acceptodds
Under review as a conference paper at ICLR 2027

Who Should Know What? Empirical Information Design for Multi-Agent LLM Economies

Abstract

Multi-agent LLM systems are shaped not only by their component models but also by an information policy that determines what each agent observes and at what level of detail. Can giving AI agents more information lead to worse outcomes? To investigate this question, we introduce empirical information design, an experimental framework that compares disclosure policies on the same markets while holding models and interaction rules fixed. We study frontier LLMs in multi-resource trade, single-good bargaining, and double auctions, where agents interact while pursuing their individual gains. Using ground-truth utilities, we distinguish between the opportunities agents identify and the outcomes they achieve. In trade, full disclosure of exact valuations reduces total gains for GPT-5.1 and Gemini 3.7 Flash agents, but increases gains for Claude Sonnet 5 and Grok 4.3. Relative to coarse disclosure of preference rankings, GPT-5.1's total gains fall from 97.2% to 88.4% of the maximum available, while Sonnet's rise from 72.1% to 85.6%. In bargaining, exact information produces less balanced agreements for GPT-5.1, while Sonnet generally divides gains more equally. In auctions, full disclosure can change price demands enough to prevent beneficial trades, lowering GPT-5.1's total gains from 98.9% to 91.0% of the maximum. These results suggest that model selection and information design should be coupled, as models differ in how their responses to the same information support the system's objective.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Information design, Negotiation, Multi-agent LLMs, Economic games

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_aQYoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper aims to answer the question of whether providing LLM agents with more information about other agents’ preferences, values, and costs improves collective outcomes in economic interactions. To answer this question, the authors introduce an empirical information design framework that compares disclosure policies while holding markets, models, and interaction rules fixed. Experimental studies in multi-resource trade, single-good bargaining, and double auctions show that more detailed information can help agents identify better trades, but can also lead to more demanding offers, less balanced agreements, or lower total gains.

Critical strengths

This paper studies an interesting question: how the information provided to LLM agents affects their interactions and collective outcomes in economic settings. To answer this question, the authors conduct experiments across resource trading, bargaining, and double auctions. To isolate the effects of information disclosure, the authors hold markets, models, and interaction rules fixed within each comparison. Through these controlled comparisons, the paper shows how the effects of disclosure vary across models and interaction settings.

Critical weaknesses

My first concern is the assumption that the disclosed information about agents’ private values and costs is accurate. If I understand correctly, the authors make this assumption to isolate the effects of information disclosure. While I understand the purpose of this choice, in practical settings, such information may be noisy or only partially verifiable. When I read the results and try to understand their implications, I would naturally ask whether the same findings would hold even if the information were not entirely accurate. However, I failed to see the discussion and results to evaluate the inaccurate information. Without such an evaluation, it is hard for me to determine which conclusions can guide information disclosure beyond settings where truthful information is available.

My other concern is how to understand the uncertainty in the reported results. Based on my understanding, the confidence intervals capture variation across markets rather than variation across repeated model responses to the same input. But, I am not convinced they establish how consistently the models would respond to the same disclosure policy. I would appreciate repeated evaluations of a few main comparisons to help me understand the reliability of these findings.

Although the paper considers several models and interaction settings, it is hard for me to separate the effects of information disclosure from the effects of the chosen prompts and reasoning effort. If I understand correctly, the experiments use fixed prompts and low reasoning effort. Holding these configurations fixed helps make the comparisons controlled, but it does not show whether the findings would persist under other configurations. For example, would an agent respond to more detailed information in the same way if it received simpler strategic instructions or had more opportunity to reason? Without some sensitivity analyses, I am not sure how broadly to interpret the reported effects.

Questions for authors
  1. I would appreciate it if the authors could provide a small-scale evaluation showing what happens when the disclosed information is noisy or only partially accurate. This would help me understand how much the findings depend on the assumption of truthful disclosure. If such an evaluation is not possible, could the authors clarify which conclusions they expect to hold when this assumption is relaxed, and why?

  2. How consistent are the main disclosure effects when the models are evaluated repeatedly on the same markets? Based on my understanding, the current confidence intervals describe variation across markets, which does not directly answer this question.

  3. What generalizable insight should I take from the results beyond the observation? The paper already acknowledges that more information does not necessarily improve collective outcomes in strategic interactions. Could the authors clarify whether the findings support a rule for choosing disclosure policies, and whether that rule can be evaluated in markets not used to derive it?

7 comments
AuthorsOpenReview ↗

Response to Reviewer aQYo (Part-1)

We thank the reviewer for their positive assessment of our work and thoughtful suggestions. First of all, we have revised the manuscript and expanded the evaluation from five to ten models, including GPT-6 Sol, Claude Opus 5.5, and Grok 4.7. We also added experiments on noisy information, repeated model responses, alternative instructions and reasoning effort, and held-out policy selection. These additions strengthen the evidence and clarify how our findings can guide disclosure choices. Figures 1 and 3 and Appendix G now present the updated results.

Below we address each point and identify the corresponding revisions.


W1. My first concern is the assumption that the disclosed information about agents’ private values and costs is accurate....I would naturally ask whether the same findings would hold even if the information were not entirely accurate.

Thank you for raising this thoughtful point. Our study asks whether even accurate, relevant information can hurt collective outcomes of these agentic marketplaces. Therefore, a platform with reliable information must still decide what to reveal and to whom. We isolate this information-design decision by holding utilities and trading opportunities fixed while varying disclosure, so any losses cannot be attributed to incorrect information. We agree, however, that applying these findings also requires understanding what happens when information is imperfect.

We therefore added noisy-information experiments with GPT-6 Sol and Claude Sonnet 5 across all three games (Appendix G.5, Table 10), using 18 market instances per game and information condition. Here, we multiply each counter-party value by a random factor drawn uniformly from 0.8 to 1.2, so a true value of 100 becomes a perturbed value between 80 and 120. These are synthetic perturbations, not model-generated estimates. Full disclosure presents the resulting numbers; coarse disclosure presents rankings or profitability indicators derived from those same numbers. We compare accurate information with perturbed information presented either with a warning explaining the noise or without one. Perturbations are fixed per market and shared across models, policies, warning conditions, starting roles, and turns. Agents’ own utilities and the true utilities used to evaluate outcomes remain unchanged.

