acceptodds
Under review as a conference paper at ICLR 2027

Emergent Discrimination and Polarization in Multi-Agent Systems

Abstract

We study theoretically how discrimination and polarization emerge in a population of learning agents. Each agent uses an adaptive linear decision layer to map a language-model embedding of an issue to a binary opinion and maintains an evolving estimate of how much it trusts every other agent. Through pairwise interactions, agents update both their decision weights and trust estimates using an online Bayesian rule derived from teacher–student learning over a noisy channel. In-context-learning experiments with large language models qualitatively reproduce the predicted opinion and trust updates. We then assign each agent one of two arbitrary group labels, A or B, and model group bias by allowing a fraction of agents to shift its perceived agreement with each message by a strength , depending on whether the sender and receiver share a label. We use order parameters from statistical mechanics to connect these local dynamics to collective states and transitions. Starting from uniformly random initial conditions, simulations reveal a rich phase diagram in the plane, comprising four collective states: spin-glass frustrated out-group favoritism, nondiscriminatory polarization, discrimination with opinion alignment, and discrimination without opinion alignment. Which state emerges depends on both the strength and prevalence of the bias. These results show how simple local learning dynamics can generate qualitatively distinct forms of polarization and discrimination, underscoring the need for population-level safety analysis of multi-agent systems. They also suggest measurable order parameters for characterizing analogous collective states in human societies.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multi-agent systems, AI safety, opinion dynamics, group bias, polarization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_fgtUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how polarization and discrimination develop in a population of agents that update both their opinions and their trust in others. It extends an existing Bayesian opinion-learning model by assigning agents arbitrary group labels and making some agents favour their own group. Simulations identify four collective states, including discrimination with and without opinions aligning with group membership. Separate experiments with one LLM show qualitative agreement with the local opinion and trust updates. The collective results, however, are demonstrated using mathematical agents rather than interacting LLMs.

Critical strengths

The distinction between discrimination with and without opinion alignment is the most interesting finding. It shows that group-based distrust can persist without a corresponding divide in opinions. This adds a useful insight beyond simply observing polarization, and the longer simulations and perturbation tests provide support that this behaviour is persistent.

The model provides a clear way to study how bias strength and prevalence jointly shape collective behaviour. Tracking opinions, trust, and group membership separately allows the authors to distinguish outcomes that a single polarization measure would miss. This makes the framework useful for studying collective behaviour, even though its applicability to LLM populations still needs validation.

Critical weaknesses

The connection to interacting LLM populations is not sufficiently established. The LLM experiments examine opinion and trust updates separately, using one model, whereas the main findings depend on their coupled evolution through repeated interactions. No experiment demonstrates the collective states in an actual LLM population. This leaves a substantial gap between the evidence and the paper’s broader claims about multi-agent AI safety.

The Bayesian justification needs clarification. Appendix A states that the covariance remains block diagonal, although the likelihood generally introduces dependence between opinion weights and distrust. Maintaining independence therefore appears to require an additional approximation. This matters because the derivation and claimed optimality are central to the paper’s justification of its update rules; the authors need to explain the approximation and establish whether the findings are robust to it.

The evidence for sharp phase transitions is limited. The main phase diagram averages only two realizations per parameter setting, and the population-size and duration checks cover selected points rather than the transition boundaries. These results support the existence of different collective states, but do not establish how reliably their boundaries can be located or whether the transitions remain sharp in larger populations.

Questions for authors
  1. Clarification: Appendix A states that the covariance remains block diagonal. How does this follow, given that the likelihood couples opinion weights and distrust? If cross-covariances are discarded as an approximation, how does this affect the claimed optimality of the update rules?

  2. Evidence supporting the scope of the claims: Beyond the separate single-interaction experiments, what evidence supports using these rules to predict the coupled evolution of opinions and trust in LLM populations? Have you tested repeated interactions where both evolve together? I am asking whether such validation supports the current claims, rather than suggesting a broader future research programme.

5 comments

We thank the reviewer for the helpful comments. The new results address repeated interactions in LLM populations, robustness of the learning dynamics, and the analytical basis for the collective states.

Additional results

  • R1. LLM simulations and consensus. We simulated LLM multi-agent systems. We ran ten simulations, each with N=40N=40 agents, using Claude Haiku 5.5 and GPT-6 Luna. Every ordered pair interacted five times, giving 7,800 updates per simulation. The striking result is that these LLM societies do not polarize, they approach consensus. In Appendix C, we studied other components of the group-bias tensor, in particular, the (a,fa)(a,f_a) slice has a consensus regime for positive aa: this component shifts perceived agreement upward regardless of group. Therefore, our results suggest that LLMs behave like agents with positive aa, providing a theoretical model for LLM agreeableness.

