acceptodds
Under review as a conference paper at ICLR 2027

Learning to Elicit: Co-Evolving Routing and Information Disclosure in Societies of Language Agents

Abstract

Multi-agent language models increasingly operate in societies of agents characterized by information asymmetry, where the information required to solve a task is distributed across agents. Effective coordination therefore depends not only on solving the task but also on deciding what information to elicit and from whom. We study this problem in a multi-agent meeting environment where agents possess private facts and must communicate over a shared public record to reach a verifiable settlement. We introduce ElicitBench, a benchmark in which agents hold distributed private information and must reveal the relevant facts through interaction before a designated decision maker can reach the correct outcome. We then propose Co-Adapt, a co-evolutionary training framework that jointly learns routing and agent information disclosure through reinforcement learning and privileged on-policy distillation. The router learns which agents to engage, while agents learn which private information to reveal. On the held-out test set, Co-Adapt raises the fraction of settlement checks passed from 0.65 to 0.81 and the strict success rate from 14% to 25%. Gains are largest on domains unseen during training, where the fraction of settlement checks passed increases from 0.57 to 0.78. Ablation studies further demonstrate that the gains arise from jointly learning information routing and information disclosure, rather than from either component alone.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Co-evolving Agents, Information Asymmetry, On-policy Distillation, Reinforcement Learning, Society of Agents

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_N159OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

ElicitBench tests group decisions requiring private facts. Co-Adapt uses alternating training for information elicitation. The authors report higher strict success on held-out domains.

Critical strengths

The work is comprehensive and includes extensive experiments.

Critical weaknesses
  1. The main contribution of this paper should be the proposed benchmark. I believe the proposed method "co-adapt" lacks of innovation because there is no new algorithm but a mixture of existing technologies. So I believe the correct "primary area" should be "datasets and benchmarks", rather than currently chosen "other topics in machine learning (i.e., none of the above)".
  2. I find the significance of the proposed elicitation problem limited by its incentive structure. All participant policies optimize the same training reward (Appendix C.2). Disclosure is rewarded, while free actions taken with unrevealed private facts are penalized (Appendix C.1). No private disclosure cost or individual benefit from secrecy is modeled. For example, withholding may be individually optimal if disclosure carries a private cost exceeding its expected benefit to the speaker, or if revealing a colleague's contribution would reduce the speaker's own bonus. The participants therefore have no conflict of interest over disclosure: pooling their decision-relevant facts provides a straightforward cooperative solution to the information-access problem. The framing does not adequately reflect this simplification. Lines 48–49 pose learning to elicit unobserved information as an open question; Section 3.1 emphasizes reasons people might not volunteer information. However, these narrative motives are not represented as conflicting incentives in the formal objective. Limited attention, communication, and reasoning may still make execution difficult, but the benchmark does not establish a substantive strategic disclosure challenge. I therefore find the motivation for this particular setting insufficiently convincing.
Questions for authors

None

Reviewer_TKo3OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces ElicitBench, a benchmark of 1,100 multi-agent scenarios across 11 domains designed to evaluate information elicitation under distributed private information, and Co-Adapt, a training framework that alternates between optimizing a chair agent (via RL and privileged on-policy distillation) and participant agents (via RL). The authors validate the benchmark through static LLM verification, lone-vs-group rollout certification, and human audits, and demonstrate empirically that Co-Adapt substantially improves programmatic content and provenance pass rates across seen domains, unseen domains, and multiple model backbones.

Critical strengths
  1. Programmatic verification of content and provenance: Evaluating multi-agent outcomes and information grounding deterministically without relying on an LLM judge avoids evaluation bias and separates guessing the right decision from explicitly acquiring evidence.

  2. Comprehensive empirical execution: The authors evaluate across multiple model families (Qwen, Gemma, Phi, Llama), include out-of-distribution testing on unseen domains and HiddenBench, and provide thorough seed-robustness and ablation studies.

Critical weaknesses
  1. Structural collapse of the benchmark and reward into a trivial query-all and reveal-all protocol: While it is self-evident that a lone agent cannot pass checks requiring unobserved private facts, the critical question is whether the multi-agent setting actually requires nontrivial coordination or selective elicitation. By construction, ElicitBench guarantees that every non-decision-maker holds at least one decisive fact (no exceptions or bystanders, Prompt 1 and App. B.1) and forces the chair to give every participant at least one turn before settling (App. B.2). Furthermore, the training reward R_train (App. C.1, Eq. 3) imposes no penalty for over-disclosure, directly rewards revealing facts, and explicitly penalizes any conversational say action whenever an agent still holds any unrevealed private fact (via the N_free penalty). Consequently, the optimal policy mathematically reduces to a trivial protocol: the chair queries each of the 2 to 4 participants once in round-robin order, and each participant immediately executes the reveal action on their private fact without discussion.