Each cell below shows coarse / full information performance, with the higher mean in bold. Trade and auction scores are percentages of maximum possible gains; bargaining uses normalized Nash scores, including zero for disagreements.

Table R1. Coarse and full disclosure under accurate and noisy information.

SettingModelNo noiseNoise with warningNoise without warning
TradeGPT-6 Sol98.6 / 90.997.7 / 93.795.1 / 89.7
TradeClaude Sonnet 565.5 / 87.567.5 / 76.876.2 / 89.3
BargainingGPT-6 Sol0.729 / 0.8240.695 / 0.7500.768 / 0.664
BargainingClaude Sonnet 50.668 / 0.9810.664 / 0.6840.610 / 0.753
AuctionsGPT-6 Sol96.8 / 90.196.9 / 74.996.6 / 88.0
AuctionsClaude Sonnet 591.5 / 91.691.5 / 91.991.3 / 91.9

These results show that:

  • First, more information can still reduce collective gains when it is noisy. Even with a warning, full disclosure lowers Sol’s trade and auction gains relative to coarse disclosure by 4.0 and 22.0 percentage points. Sonnet’s mean trade gains remain higher under full disclosure, but its bargaining advantage shrinks.

  • Second, full-information outcomes vary more across the tested conditions in some settings.. For example, Sol’s auction gains fall from 90.1% without noise to 74.9% with noise and a warning, while coarse gains remain near 97%. Coarse signals can be less sensitive to numerical errors because perturbations need not change a ranking or profitability category. The practical implication is to test disclosure policies under expected information errors, rather than assume that the policy preferred with accurate information remains best.

AuthorsOpenReview ↗

Response to Reviewer aQYo (Part-2)

W2. My other concern is how to understand the uncertainty in the reported results. Based on my understanding, the confidence intervals capture variation across markets rather than variation across repeated model responses to the same input... I would appreciate repeated evaluations of a few main comparisons to help me understand the reliability of these findings.

We thank the reviewer for identifying this distinction. In response to this point made by the reviewer, we added repeated runs to test whether the findings hold. Each of four models completed three fresh runs on the same 30 markets under both coarse and full disclosure, producing 720 new negotiations. Prompts, roles, message limits, and generation settings stay fixed. Overall, we see that the repeated averages agree in direction with the original single-run results.

The table below compares original 90-market averages with repeated 30-market averages. Entries show coarse → full gains; differences and intervals use the repeated runs.

Table R2. Original trade results compared with averages across three fresh runs.

ModelOriginal single run (90 markets)Three fresh runs (30 markets)Repeated full minus coarse (percentage points)Repeated 95% CI
GPT-5 mini93.9% → 83.7%91.1% → 84.0%-7.2[-12.3, -1.8]
GPT-6 Sol97.7% → 90.8%97.6% → 91.6%-6.0[-9.2, -2.9]
Gemini 3.7 Flash93.7% → 92.2%93.4% → 92.2%-1.2[-4.9, +2.8]
Claude Sonnet 572.1% → 85.6%70.5% → 83.2%+12.8[+5.4, +21.3]

To show variation between runs directly, we also report each run’s average disclosure effect. Positive values favor full disclosure; negative values favor coarse disclosure.

Table R3. Full-minus-coarse trade gains in each repeated run (percentage points).

ModelRun 1Run 2Run 3
GPT-5 mini−1.9−4.6−14.9
GPT-6 Sol−5.6−5.6−6.8
Gemini 3.7 Flash+2.0−1.9−3.6
Claude Sonnet 5+16.5+8.9+12.9

Here, we see that full disclosure lowers GPT-5 mini’s and GPT-6 Sol’s average gains and raises Sonnet’s in every run, matching the original findings. The effects averaged across runs are statistically significant, with all three 95% confidence intervals excluding zero. Only Gemini’s smaller effect changes direction across runs and is not statistically significant, where its coarse to full information difference was already low in the initial experiments.

Further details abotu these repeated-run results appear in Appendix G.2, Table 7, and Figure 23; the original results appear in Appendix G.1, Table 5.


W3. Although the paper considers several models and interaction settings, it is hard for me to separate the effects of information disclosure from the effects of the chosen prompts and reasoning effort.

We thank the reviewer for this suggestion. To adress this, we tested whether disclosure effects persist with simpler instructions or greater reasoning effort. For GPT-6 Sol and Claude Sonnet 5, the simplified variant removes the six numbered decision-process steps, retaining the objective to maximize own utility and the complete game rules, information cards, and response format.

The other variant retains the original prompt and raises reasoning effort from low to medium. Both use the same 30 trade markets. Each variant uses one run per disclosure condition, while the matched baseline averages three runs (Appendix G.2, Table 8).

Table R4. Disclosure effects under simpler instructions and greater reasoning effort.

ModelInstructions / reasoningCoarse (%)Full (%)Full minus coarse (pp), 95% CI
GPT-6 SolOriginal / low97.691.6−6.0 [−9.2, −2.9]
GPT-6 SolSimplified / low98.190.0−8.1 [−13.1, −3.5]
GPT-6 SolOriginal / medium97.492.9−4.4 [−8.9, −0.3]
Claude Sonnet 5Original / low70.583.2+12.8 [+5.4, +21.3]
Claude Sonnet 5Simplified / low58.865.5+6.7 [+0.2, +13.2]
Claude Sonnet 5Original / medium81.985.4+3.5 [−1.5, +10.0]

Both models retain the direction of their original disclosure effects under simpler instructions and medium reasoning: full disclosure lowers Sol’s mean gains and raises Sonnet’s. Sol’s losses remain statistically significant in both variants. Sonnet’s medium-reasoning effect also remains positive, and the matching directions support the original pattern.


AuthorsOpenReview ↗

Response to Reviewer aQYo (Part-3)

Q1. I would appreciate it if the authors could provide a small-scale evaluation showing what happens when the disclosed information is noisy or only partially accurate. This would help me understand how much the findings depend on the assumption of truthful disclosure.

The experiment in W1 shows why coarse disclosure can be less sensitive to numerical errors. Perturbations change the numbers shown under full disclosure but often preserve the rankings or profitability indicators shown under coarse disclosure. Larger errors can change these signals too, so this stability does not guarantee that coarse disclosure always produces more robust outcomes.