  • R2. Perturbed modulation functions. We perturbed the modulation functions and repeated the simulations. The qualitative pattern of the phase diagram remained for most perturbations, suggesting substantial robustness to perturbations in the modulation functions.

  • R3. Landau–Ginzburg analysis. We derive an analytical approximation to the IIIa–IIIb boundary from the microscopic update equations which is given by

\qquad r(x)=\frac{\phi(x)}{\Phi(x)}.$$ Here $b$ is bias strength, $f_c(b)$ is the critical fraction of biased agents, and $\phi,\Phi$ are the standard normal density and cumulative distribution. Overlaid on the numerical phase diagram, the derived curve separates the two phases with no adjusted parameters. Similar analytical expressions can be derived for the other boundaries using the microscopic update equations. - **R4. Stability and mechanism of IIIb.** From the microscopic update equations, we show under explicit conditions that the probability that a phase IIIb configuration survives all future interactions is bounded from below by: $$1-\sum_r\frac{D_r^2+\operatorname{tr}S_r}{d_r^2}.$$ Here $r$ labels agents, and $D_r$, $S_r$, and $d_r$ are quantities constructed in the proof from the agents’ microscopic states, the discussion agenda, and the bias parameters. Applied to 20 randomly sampled IIIb configurations, it guarantees persistence with probability at least 90%. ## Responses to weaknesses - **W1.** R1 directly addresses the absence of interacting LLM populations. It tests repeated exchanges in two models, with both opinions and trust updated throughout, rather than measuring the two updates separately. The resulting tendency toward consensus connects these experiments to the positive-$a$ regime in Appendix C. We will give these population experiments a central role in the revision and distinguish this observed behavior from the other collective states predicted by the mathematical model and observed in human societies, as one can attest in the recent elections. - **W2.** Independence is an explicit modeling choice: agents represent beliefs as a product of Gaussians. The general entropic derivation applies to this restricted family, within which the updates are optimal under its information criterion. Correlations in the unrestricted Bayesian posterior do not invalidate that result. We will clarify that block diagonality is imposed by the representation. R2 also shows robustness to perturbations of the modulation functions. - **W3.** R3 provides an analytical consistency criterion from the microscopic equations, and R4 gives sufficient conditions for IIIb persistence over all future interactions. These strengthen the evidence beyond the numerical diagram. ## Responses to questions - **Q1.** Block diagonality is imposed when choosing the belief family; it does not follow from the likelihood. The update retains each sector’s means and covariances, leaving out cross-sector covariances. Optimality holds within this representation, as explained in W2. - **Q2.** Yes. R1 tests ten societies of 40 LLM agents, across two models, with 7,800 joint opinion–trust updates per society. Their tendency toward consensus connects to the positive-$a$ regime in Appendix C and provides the requested evidence from repeated, coupled interactions.

I appreciate the authors conducting new experiments. Also, your explainaition of the independence assumption answers my question.

I was wondering what happened to trust as the opinions converged in your experiments. Did the agents continue to trust the members of their own group?

In the LLM experiment both trust and opinion converge. The collective do not polarize into two camps: agents from both groups trust on another, and their opinion converge to consensus.

I updated the manuscript with this changes.

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

The paper studies how discrimination and polarization emerge in an online-learning multi-agent setting with and without bias. From random initial conditions, four states can emerge, depending on the strength and type of bias present: frustration, polarization and discrimination with and without opinion alignment. The paper shows that collective behavior can emerge from local learning dynamics that cannot be seen when only looking at individual agents.

Critical strengths

The paper is interesting and shows how individual bias affects collective behavior like polarization and discrimination, which is relevant and important for understanding multi-agent systems.

Critical weaknesses

Presentation: The paper is quite technical and difficult to understand, partly also because the preliminaries are vague in some points and use concepts not previously defined. This affects how well other researchers can understand the paper.

Setting: The setting seems to be limited to the simulated agents and it is unclear whether the observed dynamics also happen in real-world agent systems. I am also not sure, why LLMs are used in this paper, as they only seem to provide the issue representations, which are the same for all agents.

