acceptodds
Under review as a conference paper at ICLR 2027

Scaling Capabilities and Competition in Strategic Interactions

Abstract

LLM agents are becoming autonomous and ubiquitous, moving us toward a world where agent-agent interactions carry real consequences for the people and institutions they represent. These interactions will be strategic, mixed-motive, and structurally diverse, involving agents that differ in capability and number. We introduce three new multi-turn negotiation games grounded in real-world scenarios, designed to enable experiment sweeps that systematically scale agent capability, number of agents, and degree of competition. Across runs and models, more capable agents reliably achieve higher payoff, with greater inequality in capability-diverse groups than homogeneous ones. For models with an Elo above , we observe the emergence of a ruthless coalitioning behaviour, where agents form the minimum supermajority needed to complete the negotiation while allocating zero payoff to excluded agents. Moreover, we find that a team of weaker agents can outcompete a strong agent, but the advantage degrades as the number of agents scales, as larger, more complex games magnify capability gaps and facilitate deceptive exploits such as falsely claiming team membership. Finally, we also scale test-time compute, finding negotiation payoff gains comparable to - Elo points in GPT-5 and Claude Sonnet 4.6.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
agentic inequality, multi-agent systems, cooperative AI, LLM agents, negotiation benchmarks, strategic AI

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_GyM7OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper analyses the impact of LLM capabilities and test time compute in strategic multi agent negotiations. Considered domains for this LLM benchmark are item allocation, diplomatic negotiation and public good funding. Overall 30 models are considered in 7160 runs. Highly strategic patterns like exclusion occur only in highly capable models. Further coordinated weaker agents can outperform stronger agents, however this advantage does not increase with team size.

Critical strengths

The choosen environments cover competition, capapbitly and group size, three differnt relevant axis across 30 models and 7160 runs which is a substantial amoutn of experimentation.

Consistent capability and test time compute relationships established.

Critical weaknesses

Observing the behavior of minimal coalitions and exclusions in LLM is interesting. The concepts are well know and within the training data. So is this amplified by certain methods/ optimisaiton pressure?

Models of weaker capabilities might be overwhelmed with too many items within their context. How much of their degredation is due to capabilities (in context) vs actual strategic deliberations?

Questions for authors

Thanks for the llama baseline that is a first step towards cross provider negotations (at matched elo) Can you please extend this to within lab and cross-lab negotiations.

Can you please ablate the impact of choosen prompt for different model capabilities and see when exclusion of certain agents or formation of minimum winning coalitions emerge.

Please attribute the capability payoff relationship to a) strategic decision makind and b) miscellaneous (arithmetic errors, invalid, parsing etc)

How to make teams of weak agents scale with team size?

2 comments
AuthorsOpenReview ↗

Response to Reviewer GyM7 [Part 1 of 2] Free-form Negotiation Scenarios and Teams of Weaker Agents [W1, W2]

Dear Reviewer GyM7,

Thank you for the feedback! Please see the supplementary material, ICLR 2027 Rebuttal Figures, for Rebuttal Figures 1–19 referenced below.

Response to Critical Weakness 1:

Observing the behavior of minimal coalitions and exclusions in LLM is interesting. The concepts are well know and within the training data. So is this amplified by certain methods/ optimisaiton pressure?

We agree that minimum-winning coalitions are established concepts that may have appeared in the training data when the models were trained. The goal and purpose of this paper was to take those finished models that have already been trained, and to study their behavior and scaling properties when placed in a wide range of negotiation settings. How the characteristics that we observe these models exhibit during negotiation came about through its training process is both (1) unobservable to us (since we do not have access or visibility to the training process and intermediary checkpoints of these models during their training/optimization), and (2) it is a question which is outside of the scope of this paper.

Response to Critical Weakness 2:

Models of weaker capabilities might be overwhelmed with too many items within their context. How much of their degredation is due to capabilities (in context) vs actual strategic deliberations?

This is a great point! We find that much of the degradation is indeed driven by the increased complexity of reasoning over more items. We ran an additional experiment comparing (1) scaling the number of items alongside the number of weak agents, as in our original setup, vs. (2) holding the number of items constant at 10 as the team grows. As shown in Rebuttal Figure 15, the degradation disappears when the number of items is held constant. This supports our hypothesis that weaker agents struggle with increased game complexity, but it does reveal a new learning that a significant portion of this complexity comes from juggling more items rather than just the raw number of agents itself.

