acceptodds
Withdrawn from ICLR 2027

The Partners You Keep Decide What You See: When Reputation-Driven Partner Choice Reshapes Exposure in LLM Societies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language model (LLM) agents can evaluate peers and select future partners, thereby choosing whose behavior remains observable. This makes behavioral trajectories ambiguous because they can reflect both learning from others and changes in exposure. We formalize this exposure-selection confound in a reputation-based partner market with finite partner budgets and observer-specific trust. Cross-role exposure is the primary structural variable. Matched-random retention isolates selection, and forced exploration intervenes on exposure. In a one-shot model with role-balanced reputation supply and role-aligned compatibility, Top- retention yields no more cross-role exposure than matched-random retention. Lower exposure then preserves a larger stationary behavioral gap. A broadcast cross-role subsidy reverses this ordering once its strength exceeds the compatibility bonus. It reaches the integration oracle when cross-role candidates also have uniformly high reputation. Across 1,749 reference-process runs, activation follows partner scarcity and selection precision rather than population size. The result extends to four groups and a continuous trait. Across ten paired seeds per family, DeepSeek shows a modest but detectable exposure reduction, whereas MiniMax and GPT remain near matched-random retention. DeepSeek also follows its reported trust ranking more closely. Instruction-conditioned agents exhibit strong exposure lock-in in all three families. Perturbation-removal experiments separate persistent individual memory from network-mediated persistence, but do not establish a self-sustaining exposure–behavior loop. These results provide an identification framework and measurable activation conditions for determining when reputation-driven selection changes what agents see.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, multi-agent systems, endogenous networks, partner selection, selective exposure, reputation, social learning, exposure lock-in

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_FdKLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how reputation-driven partner selection among LLM agents can affect the diversity of their interactions and the information they receive. The authors introduce a mathematical framework in which agents select partners based on reputation and compatibility, and analyze how this selection mechanism can reduce cross-group exposure compared with random partner selection. They further establish theoretical results showing that reduced exposure can preserve behavioral differences between groups, and propose a cross-group subsidy to encourage more diverse interactions.

Critical strengths

The paper studies an interesting phenomenon in LLM-agent societies: when agents select interaction partners based on reputation and compatibility, they may increasingly interact with similar agents, leading to reduced diversity of exposure. The authors attempt to establish a mathematical framework to explain this phenomenon and investigate possible interventions to mitigate it.

Critical weaknesses

My main concern is that the theoretical analysis and the LLM-agent experiments seem largely disconnected, making it difficult to assess the actual contribution of the paper.

  1. Limited novelty of the theory and unclear connection to LLM agents: The mathematical framework appears to be fairly standard, and LLM agents seem to be just one possible instantiation rather than an essential part of the analysis. Moreover, the theoretical results are not particularly surprising given the assumptions. For example, the model explicitly assumes that agents prefer compatible partners and that compatibility aligns with group membership. Under these assumptions, it is expected that selecting the top-ranked partners would lead to reduced cross-group exposure. More importantly, it is unclear to what extent the interactions among actual LLM agents satisfy the assumptions of the mathematical model, or whether the theory provides meaningful insights into their behavior beyond what is already built into the model.

  2. Unclear purpose of the empirical experiments: The experiments seem more like demonstrations of the proposed mechanism than convincing empirical evidence supporting the theoretical findings. In particular, the strongest results are obtained when LLM agents are explicitly instructed to follow the prescribed partner-selection rule. This shows that LLM agents can implement such a mechanism, but it does not necessarily establish that the phenomenon emerges naturally from their interactions. In fact, the results without these explicit instructions are much weaker and vary across models. It is therefore unclear what conclusions the authors intend to draw from these experiments, or how they substantiate the theoretical analysis.

  3. Limited validation of the proposed remedies: The paper proposes interventions such as a cross-role subsidy to encourage more diverse interactions. However, their effectiveness is primarily justified within the simplified mathematical framework. Although the authors also test the subsidy with LLM agents, the predicted threshold behavior is not observed in these experiments. Furthermore, it is unclear whether increasing cross-group exposure actually improves the agents' behavior or overall system performance. Consequently, the practical effectiveness of the proposed interventions remains insufficiently established.

Questions for authors

See Weaknesses section

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

This paper investigates how partner selection based on peer evaluations changes the information available to LLM agents. The authors assign agents experimental role treatments and compare different partner-retention rules while holding partner capacity fixed. A simplified one-step model establishes conditions under which partner selection reduces cross-role exposure, and a separate behavioral model relates exposure to stationary differences between groups. The authors investigate these mechanisms through reference-process simulations and LLM experiments, distinguishing changes in network composition from changes in behavior.

Critical strengths
  • The paper investigates an interesting aspect of interactions between LLM agents: how their evaluations of one another influence the formation of relationships and, consequently, the information available to them. Comparing different retention rules while keeping partner capacity fixed provides a useful approach to studying these effects.
  • The connection to social science results is described and adapted. I appreciate the distinction between changes in agents' behavioral responses and changes in which partners they observe. The exposure-selection confound is described clearly and provides a motivation for the study.
Critical weaknesses
  • Overall, this paper, while focusing on an interesting direction of research, is not ready for peer review at a conference such as ICLR. The paper, e.g., Sec. 3.1, is not connected to the text and makes the paper not legible without trying to read beyond and back and forth to understand the story. See comments on ``suggestions for future work''. There is often little intuition for the modeling choices or explanation of what a quantity will subsequently be used for.
Questions for authors

Questions

  • Since the theoretical section reduces the setting to one time step, how do Theorems 1 & 2 and Corollary 1 relate to the subsequent sections with multiple rounds?
  • How do the findings of this paper for LLMs compare to the cited social science literature? Within the literature, I would assume similar topics have been studied before the existence of LLMs; could you provide a more critical analysis of the differences between humans and artificial agents?