Questions for authors

Q1 (or rather some suggestions for improving the presentation): In the preliminaries: What are the "labels" (are they the opinions of the agents) and what are the agents "own weights"? Is it necessary to overload the definition of agreement (an agent agrees with an issue, an agent agrees with another agent)?

Q2: The finding polarization happens when there is no bias appears to be intuitive, given that, in the update rule, opinion alignment leads to trust. What new insights can be gained from the proposed model?

Q3: How do your findings differ from those in [1]?

[1] Adams et al., The Role of Mistrust in the Modelling of Opinion Adoption, JASSS, 2021

3 comments

We thank the reviewer for the helpful comments.

Additional result

  • R1. LLM simulations and consensus. We simulated LLM multi-agent systems. We ran ten simulations, each with N=40N=40 agents, using Claude Haiku 5.5 and GPT-6 Luna. Every ordered pair interacted five times, giving 7,800 updates per simulation. The striking result is that these LLM societies do not polarize, they approach consensus. In Appendix C, we studied other components of the group-bias tensor, in particular, the (a,fa)(a,f_a) slice has a consensus regime for positive aa: this component shifts perceived agreement upward regardless of group. Therefore, our results suggest that LLMs behave like agents with positive aa, providing a theoretical model for LLM agreeableness.

Responses to weaknesses

  • W1. We will put more effort into improving the presentation. Can you specify what was vague or which concept we used without previously defining?

  • W2. R1 extends the evidence to repeatedly interacting LLM agents that update both opinions and trust. Their increasing agreement is consistent with an effective positive aa, although we have not fitted this parameter or established all predicted phases in LLM populations.

Responses to questions

  • Q1. Each agent has an opinion σI=sign⁡(wI⋅x)\sigma_I=\operatorname{sign}(\mathbf w_I\cdot\mathbf x) on an issue x\mathbf x. Its “own weights” wI\mathbf w_I are the adaptive coefficients mapping the shared issue representation x\mathbf x to that opinion. The arbitrary group labels are κI=±1\kappa_I=\pm 1 (A and B), which identifies the two groups of agents.

  • Q2. We agree that polarization through opinion–trust feedback is intuitive. In III. Phase diagram boundaries, stability and robustness of our "Revised Manuscript" official comment, we mention some applications of the proposed model. But more important, we find that interacting LLM populations approach consensus, a regime our model describes through positive uniform agreeableness aa.

  • Q3. Adams et al. (2021) already show how mistrust can lead to polarization. Our model also asks how arbitrary group membership affects this process. Agents update both their opinions across several issues and their trust in each other. We find that trust can become divided along group lines even when opinions do not. This distinction between opinion polarization and group-based discrimination is a central result of our paper.

As reported in our "Revised Manuscript" comment, we have now fitted the parameters (a,b)(a,b) from LLM single-agent interactions, and confirmed that they land in the "Consensus" region of the (a,b)(a,b) phase diagram. This results gives a predictive validation to our model, confirming that LLMs are best described by agents with positive aa, which serves as a theoretical model for LLM agreeableness.

We are happy to address any other concern from the reviewer before the public discussion period ends.

I thank the authors for addressing my questions and also providing more analysis on whether the behavior is shown in (unsteered) LLMs. Therefore, I am keeping my positive score.

Reviewer_jrgTOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how polarization and discrimination emerge in multi-agent systems through repeated interactions. The authors adopt a Bayesian opinion–trust learning model, where agents update their opinions and trust in others based on received messages. They first conduct in-context learning experiments to show that the opinion and trust updates of an LLM qualitatively agree with the theoretical model. The authors then introduce a group-conditioned bias term and investigate how the bias strength and the fraction of biased agents affect collective dynamics through numerical simulations. The simulations identify four collective states: frustration, polarization without group discrimination, discrimination with opinion alignment, and discrimination without opinion alignment. The paper uses these findings to highlight potential collective safety concerns in multi-agent systems.

Critical strengths
  1. Interesting perspective on LLM behavior. I like the idea of using opinion dynamics to study LLM agents and their collective behavior. Although the underlying model is based on previous work, I think applying this perspective to LLM systems is interesting and provides a different way to understand their interactions.

  2. The experiments in Section 3.2. The experiments in Section 3.2 show encouraging qualitative agreement between the LLM's single-step opinion and trust updates and the theoretical update functions. I think these results provide an interesting initial connection between opinion dynamics and the behavior of black-box LLMs.

