acceptodds
Withdrawn from ICLR 2027

Critical Phase Transitions in Bias of MAS: From Individual Thresholds to Collective Cascades

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-agent systems (MAS) may amplify individual biases through collective interaction. Although previous studies have documented this phenomenon, quantitative analyses of how system bias responds to varying levels of individual bias, and whether critical points can be calculated in advance, remain limited. Using the bias injection level as a control parameter, we conduct experiments with five models to examine how six network topologies influence collective behavior: ring, sparse random, moderately connected random, fully connected, clustered dual-community, and star. The experiments use different levels of persona prompts as anchors, supplemented by numerical stance injection to refine the parameter grid, and compare publicly expressed and privately reported stances using anonymous private probes. Individual responses to opposing views exhibit threshold-based conformity, with model-dependent switching thresholds ranging from 0.48 to 0.63. In homogeneous groups, the consensus probability jumps from 0 to 1 at an injection level of approximately 0.27. This critical value can be calculated in advance from the dose-response curve and drift rate using the dose-drift crossing law. In mixed groups, the critical fraction of randomly placed biased agents required to trigger a global cascade is approximately 0.44, and varies by approximately 0.3 across spatial layouts. In opposing camps, the collective state shifts between compromise and polarization as topology density varies. These results characterize critical phase transitions in bias from individual thresholds to collective cascades and provide a testable procedure for predicting critical points in advance.

Primary area:
interpretability and explainable AI
Keywords:
multi-agent systems, large language models, bias amplification, critical phase transitions, critical mass

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_ka6BOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how collective stance dynamics in LLM-based multi-agent systems respond to controlled bias injection. Experiments use five models, groups of eight agents, six network topologies, and repeated random seeds. The study examines individual conformity responses, homogeneous-group consensus thresholds, mixed-group cascades, and polarization between opposing camps. Additional private-report prompts assess differences between publicly expressed and privately reported stances. The paper reports a homogeneous-group transition between injection levels 0.25 and 0.30, summarized by an approximate threshold of 0.27. It proposes a dose–drift crossing relation to locate this transition interval. In mixed groups, the fraction needed for a cascade depends on network structure and biased-agent placement. The contribution is a useful empirical characterization, although the predictive interpretation and measurement protocols require clarification.

Critical strengths

1.The controlled response-curve approach adds information beyond documenting bias amplification. Scanning injection levels and examining both continuous stance changes and consensus frequencies helps distinguish gradual drift from sharp changes. Comparing homogeneous groups, mixed groups, and opposing camps also reveals that these settings exhibit different collective dynamics.

2.The findings offer useful insights into the competing effects of network structure. Community separation can delay global cascades while sustaining local polarization, whereas dense interaction can encourage compromise under lower injection levels. The paper also appropriately qualifies “phase transition” as an operational finite-group phenomenon and acknowledges that the proposed early-warning indicators lack demonstrated runtime lead time.

Critical weaknesses

1.The dose–drift crossing relation has limited evidence as an a priori predictive law. Section 4.2 estimates drift from a nearby condition and combines it with the measured dose-response curve to locate the same transition interval. This is a useful approximation, but the manuscript does not establish predictive accuracy on held-out model or protocol conditions. The mean-stance crossing is also only a surrogate for the actual three-quarters consensus rule. Furthermore, fine-grid injection is constructed through numerical stance prompts, and Appendix B.1 reports different drift behavior for textual and numerical-anchor conditions. Their compatibility is important to the prediction claim.

2.Some measurement interpretations and protocol details remain unresolved. A private-report prompt measures a response under different instructions; it does not independently establish an underlying private belief or superficial compliance. There are also concrete inconsistencies: the abstract reports switching thresholds of 0.48–0.63, while Section 4.3 uses approximately 0.72; Sections 3.3 and the Figure 3 caption describe a final-round private probe, while Appendix C.1.5 describes probing after every round. These distinctions affect the interpretation and reproducibility of the collective-dynamics analysis.

Questions for authors
  1. Injection-protocol clarification: Which scan points use persona text and which use numerical stance injection? Can you show that the transition and drift estimates remain consistent when the two injection procedures are analyzed separately?

  2. Predictive-validation clarification: Do you have predictions made using independently estimated dose-response and drift parameters for held-out models or interaction protocols? What prediction uncertainty follows from the parameter estimates and grid resolution?

  3. Reporting clarification: Please reconcile the 0.48–0.63 versus approximately 0.72 switching thresholds and the conflicting descriptions of private-probe timing. What evidence distinguishes the public–private response gap from a direct prompting effect?

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

This paper investigates how individual bias affects collective behavior in small groups of LLM agents. It varies bias injection and interaction structure, examining consensus formation, cascades, and polarization. The authors also propose a dose–drift relation to estimate the consensus threshold and compare publicly expressed stances with anonymous private reports.

Critical strengths
  1. Systematically varying the initial bias provides a useful picture of how collective outcomes depend on individual predispositions.

  2. The individual-level experiments reveal both reinforcement and conformity, offering an interesting characterization of responses to peer opinions.

  3. The comparisons across network structures and the use of private probes broaden the empirical analysis.

Critical weaknesses
  • The theoretical basis for the phase-transition claims is unclear. The paper adopts concepts from statistical physics and refers to threshold and cascade models in Section 2, but does not derive a critical condition from an interaction model or stability analysis. Eq. 2 is a phenomenological approximation based on constant drift and threshold crossing rather than a critical condition, and the mixed-group analysis invokes individual thresholds and local exposure without deriving pcp_c from them. Although the authors adopt an operational definition of a phase transition, experiments with eight agents and a fixed interaction horizon do not distinguish a collective phase transition from a sharp crossover or an effect of the predefined consensus criterion, and the field-driven crossover interpretation of De Nobili (2026) remains a plausible alternative. The empirical characterization is useful, but the phase-transition claims require theoretical and empirical support or should be narrowed to reflect the scope of the analysis.

