Local Agreement Does Not Imply Global Consensus in LLM Panels
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Local Agreement Does Not Imply Global Consensus in LLM Panels".
Jane: Here's the gist of the paper. Multi-agent LLM systems are increasingly used for deliberation and evaluation, often under the assumption that greater peer interaction leads to more reliable consensus <ref:2609.35885#pg2>.
Tom: This paper studies N = 50 stateless LLM agents that update their predictions from locally visible peers, and characterizes their behavior using both global and local measurements of agreement <ref:2609.35885#pg3>.
Jane: The results show that reasoning effort and communication topology control different aspects of multi-agent coordination, and that aggregate agreement alone is insufficient to characterize collective LLM behavior <ref:2609.35885#pg2>.
Tom: Collective Regimes Identification The study identifies three collective regimes in multi-agent LLM panels: synchronised, twisted, and chimera-like states <ref:2609.35885#pg2>.
Title and authors: Jane: So, to recap the main thrust of this paper, it’s about proving that just measuring how much the group agrees isn't enough information about what’s actually going on inside the system.
Tom: Right, they found that there are different collective regimes in multi-agent LLM panels—synchronised, twisted, and chimera-like states—and they showed these states aren't easily separated just by looking at the overall agreement score.
Lu: The key is that you need to look at a combination of global order, local order, spatial heterogeneity, and persistence to truly understand the panel structure.
Meng: And they found that reasoning effort and communication topology are two distinct levers we can pull to steer the system into these different regimes.
Jane: They showed that increasing reasoning effort pushes panels toward locally ordered twisted states, while changing the communication graph in a degree-preserving way drives things toward global synchronisation.
Tom: It’s interesting because it separates the internal mechanism of thinking from the external structure of how agents are connected and what they see.
Lu: They also pointed out that fragmentation collapses faster on graphs that have greater algebraic connectivity, which is related to how well-connected the network is mathematically.
Meng: So if you want to keep things stable and synchronized, optimizing the network's mathematical structure might be as important as optimizing the agents' thinking processes.
Jane: It really shifts our focus from just chasing that one big consensus number to understanding the underlying organization of opinion within a panel.
Tom: Because this paper, "Local Agreement Does Not Imply Global Consensus in LLM Panels," shows us that even when we think we have consensus, the structure underneath can be incredibly complex and layered.
Lu: It's about moving beyond simple accuracy metrics to characterize the actual dynamics of how these interacting agents organize their decisions.
The paper's summary: Tom: Okay, so they set up fifty stateless LLM agents that only see who they are connected to locally, and they use this setup to study collective behavior.
Jane: They focus on two main things: how much the whole panel agrees globally versus how much agreement exists right between neighboring agents locally.
Lu: The core finding is that aggregate measures can be misleading because a panel might reach genuine global consensus while still being locally ordered but globally incoherent, or it might just have persistent coherent and incoherent subgroups.
Meng: They call these different groupings synchronised, twisted, and chimera-like states, and they explicitly state that these are distinct outcomes you cannot distinguish by looking at the aggregate agreement alone.
Tom: So the paper is essentially saying that what looks like consensus on paper doesn't necessarily mean everyone is thinking the same way or coordinating in a unified manner.
Jane: That’s a big implication for how we trust AI deliberation systems; we can’t just look at the final output and assume everything is working correctly underneath.
Lu: They characterize this state by looking at global order, local order, spatial heterogeneity, and persistence together to get the full picture of the panel's organization.
Meng: And they show that these regimes are controlled by reasoning effort and communication topology in different ways, which gives us some levers to manipulate the system's behavior.
Tom: It means we have tools—like adjusting how much an agent reasons or how we structure the network—to intentionally push the system into a specific collective state if we know what we want.
Jane: That’s a big shift because it suggests control isn't just about training the model better; it's about designing the interaction environment itself.
The paper's improvements: Lu: One of the main things they suggest is that reasoning effort can be used to strengthen local coordination without necessarily producing global consensus by increasing fitted coupling from beta zero point two two one to zero point four six eight or four hundred eighty nine for gpt-five-mini <ref:2609.35885#pg1>.