Critical weaknesses
  1. Theoretical novelty is relatively limited. The Bayesian opinion–trust learning dynamics are largely inherited from existing opinion dynamics work, such as Alves and Caticha [1]. The main theoretical addition of this paper over the previous works is adding a group-conditioned bias term to an existing dynamical model, and then obtaining phase diagrams through numerical simulations. However, the paper lacks mathematical analysis of the identified collective states, providing neither sufficient or necessary conditions for their occurrence nor analytical phase boundaries. Also, the stability of IIIb is supported only by long-duration simulations and perturbation tests, without theoretical guarantees. Overall, the results are mainly numerical observations from an extension of an existing model, rather than providing substantial new theoretical understanding of the collective dynamics.

  2. Lack of validation in real LLM multi-agent systems and analysis of practical significance. Section 3.2 validates the qualitative agreement between single-step opinion and trust updates of an individual LLM agent and the theoretical model. I understand that this experiment is intended to evaluate whether the proposed update functions can describe the behavior of a black-box LLM. However, I still doubt whether these dynamics hold over repeated interactions between two or more LLM agents, since errors in local updates may accumulate or be amplified over long-term interactions. In addition, the experimental results in Section 4 are obtained from numerical simulations rather than experiments with real LLM agents, despite the paper's emphasis on potential risks in LLM multi-agent systems. The authors neither validate the proposed group-bias mechanism in real LLMs nor observe these collective states in LLM multi-agent systems. Finally, the paper does not evaluate the effects of these collective states on practical tasks or decision quality. For example, I understand that polarization is considered a potential safety concern, but I am not sure whether polarization among LLM agents is necessarily undesirable. Could it sometimes be beneficial, depending on the task? Although this paper puts these directions in the Discussion and Future Work, I think at least some of these gaps should be addressed in this paper to establish a stronger connection between the results in Section 4 and real LLM systems.

[1] Alves, F., and Caticha, N. (2021). Frustration, Glassy Behavior and Dynamical Annealing in Societies of Neural Networks.

Questions for authors
  1. Theoretical novelty and phase boundaries. Compared with Alves and Caticha [1], what new theoretical insights does the proposed group-conditioned bias mechanism provide beyond the numerical phase diagrams? In particular, can the authors derive sufficient or necessary conditions for the emergence of the four collective states, or provide analytical explanations for the phase boundaries?

  2. Stability of IIIb. Section 4.2 uses long-duration simulations and perturbation tests to support the stability of IIIb. Is it possible to provide theoretical analysis or guarantees for its long-term convergence or stability?

  3. Multi-step LLM dynamics. Section 3.2 validates single-step opinion and trust updates in an individual LLM agent. Have the authors examined whether the proposed update functions can still describe the LLM's behavior over multiple interactions? I wonder whether small errors in individual updates could accumulate and lead to different long-term dynamics.

  4. Validation in real LLM multi-agent systems. Can the authors provide evidence that the group-bias mechanism in Eq. (8) exists in real LLMs? More importantly, can the collective states identified in Section 4, particularly IIIa and IIIb, be observed in actual LLM multi-agent systems? Would their trajectories agree with the predictions of the proposed model?

  5. Practical significance of collective states. How do the identified collective states affect practical task performance or decision quality? For example, is polarization among LLM agents always undesirable, or could it sometimes be beneficial by preventing a consensus on an incorrect answer?

6 comments

We thank the reviewer for the helpful comments. We believe the new results motivated by these comments significantly strengthen the paper’s contribution.

Additional results

  • R1. LLM simulations and consensus. We simulated LLM multi-agent systems. We ran ten simulations, each with N=40N=40 agents, using Claude Haiku 5.5 and GPT-6 Luna. Every ordered pair interacted five times, giving 7,800 updates per simulation. The striking result is that these LLM societies do not polarize, they approach consensus. In Appendix C, we studied other components of the group-bias tensor, in particular, the (a,fa)(a,f_a) slice has a consensus regime for positive aa: this component shifts perceived agreement upward regardless of group. Therefore, our results suggest that LLMs behave like agents with positive aa, providing a theoretical model for LLM agreeableness.

  • R2. Perturbed modulation functions. We perturbed the modulation functions and repeated the simulations. The qualitative pattern of the phase diagram remained for most perturbations, suggesting substantial robustness to perturbations in the modulation functions.

