. Since the framework exhaustively evaluates all candidate actions during the inference phase, the action with the minimum regret could have been easily seen. The paper lacks a counterfactual greedy baseline, a comparative experiment based on matched-split (i.e., a comparison with and without conformal acceptance), and empirical analysis on coverage and action rejection rate. Therefore, the necessity and effectiveness of this core authentication mechanism have not been convincingly demonstrated.
  2. There is limited empirical evidence regarding fairness and generalization capabilities. QUORUM requires extensive counterfactual reasoning for candidate actions, which introduces a significant amount of additional inference computation. However, this expense was not fully reflected in its communication cost metric, nor was it effectively controlled in baseline comparisons. Therefore, a stronger baseline model with comparable computational cost is needed to distinguish the benefits of adaptive coordination from those of additional inference and structured hints. Furthermore, the evaluation of this study focuses on the structured hidden-profile task, and the standard MLP controller it uses is trained on training data generated from all benchmark tasks. Although the paper includes a held-out evaluation, the fact that the test set used for this evaluation is different from the test set used by the full-task controller weakens the persuasiveness of its conclusions regarding generalization ability.
Questions for authors
  1. Since QUORUM exhaustively evaluates all candidate actions during inference, so what is the practical advantage of conformal acceptance over directly selecting the minimum-regret action?
  2. How is the round-wise exchangeability assumption maintained when conformal action replacements alter subsequent discussion trajectories?
  3. What is the actual computational overhead of exhaustive counterfactual rollouts, including additional LLM calls and token consumption?
3 comments

We sincerely thank the reviewer for the constructive feedback and recognition of QUORUM's closed-loop framework and ablation studies. We have strengthened the manuscript with additional experiments and clarified the theoretical assumptions, computational costs, and generalization evidence.

1. Conformal Acceptance and Greedy Selection

We agree that exhaustive evaluation identifies the minimum-regret action. However, QUORUM preserves the controller's proposed action when its regret is within the calibrated threshold, rather than always minimizing immediate surrogate loss.

We added a controlled Qwen3-8B comparison (Appendix D, Table 10):

PolicyTask Accuracy
QUORUM (LLM+Vote)91.7%
Fixed evidence seeking79.2%
Fixed action schedule58.3%
Greedy minimum-JJ70.8%
Random conformal selection70.8%

These results support the value of controller-guided selection over greedy and fixed policies, although they do not independently isolate conformal screening.

We additionally report before/after conformal accuracy on 45 held-out tasks: Rules: 91.1%→93.3%; MLP: 75.6%→84.4%; LLM: 86.7%→86.7%; LLM+Vote: 82.2%→84.4%.

New diagnostics (Table 13) report 97.2% proposal acceptance and 2.8% rejection across 248 decision states. We clarify that conformal coverage concerns one-step surrogate regret, not final-answer correctness or global policy optimality.

2. Round-Wise Exchangeability

We appreciate this important observation. We clarify that the calibration procedure incorporates conformal action selection during calibration trajectories, so earlier interventions influence subsequent states as they do during evaluation (Section 3.2).

Nevertheless, identical procedures do not automatically establish exchangeability, particularly with data-dependent thresholds. Proposition 1 explicitly assumes round-wise exchangeability of calibration and evaluation scores, and we make no unconditional sequential or trajectory-level guarantee. Appendix A clarifies these conditions and distinguishes statistical proposal coverage from the evaluated regret of executed actions.

3. Evaluation Fairness and Generalization

Fairness: Hidden-Post and QUORUM share a maximum of 10 communication rounds (Section 4.1). The controlled-policy experiments retain the Qwen3-8B backbone and communication templates, while action and routing ablations (Tables 7–10) further examine individual components. We acknowledge that differences in realized discussion length and prompting limit complete causal isolation.

Generalization: We added a 13-split held-out evaluation (Section 5.3, Appendix B.3). Each MLP is trained on 5 of 65 HiddenBench tasks and evaluated on the other 60. The held-out controller achieves 0.73 accuracy versus 0.77 for the all-task reference, supporting transfer to unseen tasks while acknowledging the minimal gap. Heterogeneous-agent and generated-benchmark evaluations provide additional evidence across configurations.