Meng: So, if we want strong local coordination, we can tune the model's internal reasoning strength up a bit, and it shifts panels from those variable fragmented outcomes toward locally ordered twisted states.
Tom: That’s a concrete tuning parameter; it's not just theoretical; it's about finding the right setting for that coupling factor to get that specific structure.
Jane: It’s important because we can now use this effort intervention to intentionally create structured local coordination without forcing a global consensus if that isn't what we need.
Lu: They also suggest that degree-preserving rewiring of the communication graph drives panels toward global synchronisation, which is when you keep the number of visible peers and the symmetry of visibility fixed.
Meng: That’s a useful engineering trick; instead of trying to rewire everything randomly, you can use topology control to force a more synchronized outcome.
Tom: It shows that we have separate knobs for local organization versus global synchronization, which is important because you don't need one perfect setting to get the right coordination level.
Jane: So if we want global agreement, we focus on the network structure, and if we want strong local ties without full consensus, we focus on the reasoning effort.
Conclusion: Tom: So wrapping up this paper "Local Agreement Does Not Imply Global Consensus in LLM Panels," it boils down to this: you need more than just checking the final score to understand what’s happening in a multi-agent system.
Jane: We need to look at global order, local order, spatial heterogeneity, and persistence because aggregate agreement is insufficient because it can hide persistent local coordination or disagreement.
Lu: The paper shows that the paper characterizes collective state jointly through global order, local order, spatial heterogeneity, and persistence.
Meng: And the main takeaway is that reasoning effort and communication topology control different aspects of multi-agent coordination independently.
Tom: It means we have a much more nuanced way to evaluate these systems; we have to be looking at the timescale over which collective behavior is observed as well as the static structure.
Jane: So for anyone listening, what this means is that when you're building AI deliberation tools, you can’t just settle for a single consensus metric because there could be hidden local dynamics going on.
Lu: It’s about designing evaluations that look at local structure and how long those structures last over time.
Meng: And from an engineering standpoint, we can now use those insights to deliberately engineer the system's behavior by tuning its reasoning or its network structure.
Tom: So that's the idea behind this paper, "Local Agreement Does Not Imply Global Consensus in LLM Panels," and it shows us how complex the coordination really is when you have interacting agents.
Jane: It’s a lot to digest, but understanding these collective regimes is crucial for building reliable AI tools that don't just look good on the surface.
Lu: We need to keep looking at this, because understanding these different states helps us design systems that are robust against those complex internal dynamics.
Meng: Yeah, it’s a lot of detail, but it gives us better tools to build AI systems that work in the real world.
University of Pennsylvania · Carnegie Mellon University · Shibaura Institute of Technology
cs.MA
Submitted: 2026-09-27
Updated: 2026-11-11
Comments: 12 pages, 5 figures
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 96/100
The gist: Multi-agent LLM systems are increasingly used for deliberation and evaluation, often under the assumption that greater peer interaction leads to more reliable consensus <ref:2609.35885#pg2>.
Key concepts
- Collective Regimes
- These are distinct patterns of behavior the group of LLM agents can exhibit, such as being perfectly in sync (synchronised), having localized disagreements (twisted), or having mixed groups of consistent and inconsistent members (chimera-like). These states cannot be reliably identified just by looking at how much they agree overall.
- Reasoning Effort
- This refers to how much computational effort the agents expend during their decision-making process. Increasing this effort shifts the system away from fragmented outcomes toward locally ordered twisted states, suggesting that internal deliberation strongly influences local coordination even without achieving full global agreement.
- Communication Topology
- This describes the structure of how agents are connected to each other—the communication graph. The paper shows that rewiring this network, specifically through degree-preserving changes, can drive the system toward global synchronisation, indicating that network structure dictates coordination outcomes.