  • R3. Landau–Ginzburg analysis. We derive an analytical approximation to the IIIa–IIIb boundary from the microscopic update equations which is given by

\qquad r(x)=\frac{\phi(x)}{\Phi(x)}.$$ Here $b$ is bias strength, $f_c(b)$ is the critical fraction of biased agents, and $\phi,\Phi$ are the standard normal density and cumulative distribution. Overlaid on the numerical phase diagram, the derived curve separates the two phases with no adjusted parameters. Similar analytical expressions can be derived for the other boundaries using the microscopic update equations. - **R4. Stability and mechanism of IIIb.** From the microscopic update equations, we show under explicit conditions that the probability that a phase IIIb configuration survives all future interactions is bounded from below by: $$1-\sum_r\frac{D_r^2+\operatorname{tr}S_r}{d_r^2}.$$ Here $r$ labels agents, and $D_r$, $S_r$, and $d_r$ are quantities constructed in the proof from the agents’ microscopic states, the discussion agenda, and the bias parameters. Applied to 20 randomly sampled IIIb configurations, it guarantees persistence with probability at least 90%. ## Responses to weaknesses - **W1.** R3 provides an analytical expression for phase boundaries. R4 provides a stability guarantee for IIIb. R1 strengthens the significance of adding a group-conditioned bias term to the existing dynamical model by connecting LLM agreeableness and consensus to positive $a$. In light of R1, we plan to give the LLM experiments a more central role in the revised paper, emphasizing their connection to the consensus region of the $(a,f_a)$ phase diagram. - **W2.** R1 tests repeated interactions in actual LLM populations; R2 tests robustness to update errors. We agree polarization can be useful. The consensus tendency that we found in R1 may have contributed to the [Hugging Face incident](https://openai.com/index/hugging-face-incident-and-the-road-ahead/), where dissent failed to stop misaligned coordination. In the revised manuscript we will add a discussion about that point. ## Responses to questions - **Q1.** R3 derives phase boundaries from microscopic equations, and R4 gives sufficient conditions for IIIb persistence. These go beyond numerical phase diagrams. - **Q2.** R4 gives a stability guarantee over all future interactions, with probability at least 90% for the 20 tested configurations. It applies once IIIb has formed. - **Q3.** R1 tests repeated LLM interactions; R2 tests whether update errors accumulate into different collective states. - **Q4.** R1 finds consensus formation for LLM multi-agent systems, which is consistent with the phase reported Appendix C for positive $a$. - **Q5.** We agree polarization among LLM agents might not be undesirable, and R1 shows that they do not polarize. We have not measured task performance. See W2.

Thank you for the response and additional results. Some of my questions are solved, but the observed consensus behavior in the LLM simulations (R1) deepen my concern to the scope of this paper. Here are my two main concerns.

Gap between the proposed mechanism and real LLM behavior

I think the main theoretical contribution of this paper is to show how group-conditioned bias bb can lead to different collective states, including polarization and discrimination. However, R1 shows that real LLM agents reach consensus, which is consistent with the positive-aa regime of the model. Could the authors clarify how R1 supports the main contribution (group-conditioned bias bb) of the paper, beyond showing that the model also has a consensus regime?

For example, can the model predict the behavior of independently interacting LLM agents using parameters estimated from their individual updates? This experiment is not mandatory. I am just using it as an example to help clarify my concern. Given the limited time, the authors are welcome to address this concern with other experiments or evidence they consider more appropriate.

Insufficient details to verify the new rebuttal results

I appreciate the additional experiments and theoretical results, but several essential details are still missing, which makes it difficult for me to be convinced by the results. Here is a list of details I would like to know (including, but not limited to):

  • R1: Do the LLM agents update their opinions and trust independently, or are their updates explicitly governed by the proposed Bayesian update equations?
  • R2: What types and magnitudes of perturbations are applied to the modulation functions? How is robustness quantified?
  • R3: What are the key assumptions and derivation steps leading to the IIIa–IIIb boundary?
  • R4: I would like to see the complete proof.

These details are provided neither in the rebuttal comments nor in the manuscript. Just a reminder that ICLR allows authors to update their manuscripts during the rebuttal period, so the authors could add these details in a different color to help me better understand and evaluate their experiments and results.

Thank you for your comment. We just uploaded the updated manuscript which will hopefully address most of your concerns.

Gap between the proposed mechanism and real LLM behavior.

The observation that LLM populations approach consensus is interesting in its own right, particularly given the tendency of many human collectives to polarize, as electoral politics illustrates. Our main contribution, however, is the theoretical framework for understanding how different collective states emerge in multi-agent systems.