Furthrmore, information reliability does not remove the strategic question of what to disclose. Information-design theory accommodates imperfect signals [1], and Morris and Shin [2] show that more precise public information can reduce welfare under particular coordination incentives. Although their mechanism differs from ours, this reinforces the need to evaluate how agents respond to information, rather than assume that greater detail or accuracy improves collective outcomes.

We thank the reviewer for motivating this extension. Appendix G.5 now reports the noisy-disclosure results and explains their limits, while Section 5 (conclusion) also identifies truthful elicitation and verification as future work.


Q2. How consistent are the main disclosure effects when the models are evaluated repeatedly on the same markets? Based on my understanding, the current confidence intervals describe variation across markets, which does not directly answer this question.

Please see W2 above for the repeated evaluation results compared to our original results. The directions agree across all runs for GPT-5 mini, GPT-6 Sol, and Sonnet; whereas Gemini's smaller effects change sign, although on average the direction matches. The confidence intervals for the average effects exclude zero for the first three models and is statistically significant. Appendix G.2, Table 7, and Figure 23 report the new evaluation.


Q3. What generalizable insight should I take from the results beyond the observation? ... Could the authors clarify whether the findings support a rule for choosing disclosure policies, and whether that rule can be evaluated in markets not used to derive it?

The contribution is to turn disclosure into a design choice that can be selected and tested before deployment. A designer specifies the intended outcome, compares candidate information policies on a set of markets, and selects the policy that performs best for the models involved. Evaluating that choice on unseen markets then tests whether it is useful beyond the original observations. Our framework provides the behavioral measurements needed for this procedure, without assuming how LLM agents will respond to additional information.

To test whether these measurements can guide a practical choice, we added an experiment to test whether this approach works and summarize it in a new “Held-out policy selection” paragraph in the main text. In this experiment, for each model, we select among least, coarse, and full disclosure using average total gains on the original 90 trade markets. We then fix that choice before evaluating 60 new markets with previously unused utility distribution. Appendix G.4 and Table 9 provide the construction and full results.

Table R5. Performance on unseen markets after selecting disclosure on the original markets. Bold marks the policy selected on the original markets. ✓ indicates that it also achieves the highest test mean; ✗ indicates that another policy performs better.

ModelSelected policyLeastCoarseFullBest on test?
GPT-5 miniCoarse70.896.385.8✓
GPT-5.1Coarse86.998.891.7✓
GPT-6 LunaFull55.468.574.3✓
GPT-6 SolCoarse96.298.792.1✓
Gemini 3.7 FlashCoarse79.895.497.9✗
Gemini 3.1 ProCoarse91.999.397.3✓
Claude Sonnet 5Full67.675.387.8✓
Claude Opus 5.5Coarse81.799.494.5✓
Grok 4.3Full53.760.889.3✓
Grok 4.7Coarse94.897.394.0✓

The selected policy achieves the highest test mean for nine of ten models. Coarse remains best for six models, while full remains best for Luna, Sonnet, and Grok 4.3. Gemini Flash is the exception, with selected coarse disclosure trailing full. Every model benefits from some initial counter-party information, so withholding information is not a general solution either. These results show that measured disclosure effects can guide choices beyond the markets used to select them. They support model-specific policy selection within this market family; a different market or objective would require its own evaluation.


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

This paper studies how a variety of frontier LLMs empirically perform on simple economic test problems as the information available to them varies, while keeping the interaction rules and other parameters fixed. Throughout multi-resource trade, single-resource bargaining and double auction experiments, the GPT family of models suffer degraded outcomes as the level of information available to them increases, while other frontier LLMs showing varying trends, with Sonnet/Grok improving with more information in trading/bargaining but worsening in auctions. The main message of the paper is that agent performance does not monotonically improve with information access.

Critical strengths
  • The paper provides an empirical testbed for the folklore knowledge of "more information is not necessary better" for LLM agents. Though this is an intuitive and known phenomenon, I am unaware of previous works that explicitly test this.
Critical weaknesses
  • As I mentioned in the strengths, the main message of "model performances are affected by the information non-monotonically" is already known as folklore, and beyond this known message, I find it difficult to take any model level finding to be generalizable beyond the 3 test problems in the paper.
  • I feel the paper is restrictive in the types of information access levels it tests for. A very real concern when deploying LLM agents is that they often come across information that may be orthogonal/irrelevant to their task. However, the information levels designed in this paper are only increase the granularity of task relevant information, and do not test whether such irrelevant information acts as noise/poisons agent performance.

Though I find these weaknesses to be significant, I still tend towards a weak acceptance because as I mentioned in the strengths section, I am unaware of previous works in literature that specifically test out this folklore knowledge.

Questions for authors
  1. In Figure 4, how many trades were actually completed? If I understand correctly, complete rates are provided for bargaining/double auctions but not for the multi-resource trading.
  2. Do the authors have any intuition about what exactly is it about each of the three game types that causes a model like Sonnet to sometimes gain from more information in trade/bargaining, but suffer in auctions? I found the discussion about what are the characteristics of the games that cause this to be a bit insufficient in the current manuscript.
5 comments
AuthorsOpenReview ↗

Response to Reviewer vcSw (Part-1)

We thank the reviewer for their positive assessment and for recognizing the novelty of testing information disclosure in LLM interactions. First, we have revised the manuscript and expanded the evaluation from five to ten models, including GPT-6 Sol, Claude Opus 5.5, and Grok 4.7. Based on the sugestions, from you and other reviewers, we also added repeated-run, instruction and reasoning, held-out policy-selection, and noisy-information experiments to strengthen the evidence and examine its practical implications. Figures 1 and 3 and Appendix G present the updated results.

Below, we address each point and explain the corresponding revisions.


W1. As I mentioned in the strengths, the main message of "model performances are affected by the information non-monotonically" is already known as folklore, and beyond this known message, I find it difficult to take any model level finding to be generalizable beyond the 3 test problems in the paper.

Thank you for raising this question. Most importantly, our contribution is to show how accurate, relevant information affects LLM agents pursuing distinct interests; and how these effects can guide disclosure decisions. Prior work, such as Shi et al. [5], shows that irrelevant context can distract models and reduce reasoning accuracy. On the other hand, we examine a different problem, where even the true and relavant information can be useful to an individual agent yet harmful to the interaction as a whole. A buyer’s valuation, for example, can help a seller identify an acceptable price but also encourage a higher demand. Unlike cooperative systems such as GPTSwarm [3], negotiating agents may not benefit from the same outcome. Connecting information design [1] with LLM negotiation research [2], we measure both the opportunities agents identify and the gains they realize, revealing when additional information improves discovery without improving collective outcomes.

