Emergent Coordination in Multi-Agent Language Models¶

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Information-theoretic test for collective structure emergence in multi-agent LLMs

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

This paper introduces an information-theoretic framework—grounded in partial information decomposition (PID) of time-delayed mutual information (TDMI)—to measure whether multi-agent LLM systems exhibit genuine emergent synergy rather than mere aggregation. Using a group binary search task, the authors show that prompt design (personas + theory-of-mind instructions) can systematically steer multi-agent systems from chaotic, uncoordinated aggregates to stable, goal-directed collectives with identity-linked role differentiation.

Problem¶

Multi-agent LLM systems routinely claim "greater-than-sum-of-parts" performance gains from agent differentiation, but no principled, data-driven test existed to determine whether such synergy actually arises—or whether observed coordination is spurious temporal coupling rather than genuine emergent complementarity. Without such a measure, it is impossible to know how to design prompts or roles to reliably induce collective intelligence.

Method¶

The framework applies three complementary information-theoretic tests to multi-agent guesses in a group binary search task (agents propose integers summing to a hidden target, receiving only "too high"/"too low" group feedback):

Practical criterion (\(S_\text{macro}(\ell) = I(V_t; V_{t+\ell}) - \sum_k I(X_{k,t}; V_{t+\ell})\)): tests whether the macro group signal is more self-predictive than the sum of individual agents' predictive power.
Emergence capacity (pairwise PID): decomposes \(I(\{X_{i,t}, X_{j,t}\}; T_{ij,t+\ell})\) into unique, redundant, and synergistic terms (\(\text{Syn}_{ij}\)) for all agent pairs.
Coalition test (\(G_3 = I_3 - \max(I_2)\)): measures whether triplets provide predictive information about the macro signal beyond the best pair.

A fourth pillar uses hierarchical mixed models to test agent-level differentiation (random intercepts \(m_1\) vs. random slopes \(m_2\) vs. time-only baseline \(m_0\)). Two surrogate null distributions—row-wise shuffle (breaks agent identities) and column-wise time-shift (breaks cross-agent alignment while preserving individual dynamics)—distinguish identity-locked synergy from dynamic alignment. Three prompt interventions are tested: Plain (no persona), Persona (Big Five + Schwartz values), and ToM (personas + "think about what other agents might do"). Models tested: GPT-4.1, Llama-3.1-8B/70B, Gemini 2.0 Flash, Qwen3 235B.

Key Contributions¶

A practical, data-driven framework to detect, localize, and decompose emergent synergy in multi-agent LLM systems using PID-TDMI
Two surrogate null tests (row-shuffle, column-shuffle) that separate identity-locked differentiation from spurious temporal correlation
Empirical demonstration that prompt-level interventions causally shift multi-agent systems across distinct coordination regimes
Evidence that internal coordination structure (variance, stability, adaptability) is measurable and controllable without modifying model weights

Results¶

Emergence exists: Practical criterion is significantly above null in all conditions (Fisher joint test \(p < 10^{-16}\); bias-corrected Wilcoxon: Plain \(p=1.5\times10^{-16}\), Persona \(p=6.6\times10^{-7}\), ToM \(p=0.02\)). Emergence capacity similarly significant across conditions.
ToM uniquely stabilizes collectives: Only the ToM condition shows significant positive coalition mutual information \(I_3\) (\(p=3.5\times10^{-14}\)) and Total Stability (\(p=2.9\times10^{-14}\)); Plain and Persona conditions are indistinguishable from zero (\(p>0.84\)).
Coordination mechanism is pairwise, not triadic: Triadic gain \(G_3 \approx 0\) in Persona and ToM, consistent with mean-field coupling to aggregate feedback rather than higher-order local bonds.
Personas enable agent differentiation: Hierarchical model comparison \(m_0 \to m_1\) is significant under Persona and ToM but not Plain, confirming identity-linked behavioral divergence.
Task scaling: Each additional agent reduces success odds by ~8% (OR=0.92, \(p<10^{-16}\)); each unit temperature increase raises success odds by ~50% (OR=1.50, \(p<10^{-7}\)); group size \(N=10\), \(T=1\) used for main experiments (600 total runs).
Results replicate across GPT-4.1, Llama-3.1-8B/70B, Gemini 2.0 Flash, and Qwen3 235B.

Limitations¶

Task is minimalist (binary search with only "too high"/"too low" feedback); findings may not generalize to richer communication or more complex multi-agent pipelines
Framework detects pairwise (\(k=2\)) synergy only for emergence capacity; higher-order synergy beyond triplets is not directly tested
Small-sample entropy estimation is a known bias source; mitigated with Jeffreys prior and Miller-Madow corrections, but discretization (quantile binning with 2 bins) may coarsen structure
Overall task success rate is not significantly different across Plain, Persona, and ToM conditions—structural emergence does not translate to a measurable performance advantage in this setup
Experiments fix group feedback type; results under richer feedback or direct agent communication are unknown

Relevance to Agentic AI / LLM Agents¶

This paper provides the first formal, empirical test of whether multi-agent LLM systems function as genuine collectives versus independent aggregates—a foundational question for any researcher designing agentic pipelines. The demonstration that persona assignment and theory-of-mind prompting causally shift coordination regimes (from chaotic to stable, goal-directed) gives prompt engineers a principled lever beyond trial-and-error. The PID-TDMI framework is model-agnostic and reusable, offering a diagnostic tool to audit whether a proposed multi-agent architecture actually achieves the synergy it claims. The finding that differentiation without goal alignment (Persona-only) fails to produce stable collectives directly challenges the common design heuristic of simply assigning distinct roles to agents.