We expect that, in real-world settings, larger groups of agents will generally operate in larger, more complex environments (i.e. analogous to more items). That’s why the default setting in our environment is to also scale the number of items along with scaling the number of agents, such that each agent has the same average number of items to begin. This new experiment both supports our original hypothesis and provides a valuable new insight into the root source of the degradation.

AuthorsOpenReview ↗

Response to Reviewer GyM7 [Part 2 of 2] Within-lab vs Cross-lab, Prompt Ablations, Strategic Decision-making vs Misc, Scaling with Team Size [Q1, Q2, Q3, Q4]

Response to Question 1:

Thanks for the llama baseline that is a first step towards cross provider negotations (at matched elo) Can you please extend this to within lab and cross-lab negotiations.

Yes! We run an analysis to compare the within-lab scaling trend versus the cross-lab scaling trend for all three games. Specifically, because our baseline is GPT-5 nano, we investigate the scaling trend of within-lab (GPT-5 nano vs. other OpenAI models) vs. cross-lab (GPT-5 nano vs non-OpenAI models). In Rebuttal Figure 16, we show that we see no statistically significant difference in the payoff trend across the three.

Response to Question 2:

Can you please ablate the impact of choosen prompt for different model capabilities and see when exclusion of certain agents or formation of minimum winning coalitions emerge.

Yes. Here, we ablate three different prompts. In addition to the current default prompt, we ablate a version of the prompt which contains language that encourages the model to consider being collaborative (“Pursue your own utility while also caring about each other agent's utility. Work toward an agreement that benefits everyone.”). We also use a third prompt variant where we insert language in the prompt to encourage the model to be competitive and emphasize that its goal is to maximize its own payoff without caring about others' payoffs (“Your sole goal is to maximize your own total utility after discounting; other agents' utilities do not matter to you.”).

We find a very interesting effect on the emergence of exclusionary collisions. Under the default prompt, exclusionary coalitions emerge around Elo 1422 and 1468. Under the competitive prompt, they emerge earlier and more frequently, beginning at Elo 1422. Under the collaborative prompt, the emergence of exclusionary coalitions disappears altogether. Thus, these ablations reveal that prompting can actually influence the effect of exclusionary coalitions in an interesting way.

Response to Question 3:

Please attribute the capability payoff relationship to a) strategic decision makind and b) miscellaneous (arithmetic errors, invalid, parsing etc)

We see that some percentage of it is, but it is not statistically significant. Even when removing all execution issues in terms of parsing errors or invalids, and then also when additionally removing numerical or arithmetic issues, we still see a clear positive slope. Please see Rebuttal Figure 18 for the exact results.

Response to Question 4:

How to make teams of weak agents scale with team size?

It would be hard for the performance of the weaker agent team to further scale because it already achieves near 100% performance with the 3-agent team. However, we found a way to at least ensure that the performance does not degrade as we scale up the number of agents, which is by holding the complexity of the environment roughly constant as we scale the team. In our experiment where we hold the number of items constant at 10, we do not see a degradation as we scale team size, and we actually see a small improvement scaling from 3-agents to 9-agent teams. Please see Rebuttal Figure 19 for the exact result here!

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

The paper studies agents' negotiation and interactions in strategic, mixed-motive settings. The paper studies scaling by comparing models with different Elo scores with respect to their achieved payoffs. It finds coalition behavior formation in more capable models, where agents form the minimum supermajority needed to complete the negotiation. The paper also studies how and whether teams of weaker agents can outcompete a strong agent.

Critical strengths
  • The paper is well presented and easy to follow, which is very important these days. I have minor suggestions such as adding table of context for the appendices. The appendix and the main body contain enough information to understand the setups of the games. But having for example 1 overall figure that walks through the details of the game and/or show snippets of transcripts would be useful to get a broad view and for further accessibility.

  • I found the analysis of exclusionary coalitions very interesting, including figure 6, and the detailed explanation of why this happens, that weaker agents may not be consistent in voting.

  • The paper has strong experiments with many models, variants, games with clear systematic evaluation such as homogeneous and heterogeneous groups and metrics, test time scaling, fairness metrics.