Furthermore, our revised evaluation expands from five to ten models, and harmful disclosure effects persist in newer frontier models. Full disclosure lowers GPT-6 Sol's trade gains from 97.7% to 90.8% and Claude Opus 5.5's auction gains from 94.6% to 64.1%. These findings make information access a consequential design choice for agents representing users, such as the ones in Project Deal [4]. This behaviour is not something that a designer can determine from model capability alone.

To test whether these measurements can guide a practical choice, we also added an experiment testing whether these measurements guide disclosure on unseen markets (Appendix G.4, Table 9). For each model, we select the highest-performing policy on the original 90 trade markets, then fix that choice before evaluating 60 new markets with previously unused utility combinations. All models and policies use the same test markets.

Table R1. Performance on unseen markets after selecting disclosure on the original markets. Bold marks the policy selected on the original markets. ✓ indicates that it also achieves the highest test mean; ✗ indicates that another policy performs better.

ModelLeastCoarseFullSelected policy best on test?
GPT-5 mini70.896.385.8✓
GPT-5.186.998.891.7✓
GPT-6 Luna55.468.574.3✓
GPT-6 Sol96.298.792.1✓
Gemini 3.7 Flash79.895.497.9✗
Gemini 3.1 Pro91.999.397.3✓
Claude Sonnet 567.675.387.8✓
Claude Opus 5.581.799.494.5✓
Grok 4.353.760.889.3✓
Grok 4.794.897.394.0✓

The selected policy achieves the highest test mean for nine of ten models. Every model also benefits from some initial counter-party information, so withholding information is not a general remedy. This supports using measured responses to select and validate disclosure on new markets within the same family. It does not require a model's preferred policy to remain unchanged across unrelated tasks. We added a summary to the main-text “Held-out policy selection” paragraph; testing this procedure under other mechanisms and in deployed marketplaces remains an important next step.


AuthorsOpenReview ↗

Response to Reviewer vcSw (Part-2)

W2. A very real concern when deploying LLM agents is that they often come across information that may be orthogonal/irrelevant to their task... I still tend towards a weak acceptance because as I mentioned in the strengths section, I am unaware of previous works in literature that specifically test out this folklore knowledge.

Thank you for highlighting this concern. Prior work examines distraction by irrelevant context [5] and difficulties using information in long contexts [6,7]. Our original experiment isolates a different problem: accurate, relevant information can also worsen interactions between agents with distinct interests. In this setting, better context processing does not necessarily resolve this tension, because the participants' objectives remain different.

To examine how these effects change when information is imperfect, we added noisy-information experiments across all three games with GPT-6 Sol and Claude Sonnet 5 (Appendix G.5, Table 10). Each model uses 18 market instances per game and condition. We multiply each counter-party value by an independent random factor drawn uniformly from 0.8 to 1.2, so a true value of 100 becomes a value between 80 and 120. Full disclosure presents the perturbed numbers; coarse disclosure presents rankings or profitability indicators derived from those same numbers. We compare accurate information with perturbed information presented either with or without a warning explaining the noise. The same perturbations are reused across models and disclosure policies, while agents’ own utilities and the true utilities used for scoring remain unchanged. This extension tests errors in task-relevant information, complementing prior work on irrelevant context.

Each cell shows coarse / full performance, with the higher mean in bold. Trade and auction scores are percentages of maximum possible gains; bargaining uses normalized Nash scores, with zero for disagreement.

Table R2. Coarse and full disclosure under accurate and noisy information.

SettingModelNo noiseNoise with warningNoise without warning
TradeGPT-6 Sol98.6 / 90.997.7 / 93.795.1 / 89.7
TradeClaude Sonnet 565.5 / 87.567.5 / 76.876.2 / 89.3
BargainingGPT-6 Sol0.729 / 0.8240.695 / 0.7500.768 / 0.664
BargainingClaude Sonnet 50.668 / 0.9810.664 / 0.6840.610 / 0.753
AuctionsGPT-6 Sol96.8 / 90.196.9 / 74.996.6 / 88.0
AuctionsClaude Sonnet 591.5 / 91.691.5 / 91.991.3 / 91.9

These results show that:

  • First, more information can still reduce collective gains when it is noisy. Even with a warning, full disclosure lowers Sol’s trade and auction gains relative to coarse disclosure by 4.0 and 22.0 percentage points. Sonnet’s mean trade gains remain higher under full disclosure, but its bargaining advantage shrinks.

  • Second, full-information outcomes vary more across the tested conditions in some settings.. For example, Sol’s auction gains fall from 90.1% without noise to 74.9% with noise and a warning, while coarse gains remain near 97%. Coarse signals can be less sensitive to numerical errors because perturbations need not change a ranking or profitability category. The practical implication is to test disclosure policies under expected information errors, rather than assume that the policy preferred with accurate information remains best.


Q1. In Figure 4, how many trades were actually completed? If I understand correctly, complete rates are provided for bargaining/double auctions but not for the multi-resource trading.

The original five-model trade experiment produced the following accepted exchanges, out of 90 negotiations per condition:

Table R3. Accepted exchanges under each disclosure policy in multi-resource trade.

ModelLeastCoarseFull
GPT-5 mini909089
GPT-5.1909090
Gemini 3.7 Flash909090
Claude Sonnet 5889090
Grok 4.3898890

Agreement is mostly common across all disclosure conditions. Agents can adjust both the resources exchanged and their quantities, giving them several ways to find acceptable terms before the interaction ends. This flexibility may help explain the high completion rates, even when the resulting agreements realize different amounts of the available gains.


AuthorsOpenReview ↗

Response to Reviewer vcSw (Part-3)

Q2. Do the authors have any intuition about what exactly is it about each of the three game types that causes a model like Sonnet to sometimes gain from more information in trade/bargaining, but suffer in auctions?

Each game has different strategic incentives and information structures, so the same disclosure policy do not produce the same behavior across games. For Sonnet, fuller disclosure appears to help overcome a coordination difficulty in multi-resource trade. Its coarse-information proposals leave more potential gains undiscovered, while exact preferences help it identify better bundles. Several other models already find high-quality exchanges from coarse information, leaving less scope for this benefit. In single-good bargaining, every agreement creates the same surplus, so information instead affects whether agents agree and how they divide those gains.