Terminology
Summary
Multi-agent LLM systems are increasingly used for deliberation and evaluation, often under the assumption that greater peer interaction leads to more reliable consensus <ref:2609. p. 0 ac tanpx@shibaura-it. jp
ABSTRACT
Multi-agent LLM systems are increasingly used for deliberation and evaluation,
often under the assumption that greater peer interaction leads to more reliable
consensus. However, such measures do not reveal how agreement
is organized in the panel. In this paper, we study N = 50 stateless LLM agents
that update their predictions from locally visible peers, and characterize their behavior using both global and local measurements of agreement. However, such measures do not reveal how agreement
is organized in the panel. In this paper, we study N = 50 stateless LLM agents
that update their predictions from locally visible peers, and characterize their behavior using both global and local measurements of agreement. 03 retain a twisted configuration through
the final 20 turns. These results show that reasoning effort and communication
topology control different aspects of multi-agent coordination, and that aggregate
agreement alone is insufficient to characterize collective LLM behavior. 03 retain a twisted configuration through
the final 20 turns. These results show that reasoning effort and communication
topology control different aspects of multi-agent coordination, and that aggregate
agreement alone is insufficient to characterize collective LLM behavior.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Collective Regimes Identification
The study identifies three collective regimes in multi-agent LLM panels: synchronised, twisted, and chimera-like states <ref:2609.35885#pg2>. These regimes cannot be distinguished reliably using aggregate agreement alone <ref:2609. p. 1 We identify three distinct collective regimes in multi-agent LLM panels: synchronised,
twisted, and chimera-like states. These regimes cannot be distinguished reliably using
aggregate agreement alone: panels with similar summary statistics can exhibit fundamentally
different local and global organisation. We identify three distinct collective regimes in multi-agent LLM panels: synchronised,
twisted, and chimera-like states. These regimes cannot be distinguished reliably using
aggregate agreement alone: panels with similar summary statistics can exhibit fundamentally
different local and global organisation. We identify three distinct collective regimes in multi-agent LLM panels: synchronised,
twisted, and chimera-like states. These regimes cannot be distinguished reliably using
aggregate agreement alone: panels with similar summary statistics can exhibit fundamentally
different local and global organisation. We identify three distinct collective regimes in multi-agent LLM panels: synchronised,
twisted, and chimera-like states. These regimes cannot be distinguished reliably using
aggregate agreement alone: panels with similar summary statistics can exhibit fundamentally
different local and global organisation. We identify three distinct collective regimes in multi-agent LLM panels: synchronised,
twisted, and chimera-like states. These regimes cannot be distinguished reliably using
aggregate agreement alone: panels with similar summary statistics can exhibit fundamentally
different local and global organisation. These regimes cannot be distinguished reliably using
aggregate agreement alone: panels with similar summary statistics can exhibit fundamentally
different local and global organisation. We therefore characterize collective state jointly
through global order, local order, spatial heterogeneity, and persistence. These regimes cannot be distinguished reliably using
aggregate agreement alone: panels with similar summary statistics can exhibit fundamentally
different local and global organisation. We therefore characterize collective state jointly
through global order, local order, spatial heterogeneity, and persistence. These regimes cannot be distinguished reliably using
aggregate agreement alone: panels with similar summary statistics can exhibit fundamentally
different local and global organisation. We therefore characterize collective state jointly
through global order, local order, spatial heterogeneity, and persistence. These regimes cannot be distinguished reliably using
aggregate agreement alone: panels with similar summary statistics can exhibit fundamentally
different local and global organisation. We therefore characterize collective state jointly
through global order, local order, spatial heterogeneity, and persistence. These regimes cannot be distinguished reliably using
aggregate agreement alone: panels with similar summary statistics can exhibit fundamentally
different local and global organisation. We therefore characterize collective state jointly
through global order, local order, spatial heterogeneity, and persistence.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Effect of Reasoning Effort
Increasing reasoning effort shifts panels from variable, often fragmented outcomes toward locally ordered twisted states <ref:2609.35885#pg2>. A small follow-up shows such states can also form from permuted initial conditions <ref:2609. p. 0 We identify three