In the original manuscript, we presented this framework through the (b,fb)(b,f_b) slice of the phase diagram, varying group-bias strength and prevalence. We chose this slice because it contains several interesting states, including discrimination with and without opinion alignment (IIIa and IIIb). The framework itself extends beyond that slice. The emergence of consensus in the LLM experiments is consistent with a positive agreeableness field, a>0a>0, which motivates presenting the revised results through the (a,b)(a,b) slice. This allows us to study how agreeableness and group bias jointly shape collective behavior.

Further investigations, including the experiment you suggest, could strengthen this connection. Another possibility comes from our observation that Haiku approaches consensus faster than Luna. Comparing these trajectories with the model’s convergence rates at different values of aa could, in principle, help estimate an effective agreeableness parameter for each model. We have not performed this analysis, but it illustrates another direction opened by connecting the theoretical framework to observed multi-agent behavior.

Insufficient details to verify the new rebuttal results

We have uploaded an updated manuscript containing the experimental details, derivation, and proof. Below we address each specific question.

  • R1. The LLM generates both updates itself. Each receiver sees its current opinions, the speaker’s earlier statements, and its last trust rating, then returns its revised opinion and trust. Those answers become its state for the next interaction. No Bayesian update equation is imposed. Appendix C gives the prompts, schedule, readouts, controls, and results; Figure 6 shows the convergence trajectories.

  • R2. Appendix E.1 specifies the perturbations and shows the resulting phase diagrams. Robustness was assessed visually: the displayed perturbations preserve the four collective states and their overall organization, although their boundaries shift. The analytical approach in R3 also provides a starting point for studying how perturbations change those boundaries.

  • R3. Appendix E.2 gives the assumptions and full derivation of the conditional IIIa–IIIb boundary, together with its comparison to the numerical diagram.

  • R4. Appendix E.3 contains the complete proof, states the sufficient conditions for persistence, and evaluates the resulting bounds on saved IIIb configurations.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised Manuscript

Main changes:

The reviewers identified as the main limitation of our paper the lack of LLM simulations. So we made them and added them to the revised manuscript:

I) LLM simulations:

  1. Added section "5. Toward Consensus in LLM populations", showing that consensus is the collective state that emerges.
  2. Added Appendix C where we detail the LLM simulations.
  3. We changed the text in some places to be consistent with these new result.

Because the LLM simulations shows that a different region of the phase diagram is relevant to understand their collective behavior we also reframed that part of the paper:

II) New regions of the phase diagram:

  1. Fig. 1 now displays the (a,b)(a,b) slice of the phase diagram which contains four collective phases: frustration, polarization, discrimination and consensus.
  2. Fig. 5 displays the three order parameters relevant for the (a,b)(a,b) phase diagram and Eq. 10 introduce these three order parameters.
  3. Eq. 8 now shows both the aa and bb components of the group-bias field.
  4. Fig. 4 now displays both aa and bb shifts to the modulation function.
  5. We changed the text to be consistent with the new framing, including the title that now reads "Polarization, Discrimination and Consensus in Multi-Agent Systems".

As requested by some of the reviewers, we also provide further theoretical analysis of the different phases, their boundaries and stability:

III) Phase diagram boundaries, stability and robustness

  1. We move to Appendix F the (fb,b)(f_b,b) slice of the phase diagram;
  2. We added Appendix F.1 showing that the (fb,b)(f_b,b) phase-diagram continues to emerge even if we perturb the modulation functions.
  3. We added Appendix F.2 "Landau-Ginzburg" where we derive the boundary between regions IIIa and IIIb
  4. We added Appendix F.3 "Stability of IIIb" where we derive the stability formula for region IIIb.

We believe these last results exemplify the utility of having a theoretical description of multi-agent systems. However, we view them as applications, not as the main contributions of the paper.

As requested by reviewer jrgT, we used the model to predict the behavior of real LLM agents:

IV) Predictive validation of the model on LLM agents

  1. We added Appendix D, where we estimate the agreeableness aa and the group bias bb from the single interaction between two agents.
  2. With these parameters, the model predicts the held-out LLM societies of Appendix C as we show in Appendix D.
  3. Most significantly, the fitted (a,b)(a, b) lie in the consensus region of the phase diagram in every condition, consistent with the consensus formed by the LLM societies.
  4. We summarized this result in Section 5 and in the abstract. In the camera-ready version, we will restructure the paper to give it a more central role.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.