In auctions, agents compete through bids and asks, and profitable opportunities are realized only when those prices permit trade. Additional information can encourage higher seller asks or lower buyer bids, as each seeks a larger share. Figure 6 shows these changes in price demands alongside fewer completed trades. Thus, information that helps agents find better bundles can have different consequences when gains depend on compatible prices. These patterns help explain why disclosure has different effects across games. Differences in model capabilities and post-training may also influence how agents identify exchanges, make demands, and concede, but we cannot separate their contributions from the observed behavior alone.


References

[1] Dirk Bergemann and Stephen Morris. 2019. Information Design: A Unified Perspective. Journal of Economic Literature, 57(1):44–95.

[2] Federico Bianchi, Patrick John Chia, Mert Yuksekgonul, Jacopo Tagliabue, Dan Jurafsky, and James Zou. 2024. How Well Can LLMs Negotiate? NegotiationArena Platform and Analysis. arXiv:2402.05863.

[3] Mingchen Zhuge et al. 2024. GPTSwarm: Language Agents as Optimizable Graphs. ICML, PMLR 235:62743–62767.

[4] Anthropic. 2026. Project Deal: Our Claude-Run Marketplace Experiment. Research report, April 24.

[5] Freda Shi et al. 2023. Large Language Models Can Be Easily Distracted by Irrelevant Context. ICML, PMLR 202:31210–31227.

[6] Nelson F. Liu et al. 2024. Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12:157–173.

[7] Cheng-Ping Hsieh et al. 2024. RULER: What's the Real Context Size of Your Long-Context Language Models? COLM. arXiv:2404.06654.

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

This paper studies how information disclosure affects multi-agent LLM systems in resource trading, bargaining, and double auctions. It tests five LLMs under different levels of information about other agents’ values and costs, while keeping the markets and interaction rules fixed. The results show that more information does not always improve outcomes.

Critical strengths
  1. The paper is well written and has a clear logical structure.
  2. The authors conduct thorough experiments that provide strong support for their claims.
Critical weaknesses
  1. The motivation needs a clearer explanation. It is unclear why a platform would share one user’s private information with other users. If users know that their information may be shared, would they still report it truthfully? The paper does not discuss incentive compatibility in this setting. It assumes that all users report their information truthfully, which is a strong assumption that the authors should justify more clearly.
  2. The authors use the Nash product W ( p ) to measure the outcomes of information disclosure in bargaining. This score favors an equal split of the gains, but it is unclear why a more equal split should always count as a better outcome.
  3. The paper uses the term “information design”, but mainly compares fixed levels of information disclosure. It does not clearly explain what is being designed beyond these comparisons. And there are some minor problems.

Minor Issues

  1. In line 204, the phrases “what it offers to pay” and “what it requests” use closing quotation marks on both sides. In LaTeX, opening quotation marks should be written as``.
  2. In Figure 4(b) and (c), some colors are too similar to distinguish easily, especially those for Gemini and Sonnet, and for GPT-5.1 and Grok.
Questions for authors
  1. Consider a bargaining instance with c=0.3c = 0.3 and v=1.0v = 1.0 . Without information disclosure, the seller and buyer may agree on p=0.7p = 0.7, giving the seller a gain of 0.4 and the buyer a gain of 0.30.3. Suppose a disclosure policy aimed at maximizing the Nash product instead leads to p=0.65p = 0.65, giving each side a gain of 0.350.35. The buyer benefits from this change, but the seller receives less and may oppose the policy. How does the paper address this conflict between the platform’s objective and the seller’s interest?
  2. Building on Question 1, suppose the seller fully understands the platform’s disclosure policy and goal. To increase its own gain, the seller could report c′=0.6c ′ = 0.6 instead of its true cost c=0.3c = 0.3. If the resulting deal splits the reported surplus equally, the price would be p=0.8p = 0.8, giving the seller a true gain of 0.8−0.3=0.50.8-0.3=0.5. How do the authors prevent users from benefiting by reporting false information? What ensures that users have an incentive to report their true costs and values?
  3. Why is the Nash product W(p)W ( p ) an appropriate measure of bargaining outcomes in this setting? Under what assumptions should a more equal split be considered better?
  4. What exactly do the authors mean by empirical information design? Could they provide a formal definition?
9 comments
AuthorsOpenReview ↗

Response to Reviewer wDxo (Part-1)

We thank the reviewer for their thoughtful feedback and constructive suggestions. We have revised the manuscript and expanded the evaluation from five to ten models, including GPT-6 Sol, Claude Opus 5.5, and Grok 4.7. We also added experiments on repeatability, instructions and reasoning effort, held-out policy selection, and noisy information (Figures 1 and 3; Appendix G).

Below we address each point and identify the corresponding revisions.


W1. The motivation needs a clearer explanation. It is unclear why a platform would share one user’s private information with other users.

We thank the reviewer for the opportunity to clarify this motivation. A platform may share selected information to help participants find mutually beneficial exchanges, consistent with economic models [9]. A buyer’s preference rankings can help a seller propose a suitable bundle, whereas exact willingness to pay also reveals how much the seller could demand. GPT-5.1 illustrates this distinction where agents agree in all 90 trade markets under every condition, yet achieve 88.0% of maximum possible gains under least information, 97.2% with rankings, and 88.4% with exact values. Therefore, sharing rankings improves outcomes, while revealing full information reverses much of that benefit. The relevant choice is therefore not simply whether to share information, but how much to share.

Human negotiation research and information-design theory reinforce this motivation. Information design studies how disclosure shapes decisions and collective outcomes, including when the underlying information is truthful [2,6]. Sharing interests can improve negotiation outcomes [8], and conflicting interests can support informative but incomplete communication [10]. These findings provide reasons for sharing useful information without assuming that full disclosure is best.

Furthermore, obtaining reliable information and deciding what to disclose are connected but distinct design decisions. Our study focuses on the latter. In a setting like Project Deal [1], users provide instructions and preferences to agents acting on their behalf; not every detail needs to reach their negotiating partners. When participants choose what to report, however, elicitation and disclosure must be considered together [2,3].

Our new noisy-information experiments examine disclosure when information is inaccurate (Appendix G.5, Table 10). Coarse rankings and profitability categories were less sensitive to the tested perturbations than full-information of values, motivating further study of selective disclosure to limit the influence of inaccurate reports. Our answer to Q2 below examines how this argument applies to the reviewer’s strategic-misreporting example. We have also expanded Section 5 to discuss jointly designing disclosure, truthful elicitation, and verification as future work.