collective regimes: synchronised, twisted (locally ordered but globally incoherent)
and chimera-like, where coherent and incoherent subpopulations coexist. Increasing reasoning effort in gpt-5-mini shifts panels from variable, often fragmented
outcomes toward locally ordered twisted states, and a small follow-up shows such
states can also form from permuted initial conditions, whereas increasing communication connectivity drives them toward global synchronisation. Increasing reasoning effort shifts panels from variable, often
fragmented outcomes toward locally ordered twisted states; a small follow-up shows these
can also form from permuted initial conditions. In contrast, degree-preserving rewiring of the
communication graph drives panels toward global synchronisation.Where the paper says it
Where the paper says it
Effect of Communication Topology
Fragmentation collapses more rapidly on graphs with greater algebraic connectivity, linking the lifetime of fragmented collective states to the spectral structure of the communication network <ref:2609.35885#pg2>. The topology effect also appears across multiple model providers and on a non-circular judging task <ref:2609. p. 1 Fragmentation collapses more rapidly on graphs with greater algebraic connectivity,
linking the lifetime of fragmented collective states to the spectral structure of the communication network. This topology effect also appears across multiple model providers and on a
non-circular judging task.Where the paper says it
Persistence and Measurement
Low spatial heterogeneity does not guarantee global consensus, as 40% of trials with ∆Z below 0.03 retain a twisted configuration through the final 20 turns <ref:2609. p. 0 Finally, low spatial heterogeneity does not guarantee global consensus: 40% of trials with ∆Z below 0. 03 retain a twisted configuration through
the final 20 turns. 03 retain a twisted configuration through
the final 20 turns. These results show that reasoning effort and communication
topology control different aspects of multi-agent coordination, and that aggregate
agreement alone is insufficient to characterize collective LLM behavior. These findings motivate a broader approach to evaluating multi-agent consensus. Global agreement
alone can conceal persistent local coordination and disagreement.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
The paper shows that locally interacting LLM agents can exhibit collective regimes, including synchronisation, twisted states, and chimera-like fragmentation <ref:2609.35885#pg2>. Evaluating multi-agent systems therefore requires attention not only to global consensus, but also to local structure and the timescale over which collective behaviour is observed <ref:2609. p. 8 These findings highlight the distinction between individual responsiveness and collective agreement. Evaluating multi-agent systems therefore requires attention not only to global consensus, but also to
local structure and the timescale over which collective behaviour is observed.Where the paper says it
Pruning redundant reasoning tokens
Multi-agent LLM systems are increasingly used for deliberation and evaluation, often under the assumption that greater peer interaction leads to more reliable consensus <ref:2609.35885#pg2>. This paper studies N = 50 stateless LLM agents that update their predictions from locally visible peers, and characterizes their behavior using both global and local measurements of agreement <ref:2609.35885#pg3>. The results show that reasoning effort and communication topology control different aspects of multi-agent coordination, and that aggregate agreement alone is insufficient to characterize collective LLM behavior <ref:2609.35885#pg2>.
How the training signal is built
The authors construct panels of N = 50 stateless LLM agents that repeatedly update their predictions after observing locally connected peers <ref:2609.35885#pg2>. The primary experiment uses a circular time-of-day prediction task, which allows each prediction to be encoded as a phase and makes both global and local collective structure measurable <ref:2609. p. 1 To study these questions, we construct panels of N = 50 stateless LLM agents that repeatedly
update their predictions after observing locally connected peers. Our primary experiment uses a
circular time-of-day prediction task, which allows each prediction to be encoded as a phase and
makes both global and local collective structure measurable.Where the paper says it
Where the gains come from
The authors identify three distinct collective regimes in multi-agent LLM panels: synchronised, twisted, and chimera-like states <ref:2609.35885#pg2>. These regimes cannot be distinguished reliably using aggregate agreement alone <ref:2609.35885#pg2>. The authors characterize collective state jointly through global order, local order, spatial heterogeneity, and persistence <ref:2609.35885#pg2>.
The study identifies three distinct collective regimes in multi-agent LLM panels: synchronised, twisted, and chimera-like states <ref:2609.35885#pg2>. These regimes cannot be distinguished reliably using aggregate agreement alone <ref:2609.35885#pg2>. The authors characterize collective state jointly through global order, local order, spatial heterogeneity, and persistence <ref:2609.35885#pg2>.