Critical weaknesses
  • It is unclear how the findings translate to more real-world setups beyond games with clear, explicit payoffs. For example, the insight about more capable models moving beyond fair share is interesting, but would this translate to other free-form negotiation scenarios? I think the paper is claiming in the abstract when saying that games are grounded in "real-world scenarios". But I think the paper is still showing interesting insights such as teams of agents and excluding agents.

  • I would be curious to study more qualitative or pattern analysis such as how agents organize themselves in the "Teams of weaker agents" experiment, also whether agents identify in their reasoning traces that they are moving beyond the fair share.

Questions for authors
  • Could the authors please relate the games used in the paper to previous games in the literature? Is the main contribution of the paper studying scaling or proposing new games?

  • What does "Requested reasoning effort" mean in figure 4?

2 comments
AuthorsOpenReview ↗

Response to Reviewer oicN [Part 1 of 2] Free-form Negotiation Scenarios and Teams of Weaker Agents [W1, W2]

Dear Reviewer oicN,

Thank you for the feedback! Please see the supplementary material, ICLR 2027 Rebuttal Figures, for Rebuttal Figures 1–19 referenced below.

Response to Critical Weakness 1:

It is unclear how the findings translate to more real-world setups beyond games with clear, explicit payoffs. For example, the insight about more capable models moving beyond fair share is interesting, but would this translate to other free-form negotiation scenarios? I think the paper is claiming in the abstract when saying that games are grounded in "real-world scenarios". But I think the paper is still showing interesting insights such as teams of agents and excluding agents.

We use games with explicit payoffs because they let us measure outcomes and vary competition, group size, and model capability in a controlled way. This structure makes it possible to compare settings quantitatively and identify how each factor relates to negotiation outcomes.

There is no reason why the general results wouldn’t apply to more free-form games. To show this, we run our experiments on a free-form, unstructured game, where two agents (one buyer, one seller) are allowed to negotiate in a fully free-form manner. Across six adversary models, we see the same payoff-capability scaling trend, as payoff increased by an average of +16.42 per 100 Elo (see Rebuttal Figure 13).

Response to Critical Weakness 2:

I would be curious to study more qualitative or pattern analysis such as how agents organize themselves in the "Teams of weaker agents" experiment

That’s a great question! In this experiment, we assigned the weaker agent team a process where they would propose allocations, review each other's proposals, and the captain would select the final allocation as the one to represent the team. Within that process, qualitatively, agents often decided to correct each other on incorrect calculations, which happened in 61.25% of the game, and also how sensible the proposal is, which happened in 57.5% of the games. This was effective, as 40% of the games the correction by an agent made it to the final proposal from the captain, affecting the outcome for the whole team. Because the agents were “weak” agents, 36.8% of the time, the correction proposed was actually an incorrect correction, but this disrupted a much smaller percentage of games, 11.25%. Overall, the team structure of proposal review and selection was very effective for the agents at reviewing each others’ work and maximizing the chance of proposing a solid final allocation.

whether agents identify in their reasoning traces that they are moving beyond the fair share.

Yes! Some agents actually do explicitly reason that they're moving beyond the fair share. For example, in one N=2 Item allocation game on the max competition setting, a Qwen3 Max agent stated “any split that gives one agent >50 is inherently 'unfair' in total points—but fairness isn't the goal; unanimous acceptance is”. In another, a Gemini 2.5 Pro agent stated “For me, this is still an excellent outcome. It secures my absolute top priority, the Stone (worth 58 points), and provides me with a total utility significantly higher than the 50-point 'fair' split”... So this does happen, however it is a very small percentage of runs where the agents explicitly do this, 5 out of 1500 runs (or 0.33%).

AuthorsOpenReview ↗

Response to Reviewer oicN [Part 2 of 2] Previous Games and Requested Reasoning Effort [Q1, Q2]

Response to Question 1:

Could the authors please relate the games used in the paper to previous games in the literature? Is the main contribution of the paper studying scaling or proposing new games?