AuthorsOpenReview ↗

Response to Reviewer wDxo (Part-2)

W2. The authors use the Nash product W(p) to measure the outcomes of information disclosure in bargaining. This score favors an equal split of the gains, but it is unclear why a more equal split should always count as a better outcome.

We thank the reviewer for prompting us to clarify this choice. We use the Nash product to evaluate a symmetric bargaining objective, not to claim that equal sharing is always preferable. A marketplace can complete trades while leaving one side with almost none of the gains. A designer concerned with satisfaction and continued participation may therefore examine how gains are divided alongside whether agreement occurs. Similarly, Project Deal evaluates user satisfaction [1], and human negotiation research connects subjective experience with willingness to negotiate again [5].

In single-good bargaining, every agreement creates the same surplus, (v−p)+(p−c)=v−c(v-p)+(p-c)=v-c, regardless of price. Agreement rate therefore captures normalized total gains but cannot distinguish a 99:1 division from a 50:50 division. The normalized Nash product gives both parties’ gains relative to disagreement equal weight [4], favoring balanced division in our fixed-surplus setting. Agents are nevertheless instructed to maximize their own gains; where the score is calculated after negotiation and do not guide their behavior.

This evaluation choice has precedent in automated negotiation. ANAC reports the Nash product alongside individual utilities, total welfare, and agreement rates [12], while Welfare Diplomacy uses Nash welfare to evaluate interactions between LLM agents [13], and GLEE distinguishes efficiency from fairness [7]. We similarly report complementary measures rather than treating one score as a universal definition of success. In our paper, Figure 5 already separates agreement and balance across message limits, and the new Appendix G.1, Table 6, provides the complete ten-model breakdown.

Table R1. Agreement and surplus division for the five newly added models. Entries show coarse → full disclosure, combining both starting roles. The final column includes zero scores for disagreements.

ModelAgreement rate (%)Nash score among agreementsNash score including disagreements
GPT-6 Luna83.3 → 100.00.807 → 0.8250.672 → 0.825
GPT-6 Sol83.3 → 100.00.854 → 0.8260.712 → 0.826
Gemini 3.1 Pro80.0 → 100.00.878 → 0.9700.703 → 0.970
Claude Opus 5.580.0 → 100.00.897 → 0.9230.718 → 0.923
Grok 4.780.0 → 100.00.885 → 0.1660.708 → 0.166

For Grok 4.7, full disclosure increases agreement but makes gains substantially more one-sided; for Gemini 3.1 Pro, it improves both. The preferred policy therefore depends on the designer’s objective. We have clarified this motivation in Section 5 and Appendix G.1. Importantly, the losses reported in trade and auctions are reductions in total gains themselves.


W3. The paper uses the term “information design”, but mainly compares fixed levels of information disclosure. It does not clearly explain what is being designed beyond these comparisons.

Thank you for asking us to clarify the design contribution. The object being designed is the information policy, which determines what each agent learns about its counterpart and at what precision [6]. We evaluate candidate policies through actual LLM interactions, since their effects depend on how agents use the information.

To test whether these measurements can guide a practical choice, we added a held-out policy-selection experiment to test this design procedure (Appendix G.4, Table 9). For each model, we select the policy with the highest average total gains on the original 90 trade markets, then evaluate that fixed choice on 60 new markets with previously unused utility combinations. The table below reports test gains as percentages of the maximum possible; bold numbers mark the policy selected before observing test outcomes.

Table R2. Performance on unseen markets after selecting disclosure on the original markets. Bold marks the policy selected on the original markets. ✓ indicates that it also achieves the highest test mean; ✗ indicates that another policy performs better.

ModelSelected policyLeastCoarseFullSelected policy best on test?
GPT-5 miniCoarse70.896.385.8✓
GPT-5.1Coarse86.998.891.7✓
GPT-6 LunaFull55.468.574.3✓
GPT-6 SolCoarse96.298.792.1✓
Gemini 3.7 FlashCoarse79.895.497.9✗
Gemini 3.1 ProCoarse91.999.397.3✓
Claude Sonnet 5Full67.675.387.8✓
Claude Opus 5.5Coarse81.799.494.5✓
Grok 4.3Full53.760.889.3✓
Grok 4.7Coarse94.897.394.0✓
AuthorsOpenReview ↗

Response to Reviewer wDxo (Part-3)

W3 contd.

Overall, the selected policy achieves the highest test mean for nine of ten models. Coarse remains best for six models, while full remains best for Luna, Sonnet, and Grok 4.3. Every model benefits from some initial counter-party information, so withholding information is not a general solution either. These results support model-specific policy selection within this market family; a different market or objective would require its own evaluation.


Q1. Consider a bargaining instance ... The buyer benefits from this change, but the seller receives less and may oppose the policy. How does the paper address this conflict between the platform’s objective and the seller’s interest?

Thank you for the thoughtful question. Our framework considers a designer optimizing an overall objective, rather than maximizing every participant’s individual payoff. This distinction is standard in mechanism and information design. Myerson’s optimal auction, for example, maximizes the seller’s expected payoff subject to incentive and participation constraints, not every bidder’s payoff [11]. Information design likewise allows the designer’s objective to differ from participants’ interests [6]. In your example, the seller may therefore prefer the original agreement even when the designer prefers the alternative consistent with the given literature.

We use the Nash score to study one such objective that distinguishes modest imbalances from extreme divisions. A 40:60 split scores 0.96, close to the maximum of 1, whereas a 1:99 split scores only 0.0396. Modest redistributions from equal sharing therefore receive only a small penalty. This distinction may matter to a platform if persistently one-sided outcomes discourage participation. Project Deal [1] also illustrates why monetary outcomes and satisfaction should be distinguished. There, users represented by Haiku obtained worse prices than those represented by Opus, yet reported similar perceived fairness and no statistically significant difference in user satisfaction. Monetary gains alone therefore need not capture how users evaluate an interaction.

We therefore report agreement, individual gains, and balance so designers can assess these trade-offs. The revised Section 5 now discusses satisfaction and continued participation, including the need to evaluate whether users prefer the resulting outcomes to their alternatives.


