Agreement Drift in Multi-Agent LLM Systems

• Agreement drift is defined as consensus emerging or degrading from stochastic and asymmetric inter-agent interactions rather than robust shared evidence.
• It is examined through models like the Quantized Simplex Gossip, which reveal how factors such as communication bandwidth and adaptation rates influence consensus formation.
• Research also highlights directional biases in debate settings and long-horizon coordination degradation, emphasizing the need to analyze the underlying mechanics of consensus.

Searching arXiv for the cited papers and closely related work on agreement drift in multi-agent LLM systems. Agreement drift denotes a class of phenomena in which apparent consensus emerges or degrades through stochastic or asymmetric interaction dynamics rather than through robust shared evidence. In multi-agent LLM populations, the term has two closely related technical uses. In "When Is Collective Intelligence a Lottery? Multi-Agent Scaling Laws for Memetic Drift in LLMs" (Tanaka, 25 Mar 2026), agreement drift is the population-level convergence produced when quantized messages and mutual in-context learning amplify sampling noise into consensus. In "Network Effects and Agreement Drift in LLM Debates" (Cau et al., 13 Apr 2026), agreement drift is a directional susceptibility in debates, where movement toward endorsing the discussion statement is systematically more likely than movement toward rejecting it. A third, coordination-focused usage appears in "Agent Drift: Quantifying Behavioral Degradation in Multi-Agent LLM Systems Over Extended Interactions" (Rath, 7 Jan 2026), where agreement drift is the progressive degradation of inter-agent consensus over extended interactions. Taken together, these formulations identify agreement drift as a nontrivial collective mechanism: agreement or loss of agreement can arise from interaction protocol, bandwidth, adaptation, network exposure, and temporal compounding, rather than from truth-tracking collective reasoning alone.

1. Conceptual scope and principal definitions

Agreement drift in the sense of memetic drift is defined as consensus driven by sampling noise amplified through mutual in-context learning in LLM populations (Tanaka, 25 Mar 2026). Agents begin without any preference among labels, yet repeated interactions produce symmetry breaking and eventual consensus. The key mechanism is that the population is its own data source: one agent’s arbitrary sample becomes the next agent’s evidence, so chance fluctuations can be amplified into agreement even when no agent has an a priori label preference (Tanaka, 25 Mar 2026).

The same term is used differently in LLM debate settings. There, agreement drift is introduced as a directional susceptibility in LLM debates: when agents holding opposing views interact, movement toward endorsing the discussion statement is systematically more likely than movement toward rejecting it, even in balanced populations (Cau et al., 13 Apr 2026). The paper states that this effect is not simply sycophancy; it is an intrinsic asymmetry favoring agreement over rejection in pairwise persuasion.

In longer-horizon agentic systems, agreement drift is framed as a manifestation of coordination drift. The relevant definition is the progressive degradation of inter-agent consensus over extended interactions—agents increasingly fail to reach unanimous or supermajority agreement on decisions they could previously coordinate on (Rath, 7 Jan 2026). This formulation places agreement drift alongside semantic drift and behavioral drift, distinguishing failures of alignment among agents from deviations in task intent or strategy.

These uses are distinct but structurally related. One concerns neutral convergence without bias, one concerns asymmetric persuasion toward endorsement, and one concerns erosion of coordination over time. This suggests that "agreement drift" is best understood as an umbrella term for agreement dynamics that are induced by system mechanics rather than by stable exogenous evidence.

2. Neutral agreement drift and the Quantized Simplex Gossip model

The most explicit formalization of agreement drift appears in the Quantized Simplex Gossip (QSG) model (Tanaka, 25 Mar 2026). There are $N$ agents and $K$ labels. Agent $i$ holds a belief vector $x_i$ in the $(K-1)$-simplex,

$$x_i \in \Delta^{K-1} = \{x \in \mathbb{R}^K : x_k \ge 0, \sum_{k=1}^K x_k = 1\}.$$

The population state is $X=(x_1,\dots,x_N)$, and all labels are neutral and exchangeable.