4. Computational Overhead

We agree that exhaustive counterfactual evaluation is computationally expensive. New Table 11 reports LLM calls, tokens, counterfactual evaluations, and execution time.

Full QUORUM averages 142.8 LLM calls (101.3 counterfactual), 123.0k input tokens, 13.8k output tokens, and 81.4 seconds per task. Fixed evidence seeking requires 65.5 calls and 35.3 seconds, while greedy minimum-JJ requires 160.1 calls and 90.8 seconds.

These measurements expose the accuracy–computation trade-off. Although greedy selection incurs greater measured overhead with lower accuracy in this experiment, we acknowledge that a strictly compute-matched comparison with stronger baselines remains open.

We thank the reviewer again for these valuable suggestions. The revisions provide stronger controlled comparisons, held-out generalization evidence, explicit computational measurements, and a more precise statement of conformal guarantees. We believe they strengthen the central finding that adaptive, state-aware communication control improves evidence integration in hidden-profile multi-agent reasoning, while clearly delineating the method's assumptions and limitations.

Thank you for your detailed response. Your reponse has largely solve my problems, but two remain. Firstly, current experiment results do not strongly support the independent benefits and robustness of conformal acceptance. More importantly, acknowledging the assumptions of round-wise exchangeability and independence between calibration data and the score-generating procedure does not establish that they hold under the deployed sequential policy, particularly with data-dependent calibration thresholds. I will keep my original score, although my overall assessment is closer to 2.5. I would have no objection if the AC and PC decide to accept the paper.

We sincerely thank the reviewer for their thoughtful feedback, encouraging reassessment, and openness toward acceptance.

Regarding independent benefits, we clarify that the previously reported results come from independent end-to-end runs with conformal screening enabled versus disabled on the same tasks, with all other components fixed. These experiments show improvements for Rules (91.1% → 93.3%), MLP (75.6% → 84.4%), and LLM+Vote (82.2% → 84.4%), providing evidence of independent benefits of conformal acceptance. We acknowledge that broader robustness need further evaluation.

Regarding exchangeability, we appreciate the reviewer's important theoretical observation. Our calibration procedure incorporates conformal acceptance during calibration itself, allowing trajectories to account for preceding interventions and more closely reflect the deployed sequential policy, thereby reducing calibration–deployment mismatch. While establishing exchangeability rigorously in this adaptive setting is challenging, our theoretical guarantee is explicitly conditional on the stated assumptions.

We greatly appreciate the reviewer's constructive engagement and valuable consideration.

Reviewer_nC7qOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce QUORUM, a sequential communication decision framework for LLMs with rule-based, learned, and LLM-based controllers and conformal screening of actions using one-step regret. Though not formalized as such, this amounts to a POMDP, with a centralized coordinator selecting communication actions using one-step regret measured with a heuristic for evaluation, and with a hand-crafted communication-specific action space. In stage four, the selected action is routed to an agent, typically in a rules-based way. A shared blackboard is then updated, and the process repeats at the next decision epoch. QUORUM is evaluated across several language models on HiddenBench and an LLM-generated benchmark, where it is shown to consistently improve over uncontrolled hidden-information protocols. QUORUM also matches or exceeds a full-information baseline in experiments across models, suggesting that the process of reaching a quorum is just as effective as having all the critical information itself.