  • The evidence for the predictive claims is limited. For Eq. 2, the main quantitative prediction, setting the drift to zero gives b0∗=0.30b_0^\ast = 0.30 from the reported dose curve (μ(0.30)=7.5\mu(0.30) = 7.5), at the upper end of the observed interval of 0.25–0.30. Including drift shifts the estimate to 0.27, less than the grid spacing of 0.05, so agreement with the interval does not establish an improvement over the no-drift baseline. The calibration is also of unclear applicability: ν\nu is estimated under text injection at b0=0.3b_0 = 0.3, whereas the fine grid uses numerical injection, for which Appendix B.1 reports no drift. Finally, Eq. 2 predicts that b0∗b_0^\ast decreases with TT, and Section 3.3 lists T∈1,2,4,8T \in \\{1,2,4,8\\}, but critical points are reported only at T=4T = 4; the T=6T = 6 runs serve only to estimate ν\nu, and the protocol-parameter analysis referenced in Section 4 is absent from the appendix. Tests across interaction horizons, where Eq. 2 predicts shifts larger than the grid spacing, would be the most direct way to establish the law's added value.

  • The mixed-group analysis and the experimental scope are not fully specified. The threshold used in Section 4.3 (θ≈0.72\theta \approx 0.72) differs from the values reported in the abstract and Fig. 4 (0.48–0.63) without explanation, and the placement comparisons also change the network topology, so placement and topology effects cannot be separated. It is also unclear which models and topic pairs underlie the reported critical points and how they are aggregated: Section 1 says the model is varied between groups while Fig. 1 reads 'eight agents from random five LLMs,' and Section 3.2 mentions eight topics but states that the same primary/spillover statement pair is used in every condition. Finally, although Section 3.1 states that standard errors over seeds are reported throughout, Figs. 6(a), 7, 9 and 10 show none; this matters for Fig. 7, since the phase labels in Table 1 rest on those values and one topology sits exactly on the P=0.4P = 0.4 boundary.

Minor comments

  • The labels in the figures are too small at normal viewing size.
  • Several symbols and labels are introduced without clear definitions: pcp_c, ER5, E0/E2c, and ``Extremism (EE)'' in Fig. 7. The notation νx\nu_x/ν\nu and λ\lambda/λD\lambda_D should be unified. The symbol PP is used for both persuasion strength (Section 3.3) and the polarization index (Section 4.4, Table 1).
  • Protocol descriptions should be made consistent across the main text, figures, and appendix (e.g., the private probe is 'once after the final round' in Section 3.3 and Fig. 3's caption, but 'after each round' in Fig. 3 itself and Appendix C.1.5).
Questions for authors

Please address the concerns raised in the Weaknesses above, with clarifications or supporting evidence where possible.

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

The paper investigates whether bias in individual LLM agents can lead to predictable critical transitions in collective opinion. By varying initial bias and interaction networks, the authors examine how groups move toward consensus or polarization. They report threshold-like conformity in individual agents, abrupt transitions to consensus in uniformly biased groups, and cascades triggered by a sufficiently large biased minority in mixed groups. They propose a “dose–drift crossing law” to estimate the consensus threshold from initial stances and changes accumulated during discussion. Network structure influences both the spread of bias and the persistence of opposing camps. The authors also observe differences between public stances and private reports, which they interpret as superficial compliance. These findings characterize transitions in small experimental groups rather than establish thermodynamic phase transitions.

Critical strengths
  • The paper investigates how collective outcomes change as individual bias or the fraction of biased agents varies, comparing multiple network structures and three group compositions. This provides a more informative characterization of collective bias than measurements at isolated settings.

  • Synchronous updates prevent agents from responding to others’ newly updated outputs within the same round, avoiding update-order effects within each round. The paper also explicitly defines critical transitions operationally and acknowledges the limitations of its small-group setting.

  • The opposing-camp experiments show substantially lower polarization in fully connected networks than in separated communities. This comparison illustrates how network structure can shape whether disagreement leads to compromise or persistent polarization.

Critical weaknesses
  • The bias scan combines persona prompts with numerical stance instructions, while the appendix reports different drift behavior under these methods. Because the reported transition occurs near a persona anchor, a scan using a consistent injection method is needed to distinguish collective dynamics from prompting effects.

  • The calibrated initial mean stance already reaches the consensus cutoff near the observed transition. Agreement with the reported transition interval therefore does not clearly establish an advantage over a baseline based on initial stances alone. Validation across additional conditions would strengthen the prediction claim.

  • Individual thresholds are reported inconsistently, and the cascade explanation connects neighbor proportions to changes in stance scores without specifying the conversion. It also applies thresholds measured in biased agents facing opposition to initially neutral agents, without validating that transfer.

  • The private probe changes how agents are prompted, so differences could reflect re-querying or prompt framing rather than superficial compliance. A matched public re-query control is needed to distinguish these explanations. Inconsistent descriptions of probe timing further complicate interpretation.

Questions for authors
  1. Could you clarify which injection method was used at each bias level in the main experiments? Does the reported consensus transition persist when the same injection method is used throughout the scan?

  2. Does the transition shift as predicted by the dose–drift crossing law when the number of discussion rounds changes? How does its predictive performance compare with a baseline based only on initial stances?

  3. Could you reconcile the different individual thresholds reported in the paper and explain how thresholds measured in biased agents facing opposition apply to initially neutral agents? How are changes in stance scores translated into changes in the neighbor-exposure threshold?

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

Related papers

Discussion (0)

Sign in to comment.