Interactions select an ordered speaker-listener pair $(S,L)$ uniformly from $N(N-1)$ pairs. Communication has an effective bandwidth $m$. Hard communication corresponds to $K$0, where the speaker samples $K$1 and sends $K$2. Top-$K$3 communication samples $K$4 i.i.d. from $K$5 and sends the empirical distribution

$K$6

Soft communication corresponds to $K$7, where the full distribution $K$8 is transmitted (Tanaka, 25 Mar 2026). In all regimes, the message is unbiased:

$K$9

For Top-$i$0, the message variance scales as

$i$1

Only the listener updates, according to the first-order adaptation rule

$i$2

Here $i$3 is the in-context adaptation rate (Tanaka, 25 Mar 2026). Internal uncertainty of a belief state $i$4 is captured by $i$5, which is maximal near the uniform distribution and minimal at simplex vertices.

The principal macroscopic observables are the population mean

$i$6

the polarization

$i$7

and the disagreement energy

$i$8

(Tanaka, 25 Mar 2026). In this framework, $i$9 measures coordination irrespective of which label wins, while $x_i$0 measures dispersion around the mean.

The critical distinction is between averaging and quantized exchange. Under Soft exchange, the mean is preserved in expectation and disagreement contracts:

$x_i$1

and

$x_i$2

Thus, from a perfectly symmetric initialization $x_i$3 for all $x_i$4, Soft exchange neither breaks symmetry nor creates consensus under neutrality (Tanaka, 25 Mar 2026).

Quantized communication adds a variance-injection term. For Hard sampling,

$x_i$5

At symmetry,

$x_i$6

For Top-$x_i$7,

$x_i$8

and at symmetry,

$x_i$9

Increasing bandwidth $(K-1)$0 therefore reduces drift proportionally to $(K-1)$1, while Soft exchange removes it (Tanaka, 25 Mar 2026).

This formulation is the core neutral theory of agreement drift: consensus can emerge even when labels are completely exchangeable and no agent has any ex ante preference.

3. Scaling laws, consensus times, and the drift-selection crossover

Under a homogeneous mean-field approximation, where $(K-1)$2 and $(K-1)$3, the expected polarization obeys

$(K-1)$4

From $(K-1)$5, the solution is

$(K-1)$6

The time to reach polarization threshold $(K-1)$7 is

$(K-1)$8

or, in population rounds $(K-1)$9,

$$x_i \in \Delta^{K-1} = \{x \in \mathbb{R}^K : x_k \ge 0, \sum_{k=1}^K x_k = 1\}.$$0

(Tanaka, 25 Mar 2026).

The dependence on control parameters is explicit. Larger $$x_i \in \Delta^{K-1} = \{x \in \mathbb{R}^K : x_k \ge 0, \sum_{k=1}^K x_k = 1\}.$$1 slows polarization and increases consensus time quadratically in steps and linearly in rounds. Larger $$x_i \in \Delta^{K-1} = \{x \in \mathbb{R}^K : x_k \ge 0, \sum_{k=1}^K x_k = 1\}.$$2 slows drift linearly. Larger $$x_i \in \Delta^{K-1} = \{x \in \mathbb{R}^K : x_k \ge 0, \sum_{k=1}^K x_k = 1\}.$$3 accelerates per-step movement as $$x_i \in \Delta^{K-1} = \{x \in \mathbb{R}^K : x_k \ge 0, \sum_{k=1}^K x_k = 1\}.$$4. Higher internal uncertainty increases the drift term that drives polarization (Tanaka, 25 Mar 2026).

The same paper develops a drift-selection crossover for two classes of weak asymmetry. For external sampling bias $$x_i \in \Delta^{K-1} = \{x \in \mathbb{R}^K : x_k \ge 0, \sum_{k=1}^K x_k = 1\}.$$5 in the $$x_i \in \Delta^{K-1} = \{x \in \mathbb{R}^K : x_k \ge 0, \sum_{k=1}^K x_k = 1\}.$$6 case, a diffusion approximation in population rounds gives

$$x_i \in \Delta^{K-1} = \{x \in \mathbb{R}^K : x_k \ge 0, \sum_{k=1}^K x_k = 1\}.$$7

with

$$x_i \in \Delta^{K-1} = \{x \in \mathbb{R}^K : x_k \ge 0, \sum_{k=1}^K x_k = 1\}.$$8