  2. Empirical evidence confirms protocol compliance rather than learned coordination: Appendix D.1 and D.3 confirm that this degenerate collapse is exactly what Co-Adapt learns: trained models converge to about 4 turns (matching the minimum threshold T_min = N_decisive + 1) and reveal 100% of decisive facts in every single episode with zero standard deviation across seeds (3.00 on Seen, 3.02 on Unseen). As the authors themselves note in App. D.3, the remaining error after training comes entirely from the decision-maker's final JSON settlement step rather than information elicitation. Thus, the large gains over untrained baselines primarily reflect teaching agents the mechanics of the formal reveal action (which untrained baselines fail to trigger, resulting in near-zero provenance scores) rather than learning complex social inquiry.

Questions for authors
  1. Triviality of the optimal policy (Clarification): Given that every non-decision-maker holds a decisive fact, every participant must speak at least once, and N_free penalizes speaking (the say action) whenever an unrevealed private fact remains, why is sequential reasoning or privileged distillation needed at all? What prevents a hard-coded round-robin router paired with a rule or prompt forcing participants to immediately execute the reveal action on Turn 1 from achieving the exact same 100% disclosure rate in N_decisive + 1 turns?

Reviewer_wgAgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ELICITBENCH, a benchmark of 1,100 human seeded meeting scenarios across 11 domains in which decision-relevant information is distributed among three to five agents. A designated decision-maker must produce a correct settlement, and programmatic checks verify both the settlement and whether the required information was explicitly disclosed during the interaction.

The authors propose CO-ADAPT, which alternately trains a chair policy to select participants and a shared participant policy to disclose information and produce the final settlement. The chair is trained using reinforcement learning and privileged on-policy distillation, while participants are trained through reinforcement learning. The evaluation includes held-out scenarios from familiar domains, two entirely unseen domains, several model configurations, and an external HiddenBench evaluation. For the primary Qwen configuration, the fraction of checks passed on unseen domains increases from 0.57 to 0.78, while strict success increases from 11% to 21%.

Critical strengths
  • The study is focused and controlled. The explicit allocation of private information, defined interaction protocol, and separate correctness and provenance checks make it possible to study whether agents actually acquire the information needed for a decision. The use of practitioner-derived scenarios and human auditing also provides some grounding for the benchmark construction.

  • The benchmark is challenging under its evaluation criteria. Strict success remains limited even for strong baseline models, leaving substantial room for improvement. The unseen-domain results, additional model configurations, and three-seed analysis strengthen within benchmark generalization claims.

Critical weaknesses
  • The ecological validity and structural generalization of the learned behavior remain unclear. The training rewards are closely aligned with this structure, including rewards for revealing decisive facts and selecting participants who hold unrevealed decisive information. These choices create a strong inductive bias toward consulting all participants and disclosing their information before settling. It is therefore difficult to assess how much CO-ADAPT learns about selective information elicitation beyond mastering this protocol. In less constrained settings, some participants may have no relevant information, facts may be redundant or unreliable, and disclosure may carry costs or conflict with a participant’s objectives. The unseen-domain evaluation is valuable, but those domains retain the benchmark’s underlying interaction rules. The HiddenBench results provide additional evidence of transfer, although the evaluation contains only 65 scenarios. Another eval on a different benchmark with trained CO-ADAPT checkpoints would demonstrate ecological validity.

  • The component analysis is insufficiently connected to the generalization results. Table 4 reports ablations only on the seen split of a 550-scenario setting with 90 evaluation scenarios, while the main experiment uses 1,100 scenarios. The paper does not clearly explain this difference or how the smaller setting was constructed. Consequently, the ablations do not establish which components are responsible for the improvements on unseen domains. This is particularly relevant to the incremental contributions of privileged distillation and reward shaping, where some differences are relatively small. I appreciate the three-seed results in Appendix D.3, but equivalent uncertainty estimates are not provided for the component ablations or optimizer comparison. Reporting these ablations on the unseen split, with a clearly matched full-system reference and variability across runs, would make the conclusions about co-adaptation more convincing.

  • Figures 1 and 3 need a formatting pass. Figure 1 contains small, tilted panels and dense labels that make the example difficult to follow. Figure 3 compresses several training signals, model roles, and update paths into a small area. Larger labels, consistent alignment, and a clearer separation of the chair and participant update phases would improve readability.

Questions for authors
  1. Benchmark assumptions and simpler baselines: How does CO-ADAPT compare with fixed round-robin routing and an explicit instruction to disclose private facts, using equivalent models and communication budgets? This comparison would help establish the benefit of learned routing in an environment where every non-decision-maker holds decisive information.
  2. How does performance change when some participants have no relevant information or when private facts are redundant or irrelevant?
  3. Ablation setting and unseen-domain evaluation: Could the authors explain the 550-scenario setting used in Table 4 and its relationship to the 1,100-scenario main experiment? Were all ablations and the full CO-ADAPT reference trained and evaluated on the same subsets? Please also report the component ablations on unseen domains to clarify which components support transfer.
  4. Why was Dr. GRPO not used in the main when it was demonstrated better in the appendix? Can you show that Dr. GRPO is equivalent to the main results?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.