Q2. Building on Question 1, suppose the seller fully understands the platform’s disclosure policy and goal. To increase its own gain, the seller could report .. instead of its true cost. What ensures that users have an incentive to report their true costs and values?

Our coarse information policy offers a way to address this manipulation by preventing the exaggerated cost from changing what the buyer learns, while preserving the signal that a beneficial trade is possible. Let the true surplus be S=v−c>0S=v-c>0. If the seller reports c^=c+δ\hat{c}=c+\delta, where 0<δ<S0<\delta<S, the coarse signal remains unchanged:

C(c+δ)=1{c+δ<v}=C(c).C(c+\delta)=\mathbf{1}\{c+\delta<v\}=C(c).

If the report affects negotiation only through this signal, holding other inputs fixed, the exaggeration cannot change the buyer’s response distribution or the seller’s expected payoff.

Accurate information also helps participants identify beneficial exchanges. Exaggerating the cost beyond the buyer’s valuation would make the signal indicate that no profitable trade exists, potentially preventing agreement and costing the seller its gains. There is therefore a coordination incentive to avoid reports that conceal beneficial exchanges, although this does not guarantee that reporting the exact true cost is optimal.

Under full information, the altered number is visible. Under the reviewer’s proposed rule of equally splitting reported surplus,

u_s=\frac{S+\delta}{2},\qquad u_b=\frac{S-\delta}{2}.$$ The seller gains $\delta/2$ at the buyer’s expense, and the Nash score based on true utilities falls to $$W=1-\left(\frac{\delta}{S}\right)^2.$$ Our agents negotiate prices rather than follow this rule, but the example identifies how coarse information can prevent an inflated cost from reaching the buyer through the disclosure signal. Our bargaining results also show that full information can amplify first-mover advantages even when the information is accurate. **These observations motivate evaluating disclosure not only for coordination, but also for its ability to limit gains from strategic misreports.** We have added the joint study of disclosure, truthful elicitation, and verification as future work in Section 5. ---
Reviewer_XzdtOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how varying levels of information sharing affect LLM agents in economic scenarios. Through various experiments with 5 different models, the authors examine how well the models find good deals for themselves and what gains they actually achieve. The results indicate, however, that more information does not always improve outcomes. Its effect depends on the model and type of interaction.

Critical strengths

Interesting problem: This paper investigates how much information should be shared between LLM agents. Whilst more information can lead to good deals, the findings show that it can also cause agents to become more selfish and reduce overall gains. Well designed experiments: The authors performed experiments on 5 models across various economic scenarios, including long conversations, more complex markets and interactions between different models. They also compared how well agents found good deals with how good the final outcomes were. This comparison helped highlight why finding good deals doesn’t always lead to good outcomes.

Critical weaknesses

No repeated experiments: The experiments weren’t repeated with varying model outputs. Since LLM agents can produce different outputs each run, it is difficult to know whether the results are consistent.

Not generalizable: The experiments used relatively simple economic games with clear objectives and fixed values. When the markets became more complex, the results also varied. In the 10 resource setting, all models actually improved with full information. It is therefore unclear how generalizable the results are.

Eval metrics: Within the bargaining scenario, the main metric rewards agents for sharing gains equally. A low score therefore didn’t always mean a bad economic outcome. Separating fairness and efficiency would have made the evaluation better.

Questions for authors

Would it be possible to repeat some experiments with varying random seeds or multiple model runs? Doing so would help establish whether the results are consistent.

How would the results vary if good outcomes were measured using different metrics? The bargaining score focuses on equal gains, so it would be good to know whether full information actually reduces overall gains or just changes how they are shared.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Xzdt (Part-1)

We thank the reviewer for their positive assessment and constructive suggestions. First, we have revised the manuscript and expanded the study from five to ten models, including GPT-6 Sol, Claude Opus 5.5, and Grok 4.7. We also added repeated-run, instruction and reasoning, held-out policy-selection, and noisy-information experiments (Figures 1 and 3; Appendix G). These additions strengthen the evidence for our findings and demonstrate how the framework can guide disclosure choices.

Below we address each point and identify the corresponding revisions.


W1. No repeated experiments: The experiments weren’t repeated with varying model outputs. Since LLM agents can produce different outputs each run, it is difficult to know whether the results are consistent.

Thank you for highlighting the need to test variability across model responses. In response to this, we added repeated runs to test whether the findings hold when models negotiate again on identical markets. Each of four models completed three fresh runs on the same 30 market set under both coarse and full disclosure. Prompts, roles, message limits, and generation settings stay fixed. Overall, we see that the repeated averages agree in direction with the original single-run results.

The table below compares our original 90-market run averages with the new repeated 30-market averages. Entries show coarse → full gains; differences and intervals use the repeated runs.

Table R1. Original trade results compared with averages across three fresh runs.

ModelOriginal single run (90 markets)Three fresh runs (30 markets)Repeated full minus coarse (percentage points)Repeated 95% CI
GPT-5 mini93.9% → 83.7%91.1% → 84.0%-7.2[-12.3, -1.8]
GPT-6 Sol97.7% → 90.8%97.6% → 91.6%-6.0[-9.2, -2.9]
Gemini 3.7 Flash93.7% → 92.2%93.4% → 92.2%-1.2[-4.9, +2.8]
Claude Sonnet 572.1% → 85.6%70.5% → 83.2%+12.8[+5.4, +21.3]

To show variation between runs directly, we also report each run’s average disclosure effect. Positive values favor full disclosure; negative values favor coarse disclosure.

Table R2. Full-minus-coarse trade gains in each repeated run (percentage points).

ModelRun 1Run 2Run 3
GPT-5 mini−1.9−4.6−14.9
GPT-6 Sol−5.6−5.6−6.8
Gemini 3.7 Flash+2.0−1.9−3.6
Claude Sonnet 5+16.5+8.9+12.9

Full disclosure lowers GPT-5 mini’s and GPT-6 Sol’s average gains and raises Sonnet’s in every run, matching the original findings. The effects averaged across runs are statistically significant, with all three 95% confidence intervals excluding zero. Only Gemini’s smaller effect changes direction across runs and is not statistically significant, where its coarse to full information difference was already low in the initial experiments.

Further details regarding these repeated-run results appear in Appendix G.2, Table 7, and Figure 23; the original results appear in Appendix G.1, Table 5.