The main contribution of this paper is studying scaling — it is not proposing new games. The 3 games proposed in this paper are not extremely groundbreaking / uniquely special in terms of the game itself - for example, NegotiationBench also proposes 3 similarly simple games (resource exchange, multi-turn ultimatum, and seller–buyer negotiation). Davidson et al. (2024) also proposes structured two-agent negotiations over issues with aligned or conflicting interests. Abdelnabi et al. (2024) proposes multi-agent negotiations over multiple issues with private preferences. Where our game is unique and different is that it is designed to be both simple yet flexible in a way that enables degrees of freedom to allow us to scale:

  • the number of agents
  • the number of models
  • the degree of competition
  • the complexity via the number of items, etc.

We can continuously scale across each of these dimensions, allowing us to systematically study the scaling laws governing strategic interactions between agents in negotiation games. To summarize, the main contribution of this paper lies in the scaling dynamics, findings, and insights, rather than the novelty of the games themselves. The three games are deliberately simple, and we do not claim that they are groundbreaking. Instead, we design them to be clean, simple, and flexible, enabling us to systematically and simultaneously vary multiple dimensions and study how strategic interactions scale across them.

Response to Question 2:

What does "Requested reasoning effort" mean in figure 4? The “Requested reasoning effort” means the API setting used to ask the model to spend more or less effort reasoning before responding.

The four plotted positions correspond to:

  • GPT-5 and Gemini 3 Flash: minimal → low → medium → high.
  • Claude Sonnet 4.6: low → medium → high → max.
Reviewer_ffi4OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how the capability of large language model (LLM) agents, the number of participants, and the degree of competition jointly shape individual payoffs and collective outcomes in mixed-motive negotiations. The authors construct three multi-round negotiation environments (item allocation, diplomatic treaty negotiation, and the co-funding of public projects) and, across 30 models and 7,160 experimental runs, compare conditions that vary in capability level, group composition, and reasoning effort.

The main findings are as follows. Stronger models generally attain higher payoffs, but their effect on other participants depends on the underlying structure of cooperation and competition. Some highly capable models form exclusionary coalitions. Coordinated teams of weaker models can counterbalance stronger models, yet they struggle to sustain optimal performance in larger-scale games. Finally, the payoff gains from increased reasoning effort vary across models.

Critical strengths
  1. The core capability–payoff relationship is supported by broad empirical coverage and robustness checks. The paper observes consistent trends across three distinct negotiation structures, rather than relying on a single task or a fixed set of opponents. The authors vary the baseline opponent and additionally analyze randomly paired two-agent experiments; the appendix further validates the main relationship using alternative capability metrics. Together, these checks strengthen the credibility of the central observation that stronger models generally achieve higher negotiation payoffs, and they constitute an important basis for acceptance.
Critical weaknesses
  1. The paper provides a broad set of empirical results, but its interpretation of several key findings still relies mainly on observational correlations and individual negotiation case studies. For example, the emergence of exclusionary coalitions is interpreted as a behavioral shift that occurs once a certain capability threshold is reached, while the declining performance of larger weak-model teams is attributed to coordination overhead and exploitation by opponents. The current analysis does not sufficiently disentangle the respective roles of model capability, institutional advantage, and task complexity. This does not diminish the value of the reported phenomena, but it limits the paper’s explanatory power regarding why these phenomena occur and under what conditions. More explicit mechanistic hypotheses, together with targeted analyses capable of distinguishing among alternative explanations, would strengthen the scientific contribution of this work.

  2. Although the paper covers three distinct utility structures, these environments share the same underlying protocol: supermajority voting, per-round discounting, and zero payoff in the event of no agreement. In particular, the key finding on exclusionary coalitions depends directly on the ability of a subset of participants to pass a proposal over the objections of others. Consequently, consistency across games does not yet adequately establish which conclusions reflect a more general relationship between capability and competition, and which are primarily determined by the chosen protocol. The paper should delineate the conditions under which its conclusions hold more explicitly, and clarify the scope to which these findings generalize, either through a limited protocol sensitivity analysis or by narrowing the corresponding claims.

Questions for authors
  1. Do the authors have results for a condition that holds the shared information, team objective, and binding voting power fixed, but assigns planning to a single Nano decision-maker? With reasoning budgets matched as closely as possible, does the multi-agent team still retain an advantage?