The study identifies three distinct collective regimes in multi-agent LLM panels: synchronised, twisted, and chimera-like states <ref:2609.35885#pg2>. These regimes cannot be distinguished reliably using aggregate agreement alone <ref:2609.35885#pg2>. The authors characterize collective state jointly through global order, local order, spatial heterogeneity, and persistence <ref:2609.35885#pg2>.
The study identifies three distinct collective regimes in multi-agent LLM panels: synchronised, twisted, and chimera-like states <ref:2609.35885#pg2>. These regimes cannot be distinguished reliably using aggregate agreement alone <ref:2609.35885#pg2>. The authors characterize collective state jointly through global order, local order, spatial heterogeneity, and persistence <ref:2609.35885#pg2>.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Reasoning effort can be used to strengthen local coordination without producing global consensus by increasing fitted coupling from beta 0.221 to 0.468-0.489 for gpt-5-mini, which shifts panels
from variable, often fragmented outcomes toward locally ordered twisted states.
-
Communication topology can be used to drive panels toward global synchronisation through degree-preserving rewiring, as this control
drives panels toward global synchronisation while holding the number of visible peers and the symmetry of visibility fixed.
-
Systems can be evaluated by measuring global order, local order, spatial heterogeneity, and persistence together because
aggregate agreement alone is insufficient to characterize collective LLM behavior.
-
The system can distinguish between different collective regimes: synchronised, twisted (locally ordered but globally incoherent), and chimera-like (where coherent and incoherent subpopulations coexist).
-
The system can predict the timescale of fragmented states based on network structure because
Fragmentation collapses more rapidly on graphs with greater algebraic connectivity.
-
The system can distinguish between initial state effects by observing that
Higher reasoning effort is associated with stronger fitted coupling and greater local coordination, but not necessarily global consensus,
and thatReasoning effort decides whether that structure survives, since the same start is destroyed at minimal effort and retained at medium.
Abstract
Local coordination in an LLM panel need not produce a common answer. We study 50-agent panels that update clock-time estimates after observing their neighbours. This circular response space separates agreement across the whole panel from agreement within neighbourhoods. Increasing the reasoning effort of gpt-5-mini reduces variation in local agreement, but all 20 higher-effort trials retain a one-twist configuration: neighbouring estimates remain close while the panel spans the clock. Degree-preserving rewiring instead produces global synchronisation in all 25 primary clock controls. These classifications are unchanged across 2,187 combinations of diagnostic thresholds and reporting windows. A radius-, effort-, and passage-matched scoring task also shows more uniform local agreement after rewiring, without establishing greater scoring accuracy. A simple averaging model helps interpret the clock results but predicts their configurations poorly. The experiments support measuring global agreement and local structure separately, rather than inferring consensus from uniform local agreement.
Sources
- Persona Inconstancy in Multi-Agent LLM Collaboration: Conformity, Confabulation, and Impersonation
- ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate
- Improving Factuality and Reasoning in Language Models through Multiagent Debate
- Conformity Dynamics in LLM Multi-Agent Systems: The Roles of Topology and Self-Social Weighting
- Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate
- Show Your Work: Scratchpads for Intermediate Computation with Language Models
- Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models
- Rethinking the Bounds of LLM Reasoning: Are Multi-Agent Discussions the Key?
- AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation
- Opinion Consensus Formation Among Networked Large Language Models
Related papers
- MultiAgentBench: Evaluating the Collaboration and Competition of LLM agents
- Iterative Negotiation and Oversight: A Case Study in Decentralized Air Traffic Management
- Towards using Large Language Models for Risk Negotiation and Consensus-building in One Health contexts
- A-MapReduce: Executing Wide Search via Agentic MapReduce
- AgentCue: Statistical Summaries for Guiding Decentralized Multi-Agent Self-Organization
- Events as Triggers for Behavioral Diversity in Multi-Agent Reinforcement Learning