acceptodds
Under review as a conference paper at ICLR 2027

Social Networks of LLM Agents

Abstract

Large language model (LLM) agents are increasingly deployed in interacting populations, raising the question of what they come to believe collectively. Whether a population aggregates genuine knowledge or collapses into a false consensus affects how much such systems can be trusted. Classical social-network models assume that the network structure alone determines how beliefs combine. This assumption breaks down for LLM agents, whose belief updates incorporate only part of the information they are exposed to, so these models may mischaracterize information aggregation and cannot distinguish genuine consensus from herding. We introduce SNLA, a framework that models each agent's realized influence on others rather than network connectivity. This influence depends on each agent's position in the network and on how sharply attention focuses. Theoretically, we show in a linear belief model that narrow attention leads to herding, with only a few agents shaping the collective belief as the population grows. Wide attention recovers wisdom-of-crowds behavior only on undirected, degree-regular exposure graphs. Empirically, a controlled testbed validates these predictions, and the herding–wisdom transition is demonstrated on operator-controlled variants of three multi-agent LLM benchmarks. These results identify attention breadth, together with exposure structure, as the key determinant of whether an LLM population aggregates knowledge or herds, offering a principled basis for designing and auditing multi-agent systems that are trustworthy at scale.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Social Network

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_8DaJOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the SNLA framework to address the limitation of existing social network models in accurately characterizing the collective belief aggregation process of LLM agent populations, particularly the issue where final beliefs become overly biased toward a minority of agents.

By introducing two core parameters—attention width and social power—the paper establishes a mapping model from exposure to realized influence. Furthermore, the paper proposes a decentralized equalization protocol to eliminate the monopoly of high-power nodes.

Critical strengths

Rigorous theoretical modeling and analysis. The paper conducts an in-depth analysis through a proxy model to prove the impact of attention width and social power on collective belief aggregation. Furthermore, it establishes error bounds between the non-rational updates of LLMs and the linear proxy model through both theoretical derivation and empirical validation, making significant contributions to the theoretical foundations of multi-agent systems.

Critical weaknesses
  1. Limited diversity of LLM backbones. The majority of experiments in this paper rely exclusively on the Qwen2.5 model family. This limited scope undermines the claim of generalizability; it remains possible that more capable or differently aligned models might naturally mitigate herding effects without requiring the proposed interventions. Additional validation across a broader range of model architectures is necessary to establish the universality of these findings.

  2. On both HiddenBench and Werewolf benchmarks, the study focuses solely on scenarios where the high-centrality hub node holds incorrect information. The authors should demonstrate that increasing attention width β\beta or applying the decentralized equalization protocol can improve (or at least maintain) collective belief accuracy when non-hub nodes err.

  3. The convergence of the decentralized pricing protocol is independent of initialization, whereas in most practical scenarios, the initialization of weights carries significant semantic meaning. For instance, the paper initializes hub nodes with high weights, yet after pricing, these weights revert to average levels. So what is the purpose of this initialization design? Consequently, this limits the practical value of the proposed protocol.

Questions for authors

See weaknesses.

What's more, the arrows in Figure 1 are composed of straight lines and triangles, and the arrows from A2 and A4 to A5 are incorrect, and the connections between A1 and A3 should be bidirectional. This leads me to suspect that the figure was AI-generated; however, this is not disclosed in the AI Use Statement.

Reviewer_K8voOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a formal framework for analysing influence effects in the inter-communication and decision-making capabilities of networks of LLMs, using attention "width" \beta to quantify herding vs. wisdom-of-crowd effects. In addition to establishing the theoretical validity of the proposed framework, the paper reports on experiments over a number of datasets that demonstrate that the attention width does indeed control the ability of the agent networks to accurately complete different problems, based on existing benchmarks.