  2. After controlling for average group capability, game type, number of participants, and competition level, how much of the difference in payoff variance between heterogeneous and homogeneous groups remains? Would reporting agreement success rates, together with payoff dispersion conditional on successful agreement, alter the main conclusion?

6 comments
AuthorsOpenReview ↗

Response to Reviewer ffi4 [Part 1 of 5] Role of Model Capability [W1, part 1]

Dear Reviewer ffi4,

Thank you for the feedback! Please kindly see the supplementary material, ICLR 2027 Rebuttal Figures, for Rebuttal Figures 1–19 referenced below.

Response to Critical Weakness 1: The reviewer points to two results in the paper: (1) the emergence of exclusionary coalitions at a critical scale of capability, and (2) the decline of performance when scaling the number of agents in weak teams against a stronger adversary. For these two results, the reviewer’s complaint is that they do not sufficiently disentangle the respective roles of (a) model capability, (b) institutional advantage, and (c) task complexity. The reviewer suggests adding more “targeted analyses” capable of distinguishing amongst these alternative explanations.

Thus, we do exactly that: we run a set of 6 new ablation experiments, for each of the results (1) and (2) to distinguish and investigate each of the three alternative explanations proposed (a), (b), and (c).

Exclusionary coalitionsLarger weak-model teams
Model capabilityFirst, we run a set of additional experiments to confirm whether the emergence of exclusionary coalitions is truly a function of the capability of the model, rather than say the individual properties of specific models themselves. As such, we run 75 new runs on the homogenous game setting across 3 new models GPT-4o, Claude Haiku 4.5, Claude Opus 4.6 Thinking, which are close in model capability to models we had previously run, to test if the effect of the emergence of exclusionary coalitions is truly a function of model capability (i.e. above a certain threshold of elo), or does testing on different set of models change the results. We find that the existence (and frequency) of exclusionary coalitions is indeed reliably a function of the model capability. As shown in Rebuttal Figure 1, the frequency of exclusionary coalitions among the new models closely matches that of previously evaluated models with similar Elo ratings: GPT-4o (Elo 1345) and Claude Haiku 4.5 (Elo 1407) also have 0% exclusionary coalitions, which is the same as GPT-5 Nano High (Elo 1337) and Qwen-3 Max Preview (Elo 1435), respectively. Claude Opus 4.6 Thinking (Elo 1504) does have exclusionary coalitions just like Gemini 3.1 Pro (Elo 1494), and both have similar frequencies: 44% vs 40%. All in all, Rebuttal Figure 1 adds to the evidence that the existence of exclusionary coalitions is likely due to / is a function of model capability, at some threshold in the 1400s.To test whether our results of the weak-agent team are robust, or simply due some artifact of model capability, we run an additional ablation experiment with 48 new runs where we still have the team set-up, but instead of the adversary having higher model capability than the Nano team, we configure the adversary to have the same capability (i.e. the adversary itself is also a Nano). We run this both for small teams (3-agent team) vs larger teams (7-agent team) to see if our results regarding larger teams doing worse replicate. In Rebuttal Figure 4, we see that indeed having team coordination is still beneficial compared to no coordination, and also similarly as our previous results with the GPT-5.4 adversary, here as we scale from the smaller team to the larger team, the advantage decreases by around half from +49.3 to +25.3. Thus, our two central results of the agent teams experiment are robust in this ablation experiment even when we remove the factor of the gap in model capability: (1) coordination among agents improves performance against an adversary, and (2) scaling up teams of weaker agents can diminish this advantage
AuthorsOpenReview ↗

Response to Reviewer ffi4 [Part 2 of 5] Role of Institutional Advantages [W1, part 2]