AuthorsOpenReview ↗

Response to Reviewer Xzdt (Part-2)

W2. The experiments used relatively simple economic games with clear objectives and fixed values. When the markets became more complex, the results also varied. In the 10 resource setting, all models actually improved with full information. It is therefore unclear how generalizable the results are.

Thank you for asking us to clarify the relevance of these settings to practical agent markets. Our settings study an information-design decision that also arises in practical agent markets, where agents represent different users and negotiate over competing interests. Project Deal [2] illustrates why marketplace size and the scope of an individual negotiation should be distinguished. Its marketplace contained over 500 listings in total, but reported negotiations included two agents bargaining over a single bicycle or gemstone. Our similar bilateral settings examine this transaction-level interaction, rather than the marketplace-search stage.

The ten-resource experiment expands the exchanges agents must consider. One evaluated market permits approximately 57 million distinct two-way trades, and exact preferences help agents identify better exchanges within this large space (Figure 7(a); Figure 10 and Appendix G.3). The benefit of fuller disclosure therefore depends on the trading problem, not market size alone. Auctions introduce a different strategic challenge through competition among multiple buyers and sellers, yet full disclosure generally lowers total gains, including for Claude Opus 5.5 and Grok 4.7. These contrasting findings highlight the importance of interaction rules and incentives, consistent with economic theory distinguishing information’s private and social value [1,6,7]. The expanded evaluation also shows harmful effects persist in frontier models, with GPT-6 Sol’s trade loss recurring across all three repetitions.

To test whether these measurements can guide a practical choice, we also added an experiment testing whether these measurements guide disclosure on unseen markets For each model, we selected the policy with the highest mean gains on the original 90 trade markets, then evaluated that fixed choice on 60 new markets with previously unused utility combinations (Appendix G.4, Table 9).

The table below reports held-out gains as percentages of the maximum possible. Bold marks the policy selected before testing; ✓ indicates that it also achieves the highest test mean, while ✗ indicates that another policy performs better.

Table R3. Performance on unseen markets after selecting disclosure on the original markets.

ModelSelected policyLeastCoarseFullSelected policy best on test?
GPT-5 miniCoarse70.896.385.8✓
GPT-5.1Coarse86.998.891.7✓
GPT-6 LunaFull55.468.574.3✓
GPT-6 SolCoarse96.298.792.1✓
Gemini 3.7 FlashCoarse79.895.497.9✗
Gemini 3.1 ProCoarse91.999.397.3✓
Claude Sonnet 5Full67.675.387.8✓
Claude Opus 5.5Coarse81.799.494.5✓
Grok 4.3Full53.760.889.3✓
Grok 4.7Coarse94.897.394.0✓

The selected policy achieves the highest test mean for nine of ten models. This demonstrates that measured disclosure effects can guide choices beyond the original observations. For a more complex market, the same approach could be used to learn a disclosure policy from interactions under its rules and objectives, then test that policy on separate, unseen interactions before deployment.


AuthorsOpenReview ↗

Response to Reviewer Xzdt (Part-3)

W3. Eval metrics: Within the bargaining scenario, the main metric rewards agents for sharing gains equally. A low score therefore didn’t always mean a bad economic outcome. Separating fairness and efficiency would have made the evaluation better.

First, we would like to clarify that we do not reward agents for sharing equally. Each agent is instructed to maximize its own gain, while the Nash score is calculated only after negotiation to evaluate the outcome. A lower score does not necessarily imply lower economic efficiency, which is why Figure 5 reports agreement and balance separately. We have clarified this distinction in the revised text and added a ten-model breakdown in Appendix G.1, Table 6.

Also, we use the Nash score because total gains alone cannot distinguish bargaining outcomes once a trade occurs. Every agreement within a market creates (v−p)+(p−c)=v−c(v-p)+(p-c)=v-c, regardless of price. Agreement rate therefore captures mean normalized total gains but treats a 99:1 division and a 50:50 division identically. The normalized Nash product provides a standard symmetric bargaining criterion that gives equal weight to both parties’ gains relative to disagreement [3]. This is relevant when a designer cares about serving both sides, rather than completing trades alone; it does not imply that equal sharing is universally preferable. Examining division alongside efficiency also follows experimental bargaining research and benchmarks such as GLEE [4,5].

The newly added table as reproduced below separates agreement, balance among agreements, and the combined score for all ten models. (Appendix G.1, Table 6) Entries show coarse → full disclosure at six messages, with 30 negotiations per model and condition across both starting roles. The combined score assigns zero to disagreements and equals the agreement rate, expressed as a fraction, multiplied by the mean Nash score among agreements.

Table R4. Separating agreement and surplus division in bargaining across all ten models.

ModelAgreement rate (%)Nash score among agreementsNash score including disagreements
GPT-5 mini76.7 → 93.30.810 → 0.1740.621 → 0.162
GPT-6 Luna83.3 → 100.00.807 → 0.8250.672 → 0.825
GPT-5.193.3 → 93.30.823 → 0.6090.768 → 0.568
GPT-6 Sol83.3 → 100.00.854 → 0.8260.712 → 0.826
Gemini 3.7 Flash80.0 → 100.00.882 → 0.9830.706 → 0.983
Gemini 3.1 Pro80.0 → 100.00.878 → 0.9700.703 → 0.970
Claude Sonnet 580.0 → 100.00.826 → 0.9640.661 → 0.964
Claude Opus 5.580.0 → 100.00.897 → 0.9230.718 → 0.923
Grok 4.383.3 → 100.00.761 → 0.8260.635 → 0.826
Grok 4.780.0 → 100.00.885 → 0.1660.708 → 0.166

Full disclosure generally increases agreement, but can make gains more one-sided. Figure 5 already separates agreement rates from balance among agreements across message limits; the new table makes this distinction explicit for all ten models. Our added Section 5 discussion connects distribution to a designer's interest in satisfaction and continued participation, also considered in Project Deal [2], without treating equal sharing as universally preferable. In trade and auctions, the reported losses instead concern total gains, with no penalty for inequality.


Q1. Would it be possible to repeat some experiments with varying random seeds or multiple model runs? Doing so would help establish whether the results are consistent.

Yes. We added three fresh runs on identical markets for four models, totaling 720 new negotiations. Please see our response to W1 above for the results with a comparison with the original experiments. The repeated evaluation appears in Appendix G.2, Table 7, and Figure 23 of the revised manuscript.


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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.