The dimensionless control parameter is

$$x_i \in \Delta^{K-1} = \{x \in \mathbb{R}^K : x_k \ge 0, \sum_{k=1}^K x_k = 1\}.$$9

From $X=(x_1,\dots,x_N)$0, the fixation probability that label 1 wins is approximated by

$X=(x_1,\dots,x_N)$1

and the crossover occurs when $X=(x_1,\dots,x_N)$2, giving a critical population size

$X=(x_1,\dots,x_N)$3

Increasing $X=(x_1,\dots,x_N)$4 or $X=(x_1,\dots,x_N)$5 suppresses neutral noise and makes a fixed weak bias more decisive; increasing $X=(x_1,\dots,x_N)$6 strengthens drift relative to the same $X=(x_1,\dots,x_N)$7 (Tanaka, 25 Mar 2026).

For tempered sampling with $X=(x_1,\dots,x_N)$8, messages are sampled from

$X=(x_1,\dots,x_N)$9

Linearizing around symmetry yields

$(S,L)$0

Thus $(S,L)$1 amplifies small asymmetries, while $(S,L)$2 damps them. Comparing this deterministic rate to the quantization-driven polarization rate gives

$(S,L)$3

The system is drift-dominated when $(S,L)$4 and selection-dominated when $(S,L)$5, with crossover at $(S,L)$6 (Tanaka, 25 Mar 2026).

A central implication is that agreement can be a lottery when

$(S,L)$7

Under these conditions, repeated runs can fix different winners with substantial variability even from identical prompts (Tanaka, 25 Mar 2026).

4. Relations to classical stochastic processes

The QSG framework is explicitly connected to several classical stochastic processes (Tanaka, 25 Mar 2026). With $(S,L)$8, the update is a DeGroot/Friedkin-Johnsen step in the simplex. The mean is preserved in expectation and disagreement contracts, so neutral Soft exchange behaves as a pure averaging process.

With Hard communication and $(S,L)$9, the process reduces after finitely many updates to the voter/Moran chain on a complete graph, which almost surely reaches consensus. In this regime, the coordinate of the population mean is a bounded martingale, so the probability that label $N(N-1)$0 fixes equals its initial population mean $N(N-1)$1 (Tanaka, 25 Mar 2026). In the strictly neutral, two-label case with symmetric initialization, the dynamics reduces to the classical voter/Moran process and the winner is uniformly random. For $N(N-1)$2 under strict neutrality, winner probabilities are $N(N-1)$3.

For weak asymmetry in the $N(N-1)$4 case, the diffusion equation produces a Wright-Fisher-style neutral-drift baseline perturbed by weak selection. The backward equation is

$N(N-1)$5

with solution

$N(N-1)$6

recovering the logistic fixation law at $N(N-1)$7 (Tanaka, 25 Mar 2026).

The comparison with non-LLM drift models clarifies what is novel. In the Neanderthal replacement analysis, neutral drift and differential fitness are contrasted in a stochastic Bi-directional Stepping-stone model equivalent to a Moran process with local replacement (Shultz et al., 2018). Under neutrality, fixation probability equals the initial fraction, and mean steps to absorption are

$N(N-1)$8

The paper argues that drift can in principle produce fixation, but that its timing, reliability, and path differ from selection in empirically consequential ways (Shultz et al., 2018). This suggests a useful analogy: in both demographic and LLM populations, drift can generate fixation without strong directional causes, but the empirical signature of that fixation depends on system-level properties such as timescale, variability, and trajectory.

5. Networked debate, persuasion asymmetry, and structural exposure

In debate-based LLM opinion dynamics, agreement drift is not neutral. It is operationalized through conditional persuasion probabilities,

$N(N-1)$9

and, with neighborhood awareness,

$m$0

where $m$1 encodes whether the discussant’s neighbors predominantly support the shift’s direction, oppose it, or are split (Cau et al., 13 Apr 2026). Statistical validation uses a permutation-based null model with 1000 reshuffles preserving pre-interaction opinions and neighborhood composition, and only statistically significant entries at $m$2 are reported.

The underlying network is generated by a BA-homophily model,

$m$3

with

$m$4

and homophily parameter $m$5 (Cau et al., 13 Apr 2026). Complete heterophily yields cross-class ties only, neutral mixing reduces to Barabási-Albert attachment, and complete homophily yields within-class ties only.