Critical strengths
  • a theoretically robust framework for analysing social network influence between agents in collaborative tasks, with potential utility for studying network influence effects of different network topologies/hierarchies of agents/agent coordination strategies (learned or otherwise)
  • clear empirical support for the theoretical findings the framework predicts based on experiments over a range of benchmarks
Critical weaknesses
  • somewhat limited experiments, in that they are all based on one model family (qwen-2.5; other than for the control agent with AgentsNet) with the only real variable that is (somewhat) explored being a setup where weaker and stronger models are combined (within the same model family); my point here is that the framework has the potential to support much richer exploration of network topologies/information bottlenecks/model families/learned vs. emergent interaction modalities between agents but wasn't, and I was itching to see it used in anger more (so as much a greedy hunger for more as a weakness)
Questions for authors
  • are the experimental results consistent across different model families (beyond qwen-2.5)?
  • the relevance of the "control agent" placement experiment to the broader narrative about attention width wasn't clear to me, if it is possible to tease this out more?
  • could the authors speculate more on the sorts of bigger-picture research questions that could be addressed with the proposed framework to answer broader questions, e.g. about how agents are trained to perform in swarms vs. what behaviours can actually be observed in their behaviours in different settings or using agent networks to simulate different social effects relating to influence/control?
Reviewer_zbGMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates opinion dynamics in LLM multi-agent systems, and in which cases herding behavior occurs. It is claimed that classical social models solely rely on the interaction graph between agents, and do not measure influence as a separate mechanism. The authors introduce a theoretical framework for modeling influence in multi-agent systems. They show that narrow attention leads to herding, which they claim to be a harmful phenomenon in multi-agent systems. To mitigate this behavior, they introduce a cost mechanism, which acts as a regularizer on the attention. In a controlled experiment, they show the effect of varying attention width in their setting.

Critical strengths

The paper raises an interesting question, when herding behavior occurs in multi-agent systems and how it can negatively influence performance. This is a relevant problem, as multi-agent systems are increasingly be applied to settings where agents have asymmetric or distributed information. Therefore, understanding when herding occurs and how to mitigate it can be useful.

Critical weaknesses

It is claimed that exposure and influence are separate mechanisms, and are not well modeled by classical social science models such as FJ. However, the FJ weight matrix can already be interpreted as influence weights, and in the paper these mechanisms are not sufficiently disentangled: In the theory, it is distinguished between the actual communication network (visible exposure graph) and the influence graph. However, in the experiments, the influence graph is constructed by omitting messages from the agents. It is unclear to me how this connects to the influence/exposure separation discussed in the problem formulation (Section 3). Additionally, it makes the name "exposure graph" somewhat misleading since the agents will not necessarily see (be "exposed" to) all messages from neighbors in the exposure graph.

It is also not made explicitly clear, that the observed behavior is not emerging from just letting agents interact, but from actually deciding the messages they see. This makes the results somewhat expected: If agents only see messages of prominent other agents, they have no choice than to either stick with their own opinion or move closer to that of the prominent agents.

The experiments do not provide evidence for whether herding actually emerges without restricting communication through the influence graph, and whether herding hurts in settings without a deliberately adversarial agent. The proposed approach prevents individual agents to dominate the discussion, and it is only tested on benchmarks, where this makes sense (where we, for example, have distributed information). However, there are cases where it would be beneficial to listen to individual agents, for example, if there is one expert with the correct answer and all other models are wrong. Here, the prominent agent is deliberately given an incorrect answer (see Figure 2), making it expected that narrow attention to this agent will decrease performance.

Clarity of writing can be improved significantly. The paper is very dense and generally not well-written, making it hard to follow. Notation and concepts are introduced without explanation or motivation why they are needed. The introduction uses specific concepts that are only introduced in later sections, such as "bridge theorem", "wisdom-of-the-crowd". This makes the paper unnecessarily complicated.

Questions for authors

Q1: Do the results hold without showing the agents the influence weights in the prompts?

Q2: In the experiments, the central hub receives most attention, and its answer is manually set. How does the pricing method affect accuracy when no answer is fixed beforehand, and in settings without a hub? Does influence naturally concentrate on agents that hold the (in)correct answer?

Q3: Wide attention approximates showing all messages, could you comment on the advantage of the pricing method over simply showing all messages to all agents?

Q4: [1] investigates multi-agent systems in a similar setting, and uses the Friedkin-Johnsen model to represent opinion dynamics. They use the FJ weights for influence, explicitly modelling how strongly agents influence each other. Could you clarify your claim that social models capture exposure, but not influence? The FJ weights seem to model influence explicitly. What limitations of the FJ weights motivate the distinction between exposure and influence in your setting given that exposure itself (in your framework) is a weight and not binary?

[1] Abedini et al., Don't Trust Stubborn Neighbors: A Security Framework for Agentic Networks, https://openreview.net/forum?id=ifqasRRoVw

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.