Critical strengths
  1. The hidden-profile framing using communication control makes sense and the problem is well introduced. The integration of a conformal gate to determine controller deviation from a greedy policy is interesting.
  2. Evaluation breadth. Fifteen models, two benchmarks, and multiple types of ablations presented in the main manuscript and elaborated on in the appendix. The heterogeneous-agent ablation and action-leave-one-out ablation are both telling. Also, it is significant and telling that Full Info is substantially outperformed by the introduced controllers across mode types.
Critical weaknesses
  1. The contribution bullets claim "communication efficiency", but the use of counterfactuals would seem to result in a substantial number of LLM calls for planning each round. If this is truly a sequential decision process, as the introduction alludes to, was more than a single-step lookahead considered?
  2. I find the manuscript generally challenging to read; the abstract tone is markedly different than the rest of the manuscript, and the figures are challenging to read. The introduction of stages one through five feels disjointed. Results are often presented and not explained; for example, why exactly in 5.2 are some controller designs (mostly rules) outperforming the other for certain model types in Table 1? Table 1 is without conformal acceptance - does Table 2 consider conformal acceptance?
Questions for authors
  1. The authors state: "At each discussion round, we uniformly sample an action from the conformal prediction set rather than executing the controller’s preferred action, and follow the resulting trajectory until termination. Randomly sampled actions maintain high task success across all controller variants, with only small differences between them". For a randomly sampled action from the conformed prediction set to maintain high task success across controllers seems to indicate to me that the controller itself is doing very little work - please explain.
  2. Why does Table 1 list 15 models but Table 2 is averaged over 12?
  3. Communication cost would seem to me to be a sensitive hyperparameter that when overweighted restricts information exchange and when underweighted results in long discussions that reduce accuracy. To what extent was this sensitivity considered, and is it model-specific?
1 comment

We sincerely thank you for the constructive feedback and recognition of QUORUM's formulation and experimental breadth. We have added controlled comparisons, computational measurements, and methodological clarifications to address the concerns.

1. Computational Cost and Sequential Planning

We agree that counterfactual evaluation introduces substantial overhead. New Table 11 reports 142.8 LLM calls (101.3 counterfactual), 123k input tokens, and 81.4 seconds per task for QUORUM. We clarify that communication efficiency concerns transcript length, not total inference cost. QUORUM uses one-step lookahead with closed-loop replanning; multi-step optimization remains future work due to its substantially larger search space.

2. Random Conformal Sampling and Controller Contribution

Although random conformal actions sometimes perform well, this does not imply that controller selection is unnecessary. New controlled comparisons (Table 10, Qwen3-8B) show:

PolicyAccuracy
QUORUM (LLM+Vote)91.7%
Fixed evidence seeking79.2%
Fixed action schedule58.3%
Greedy minimum-JJ70.8%
Random conformal70.8%

These results support controller-guided selection, with QUORUM outperforming random conformal selection by 20.9 percentage points. Table 9 additionally shows that random conformal sampling varies substantially across controllers (63.0%–82.75%), partly due to termination behavior. We acknowledge that discussion-length differences limit causal isolation.

3. Model Counts and Conformal Results

Table 1 evaluates 15 models on HiddenBench; Table 2 averages 12 models on the generated benchmark, excluding Phi-4 and two DeepSeek models. We clarified this distinction in Appendix B.

Table 1 reports pre-conformal results; Table 3 reports post-conformal HiddenBench results on 45 held-out tasks, following calibration on 20 tasks. Table 2 concerns the separate generated benchmark.

4. Communication-Cost Sensitivity

The surrogate coefficients are fixed before evaluation and shared across controllers, without model-specific tuning. Table 6 examines communication-cost features and their impact on accuracy and discussion length. The complete state achieves the best overall objective. We acknowledge that a dedicated coefficient-sensitivity sweep remains untested.

5. Controller Differences

Rules performs best on average on HiddenBench (0.843), while LLM+Vote leads on the generated benchmark (0.668). This suggests controller effectiveness depends on discussion dynamics and termination requirements, rather than one policy universally dominating.

We thank you again. We believe the new experiments and clarifications strengthen the evidence for adaptive communication control, while transparently acknowledging computational costs and other limitations.