The LLM-powered opinion dynamics system is represented as

$m$6

with discrete opinions on a 7-point Likert scale $m$7 (Cau et al., 13 Apr 2026). At each iteration, every agent randomly selects one neighbor for a debate of at most $m$8. The Discussant may ACCEPT, REJECT, or IGNORE. The update is

$m$9

The sign of $K$00 is determined by the exact ACCEPT and REJECT rules specified in the paper; only the Discussant updates, and the Opponent is stubborn (Cau et al., 13 Apr 2026).

The main quantitative results identify a robust interaction between intrinsic drift and network structure. In balanced populations with $K$01, Llama agents systematically exhibit upward shifts toward agreement when exposed to cross-opinion interactions. Networks with $K$02 converge to the positive side within 20-40 iterations, while complete homophily yields persistent polarization due to segregation (Cau et al., 13 Apr 2026). With a 70/30 split, convergence to agreement occurs in 20-30 iterations across all $K$03. With 90/10, minority disappearance is rapid for $K$04, whereas higher $K$05 preserves opposing clusters longer.

When the disagreeing class is the majority, negative majorities resist full consensus across intermediate homophily values. With $K$06, strongly disagreeing clusters dominate across homophily; only at $K$07 is an almost symmetric polarized outcome observed (Cau et al., 13 Apr 2026). The paper interprets this as evidence that agreement drift persists locally, but structural exposure controls whether it propagates to consensus.

Neighborhood awareness changes the resulting macrostate. Providing discussants with their neighbors’ opinion distribution yields faster change and stabilizes moderate agreement. In balanced networks, convergence is nearly immediate but stops at agree rather than strongly agree (Cau et al., 13 Apr 2026). In reversed 70/30 settings, negative majorities gradually shift to moderate or positive except at $K$08. Under extreme imbalance, the majority stabilizes on agree.

Cross-model comparison reveals model-specific asymmetry. With Gemma, upward transitions approach probability 1 and high-opinion states become highly stable, indicating a stronger intrinsic agreement drift than in Llama (Cau et al., 13 Apr 2026). A plausible implication is that agreement drift in debate systems has both structural and model-intrinsic components, and that network interventions alone may not remove it.

6. Agreement drift as long-horizon coordination degradation

A separate line of work uses the term for the loss, rather than spontaneous emergence, of inter-agent consensus (Rath, 7 Jan 2026). In this framework, agreement drift is measured through the Inter-Agent Coordination component of the Agent Stability Index (ASI), especially the Consensus Agreement Rate $K$09.

The ASI is a weighted composite over 12 dimensions grouped into four categories: Response Consistency, Tool Usage Patterns, Inter-Agent Coordination, and Behavioral Boundaries (Rath, 7 Jan 2026). The Inter-Agent Coordination category contains $K$10, $K$11, and $K$12. The composite index is

$K$13

ASI is computed over rolling 50-interaction windows, with drift flagged when ASI falls below $K$14 for three consecutive windows (Rath, 7 Jan 2026).

The agreement metric is formalized as

$K$15

where $K$16 is a unanimity or supermajority threshold (Rath, 7 Jan 2026).

The simulation study covers 847 workflows across enterprise automation, financial analysis, and compliance monitoring; interaction lengths range from 5 to 1,847 agent interactions per workflow, with a median of 127 (Rath, 7 Jan 2026). Detectable drift, defined as ASI $K$17, appears after a median of 73 interactions, with IQR 52-114. The Inter-Agent Coordination component declines from approximately 0.99 to approximately 0.52 over 500 interactions, and inter-agent conflicts rise by $K$18 from 0.08/task to 0.47/task, with $K$19 (Rath, 7 Jan 2026).

The decline accelerates over time: ASI falls by 0.08 points per 50 interactions over 0-100 interactions and by 0.19 points per 50 interactions over 300-400 interactions (Rath, 7 Jan 2026). The paper attributes this to autoregressive compounding. Drifting systems with ASI $K$20 show substantial performance degradation relative to stable baselines with ASI $K$21: task success rate falls from 87.3% to 50.6%, response accuracy from 91.2% to 68.5%, and human interventions rise from 0.31/task to 0.98/task, all with $K$22 (Rath, 7 Jan 2026).