Exclusionary coalitionsLarger weak-model teams
Institutional advantagesOne “institutional advantage” of the game that may facilitate the formation of exclusionary coalitions is its supermajority voting rule, which allows a subset of agents to pass proposals without unanimous agreement. To test whether coalition formation depends on this institutional advantage, we conduct additional ablations varying the voting threshold: strict majority, supermajority, and unanimity. In Rebuttal Figure 2, we see that the effect of emergent exclusionary coalitions occurs for all 3 ablation settings, and interestingly, the critical capability threshold needed for exclusionary coalition to emerge increases/decreases based on the difficulty to obtain a critical coalition (majority → supermajority → unanimous). For strict majority, it is easier to get an exclusionary coalition, so we see this starting to occur even at DeepSeek R1 0528, which is a lower elo model than the Gemini 3.1 Pro, which is where exclusionary coalitions occur for the traditional supermajority setting. For the unanimous setting, we see something interesting. Even though the unanimous setting seems to eliminate the possibility of “exclusionary coalition” since coalitions need unanimous vote and thus does not have anyone to exclude, we actually do see an occurrence of it, from the very most capable model, Gemini 3.1 Pro. Specifically, Agents 2 and 3, with the support of Agents 4–10, coordinate to allocate Agent 1 a payoff of only 8, and benefit themselves. Despite requiring Agent 1's approval, the coalition finds a clever way to exclude it by ruthlessly threatening to veto all alternative proposals unless Agent 1 votes to approve. This more sophisticated form of exclusionary coordination emerges only in the most capable model and at a lower frequency, suggesting that sufficiently capable agents can still perform maneuvers akin to the exclusionary coalition even when the institutional advantages / dynamics shift and change to make it more difficult. Overall, these results suggest that the emergence of novel strategies like exclusionary coalitions is robust to changes in the institutional voting rules.One institutional advantage of the weaker team is the binding captain vote set-up: once the captain chooses a proposal, every team member’s vote is automatically cast to that end, thus automatically constituting the supermajority when N >= 4. To test whether our results hold without this rule, we ran 48 new runs where instead of a binding vote by the captain, each agent votes on their own accord. In Rebuttal Figure 5, we find that our results hold robustly to this ablation, as the trend with the independent votes set-up tracks closely with the binding captain vote set-up: In both cases, the smaller Nano team (3 agents) had a significant advantage over the adversary, but as we scaled to the larger team (9 agents), this advantage decreased, with the percent of optimal runs dropping from 67% to 33% with a binding captain vote, and 83% to 25% with independent votes.
AuthorsOpenReview ↗

Response to Reviewer ffi4 [Part 3 of 5] Role of Task Complexity [W1, part 3]

Exclusionary coalitionsLarger weak-model teams
Task complexityWe are interested in whether the effect of exclusionary coalitions is affected by task complexity. Here, we run an additional 75 experiment runs to test this. Specifically, one predominant contributor to the complexity of the task in these games is the number of items (m). Here, we do an ablation experiment by running the exclusionary coalition experiments again for 3 fixed values of m: 5, 10, 20. In Rebuttal Figure 3, we see that the exclusionary coalition effect seems to occur at an increasing threshold for increasing levels of complexity. For m=5 items, we see that all models starting from DeepSeek R1 0528 (elo 1422) do have exclusionary coalitions. For 10 items, it is all models above GPT-5.4 High (elo 1484). For 20 items, it is so difficult that no model has yet done an exclusionary coalition up to the capability we’ve tested. These results support our hypothesis that to do exclusionary coalitions requires a critical level of intelligence/capability, and that as the setting becomes more complex that threshold increases.To test our hypothesis of whether the decline in weak-team performance is due to the increased complexity of larger games, we ran an ablation experiment of 100 additional runs where we held a predominant factor of complexity, the number of items m, constant. We still scale the number of agents in the team, but the corresponding complexity of the game does not increase nearly as much in this case since the number of items agents are juggling is constant. In this case, as shown by Rebuttal Figure 6, the decline in team performance disappears: teams of 3, 5, 7, and 9 Nano agents each reach the exact team optimum in 90% of runs. To test the effect of item count directly, we also hold the team size fixed at 9 agents and increase the number of items from 10 to 25. The exact-optimum rate falls from 90% to 45%, and average team payoff falls from 99.2% to 91.8% of its attainable maximum. In other words, when we minimize the degree to which complexity increases in larger games, we no longer see this decline as we scale the team size. These results suggest that the effect of increasing team sizes declining in advantage is indeed highly correlated to the growing complexity of the games that the weaker agents have to handle.

As such, we have done a comprehensive job of what Reviewer ffi4 has asked, ablating and disentangling the roles of (a) model capability, (b) institutional advantage, and (c) task complexity via “targeted analyses” / ablations of each of the two requested experiments (1) exclusionary coalitions, and (2) larger weak-model teams.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.