Reviewer_CkmwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the problem of multi-agent LLM deliberation in hidden-profile tasks, where the information shared by all agents favours a plausible but incorrect answer, while the correct answer can only be identified through discussion by revealing and combining their private evidence. The authors propose QUORUM, an inference-time framework that adaptively regulates communication rather than relying on fixed discussion schedules or unconstrained debate. At each round of communication, it constructs a six-dimensional state capturing group uncertainty, inter-agent disagreement, message redundancy, evidence gain, communication cost, and premature consensus. Then a controller, which can be rule-based, an MLP, or an LLM, proposes one of four communication interventions: CHALLENGE, CLARIFY, SEEKEVIDENCE, or ROUTE. The proposal is evaluated using counterfactual next-state outcomes and screened through a round-wise conformal procedure based on one-step regret under a surrogate discussion loss. If its regret exceeds the calibrated threshold, it is replaced by a minimum-regret admissible action. A deterministic routing policy then selects the speaker, prioritising the most dissenting agent for challenge and clarification, and the least-heard agent for evidence-seeking interventions.

The authors evaluate QUORUM on HiddenBench, comprising 65 hidden-profile tasks, across 15 language models, and on a separately generated travel-planning benchmark. On HiddenBench, the rule-based controller achieves an average task accuracy of 0.843 (before conformal screening), compared with 0.285 for fixed round-robin discussion and 0.731 for the full-information, no-discussion baseline. The experiments also examine the different controller variants, action-conditional speaker routing, the contribution of individual actions and state features, heterogeneous agent groups, and generalisation to held-out tasks. Overall, the authors conclude that adaptive communication control can substantially improve collective reasoning when relevant evidence is distributed across agents.

Critical strengths
  1. The paper shows that structured discussion substantially improves evidence pooling without retraining agents, raising average task accuracy from 0.285 with fixed round-robin discussion to 0.843 with QUORUM's rule-based controller (before conformal screening). This gain holds consistently across 15 models of different families and sizes.

  2. The authors compare rule-based, MLP-based, and LLM-based controllers while keeping the prompts and speaker-routing mechanism fixed. This makes it easier to compare the controller types with one another. The strong performance of the simple rule-based controller is particularly interesting, suggesting that complex learned policies are not necessarily needed to improve discussion. The action ablations also highlight the importance of the ROUTE and SEEKEVIDENCE interventions, which ask less-heard agents for new evidence.

Critical weaknesses
  1. The practical benefit of conformal certification is unclear. The main results are reported before conformal screening, and the held-out results show comparable or slightly lower accuracy on average. Since the authors already evaluate every admissible action, they should compare QUORUM against simply selecting the minimum-regret action at each round. The paper also does not report how often the regret bound is satisfied on held-out data or how often proposed actions are rejected. Finally, the guarantee concerns one-step regret under the surrogate loss JJ, not final answer correctness. These omissions make it difficult to assess the practical value of the certification mechanism.

  2. The source of the gains and the computational cost need clarification. QUORUM's action-specific instructions explicitly ask the speaker to share a new decision-relevant fact and to state which option it rules out, whereas the round-robin baseline appears to use the original HiddenBench protocol. A round-robin baseline with matched prompts and discussion length would help establish whether the gains come from adaptive control or from the prompts themselves. Moreover, the conformal step evaluates every admissible action at each round, which requires additional LLM calls, but the paper does not report the resulting inference cost or token usage. The evaluation also appears to rely on a single run, making small differences between controllers difficult to interpret.

Questions for authors

Questions 1–3 are clarification requests and question 4 is a suggestion for future work.

  1. Conformal certification: How does choosing the greedy minimum-loss (JJ) branch compare with the proposed conformal controller in accuracy and computational cost? Could you report the calibration level, empirical coverage, and proposal rejection rate, and clarify whether the surrogate loss JJ correlates with final answer correctness? Please also report the weights of JJ and the rule-controller thresholds.

  2. Baselines and computational cost: What prompts and number of rounds does the Hidden Post baseline use? Could you compare it with round-robin using QUORUM's prompts, a fixed action, and the same number of rounds? Please also report LLM calls or token usage per task.

  3. Inconsistencies and benchmark design: Section 5.4 suggests small differences between controllers when sampling from the conformal set, but Table 9 reports substantial gaps. Could you clarify this discrepancy? Also, why is full-information accuracy on the generated benchmark so low, and are agents informed of the weighted scoring rule defining the gold answer?