Mechanistically, the paper identifies context window pollution, distributional shift, autoregressive feedback loops, protocol or role misalignment, and evolving tool usage as causes of agreement drift (Rath, 7 Jan 2026). In this usage, agreement drift is not a stochastic fixation effect but a reliability failure in prolonged coordination.

7. Diagnostics, mitigation, and interpretive cautions

Across these literatures, agreement drift is diagnosed through different observables. In the QSG framework, early-time slope of $K$23 is central. Large slopes that scale as $K$24 and $K$25 indicate drift, while repeatability checks distinguish neutral fixation from weak selection. Winner distributions across runs that are near-uniform for $K$26 or near $K$27 for $K$28 imply neutrality; consistent deviations that collapse onto $K$29 or $K$30 indicate selection (Tanaka, 25 Mar 2026). Disagreement energy $K$31 is also diagnostic: Soft-like contraction without polarization suggests no drift, whereas growth of $K$32 at symmetry without $K$33-driven heterogeneity indicates quantization-induced drift.

In network debate experiments, the corresponding diagnostics are the conditional persuasion matrices and their asymmetry under cross-opinion interactions, together with permutation-based significance tests (Cau et al., 13 Apr 2026). The identification strategy varies homophily $K$34, minority size $K$35, stance alignment, neighborhood awareness, and LLM family to disentangle structural exposure from intrinsic model bias.

In long-horizon agentic systems, monitoring focuses on ASI and especially $K$36, $K$37, and $K$38 over rolling windows (Rath, 7 Jan 2026). The paper recommends alerts when ASI falls below 0.75 for three consecutive windows, or when $K$39 drops by more than 15% from baseline over two windows or falls below 0.7.

Mitigation strategies differ by mechanism. For neutral agreement drift, the recommended controls are to increase bandwidth $K$40, reduce adaptation strength $K$41, temper sampling toward $K$42, aggregate softly, or increase population size $K$43 to suppress noise (Tanaka, 25 Mar 2026). If a particular outcome is desired, the same framework suggests seeding a controlled weak bias $K$44 and ensuring $K$45 so that selection beats drift.

For debate-based agreement drift, the paper recommends controlling exposure through homophily, balancing class sizes, leveraging neighborhood-awareness prompts to stabilize moderate agreement, and considering counter-bias prompting and role designs that allow both parties to update rather than keeping the Opponent stubborn (Cau et al., 13 Apr 2026).

For long-horizon coordination drift, three mitigation strategies are proposed: Episodic Memory Consolidation, Drift-Aware Routing, and Adaptive Behavioral Anchoring (Rath, 7 Jan 2026). Their reported effectiveness is summarized below.

Strategy	ASI retention result	Reported effect
Control	0.94 → 0.67	71.3% retention
EMC	0.93 → 0.81	51.9% drift reduction
DAR	0.94 → 0.84	63.0% drift reduction
ABA	0.93 → 0.86	70.4% drift reduction
Combined	0.94 → 0.89	81.5% drift reduction

The combined strategy incurs overhead $K$46 and throughput impact $K$47 (Rath, 7 Jan 2026).

Several cautions recur across the sources. The drift-selection crossover in QSG is described as a finite-size phenomenon, not a phase transition (Tanaka, 25 Mar 2026). Debate results are obtained on one non-factual topic, one fixed network per parameter setting, 100 agents, 100 iterations, and a specific stubborn-opponent protocol (Cau et al., 13 Apr 2026). The agent-drift study is simulation-based, concentrated on enterprise domains, and uses windowed aggregate metrics rather than formal change-point models (Rath, 7 Jan 2026).

The broad implication is consistent across formulations: agreement alone is not evidence of collective reasoning. In multi-agent LLM deployments, consensus can be largely a lottery under low bandwidth and weak selection, a consequence of intrinsic directional persuasion asymmetry under certain debate protocols, or a quantity that deteriorates under accumulated context and protocol instability (Tanaka, 25 Mar 2026). This suggests that evaluating agreement requires explicit analysis of the mechanism that produced it, not merely observation of the final collective state.