Suggestions for Future Work

In my view, the paper is difficult to grasp as reviewer. Objects and functions are frequently introduced without sufficient explanation of their meaning, value spaces, admissible parameter ranges, or relationship to previously introduced quantities. As previously mentioned, there is often little intuition for the modeling choices or explanation of what a quantity will subsequently be used for. Closely related notation is introduced without clearly distinguishing the underlying concepts. For example when investigating Sections 3.1–3.2 only, we identify the following problems:

  • Role treatments ziz_i: Definition 1 provides a broad description, but what are the concrete possible treatments, and how are they implemented? An example at this point would help readers understand what the mathematical object represents.
  • The p^i(t)\widehat p_i(t): The text states that it aggregates valid responses aiqta_{iqt}, but how is this aggregation performed? What does the estimate measure, and what is its role in the subsequent analysis?
  • Equation (1): What is aj(t)a_j(t), and how does it relate to the previously introduced aiqta_{iqt}? What function does gg denote, what are its domain and range, and what assumptions are imposed on it? Similarly, the interpretation and admissible range of ηi\eta_i should be stated.
  • Equation (3): What exactly does the vector p(t)\mathbf p(t) represent? How does this modeled state relate to the p^i(t)\widehat p_i(t)? The transition from observed responses to the reference process needs an explicit explanation.
  • The construction preceding Equation (4): What function does σ\sigma denote? What is the interpretation and admissible range of β\beta? Why is attractiveness modeled through this particular transformation of pj(t−1)p_j(t-1), and how should the resulting values be interpreted?

I would expect a systematic revision of the mathematical exposition throughout the manuscript.

  • The different methods are not clearly described and therefore make the goal and actual work hard to follow. There is no mathematical/algorithmic procedure shown for what exactly they entail and how they differ. Experimental terminology also needs more context and proper citations as this may not be known to most reader. (Examples: The Watts–Strogatz ring-rewire initialization, Barabási–Albert initialization, what is the parameter ww, Selection-noise probability pp and its “inferred” range, and many other expressions used without being clear what they refer to.)
  • The results section reports many numbers, but it is difficult to understand which findings are central and what conclusions follow. New terms are introduced in that section, that did not appear before and confuse the reader (What is Role-only, Naive, instructed - do these describe the same mechanisms as “Retention” in Table 1? ). Lastly, I would also recommend to use the term cross-role exposure systematically, without abbreviations.
Reviewer_z7S6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper examines whether reputation-based partner selection helps LLMs retain more partners in the same role (cooperative and self-protective). It combines a fixed-capacity partner market, theoretical analysis, simulations, and LLM experiments comparing random selection, autonomous selection, role disclosure, and explicit selection instructions. The results show that DeepSeek has a modest reduction in cross-role exposure compared with random selection, and MiniMax and GPT show no clear difference.

Critical strengths

This paper investigates how agents' partner choices change the information available to them during repeated interactions. It highlights the need to measure exposure with a matched-random partner-selection control before attributing behavioral divergence to social influence. This is a useful methodological point for evaluating multi-agent systems.

Critical weaknesses
  1. The framework design undermines the interpretation of the results. The roles are fixed by ID and the framework fills the missing partner slots by trust, favoring lower IDs when scores tie. In the DeepSeek runs, 46 of 48 partner additions are supplied by the framework from candidates tied at a trust score of 0.5. The exposure drop comes from low-role observers (from 58.33% to 37.50%), while high-role observers stay at 51.67%. Therefore, it is unclear whether the result reflects agents' preferences or the framework's design. Besides, in the ten GPT's runs, both roles cooperate throughout and every agent retains its initial partner set. These observations do not establish that GPT's lack of segregation results from lower selection fidelity.

  2. The connection between the theory and experiments is unclear and weak. Most of the content focuses on theoretical mechanisms, but the connection to the LLM results is not well established. The theory assumes a same-role compatibility bonus and a specific behavioral update rule, yet the experiments do not establish that the LLMs follow these assumptions. For example, DeepSeek shows reduced cross-role exposure without clear evidence of an increased behavioral gap. Protocol details and the status of different control experiments are also scattered across the main text and appendices, making it difficult for me to follow the experimental design and distinguish theoretical assumptions, simulation findings, and experimental results.

Questions for authors
  1. Could you repeat the core experiments with randomized role-ID assignments and random tie-breaking?
  2. For each model, what is the partner switching rate per round, and what share of newly added partners comes from agent-listed preferences versus framework-filled slots?
  3. What evidence supports attributing GPT's lack of segregation to lower selection fidelity, when both roles always cooperate and partner sets never change?
  4. Could you evaluate cross-role exposure in another game, such as Stag Hunt, to see whether the findings generalize beyond the Prisoner's Dilemma, since models may already be familiar with it.

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

Related papers

Discussion (0)

Sign in to comment.