  4. Connection to Dynamic Epistemic Logic: Have the authors considered framing deliberation as resolving distributed knowledge through public announcements? This could provide principled definitions of evidence gain and redundancy, an epistemic reference for evaluating discussion efficiency, and a more meaningful stopping criterion for VOTE. It may also offer a complementary route towards guarantees about answer correctness, beyond the current surrogate-regret guarantee.

1 comment

We sincerely thank the reviewer for the thoughtful assessment, constructive suggestions, and recognition of QUORUM's contributions. We have strengthened the revised manuscript with additional experiments, computational analysis, and methodological clarifications addressing the concerns.

1. Conformal Certification and Practical Benefits

We agree that exhaustive evaluation identifies the greedy minimum-regret action. However, QUORUM retains the controller's proposal when its regret is admissible, rather than always minimizing immediate surrogate loss.

We added a controlled Qwen3-8B comparison (Table 10):

PolicyTask Accuracy
QUORUM (LLM+Vote)91.7%
Fixed evidence seeking79.2%
Fixed schedule58.3%
Greedy minimum-JJ70.8%
Random conformal70.8%

These results support controller-guided selection, although they do not independently isolate the contribution of conformal screening.

We also added Table 13, reporting 97.2% empirical proposal acceptance and 2.8% rejection across 248 decision states, with nominal coverage of 90%. The guarantee is explicitly limited to one-step surrogate regret under round-wise exchangeability, not final-answer correctness.

Additionally, Table 12 reports correlations between surrogate loss and correctness: Spearman ρ=−0.193\rho=-0.193 for final correctness and ρ=−0.306\rho=-0.306 for changes in correct-answer probability. These indicate limited but useful local alignment, rather than a correctness guarantee.

2. Baseline Fairness and Computational Cost

We appreciate the concern regarding stronger prompts. Hidden-Post uses the original HiddenBench round-robin protocol, while QUORUM uses action-conditioned instructions. Both share a maximum of 10 rounds, although adaptive termination can shorten discussions.

New controlled-policy experiments (Table 10) retain QUORUM's backbone and communication templates while changing action-selection strategies. The reductions under fixed policies provide evidence that improvements are not attributable solely to prompting. We acknowledge that realized discussion lengths differ.

We also added Table 11, reporting inference calls, tokens, and runtime. QUORUM averages 142.8 LLM calls, 123k input tokens, and 81.4 seconds per task, making its additional computational cost explicit. These comparisons use one evaluation seed, so small differences should be interpreted cautiously.

3. Conformal Sampling and Benchmark Design

We clarified the interpretation of Table 9. Random conformal sampling achieves 82.75% success for LLM/MLP controllers but 64.5% and 63.0% for Rules and LLM+Vote, respectively. These differences are substantial and partly reflect earlier termination (approximately 3.24–5.43 versus 9.54–9.87 rounds). We therefore do not claim uniformly similar performance.

For the generated benchmark, we clarified that it is independently constructed and deterministically validated, rather than a direct reproduction of GroupTravelBench (Appendix E). Gold answers follow weighted preference aggregation: must = +2, prefer = +1, avoid = -1, reject = -2.

The relatively low Full-Info accuracy (0.366) suggests that providing all evidence does not necessarily ensure correct aggregation. We acknowledge that explicitly communicating the numerical scoring rule to agents requires clearer documentation and may affect this result.

4. Objective Parameters

The surrogate coefficients are fixed before evaluation and shared across controllers. We agree that the precise coefficient values and rule-controller thresholds should be explicitly tabulated for reproducibility.

We especially appreciate the suggestion to connect QUORUM with Dynamic Epistemic Logic. Modeling communication as public announcements could provide principled representations of knowledge acquisition, redundancy, and stopping conditions. We view this as a promising complementary direction for extending the current empirical framework toward stronger epistemic guarantees.

We thank you again for these valuable suggestions. The revised experiments and clarifications strengthen the evidence for adaptive communication control while making the conformal assumptions, computational trade-offs, and remaining limitations more transparent.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.