Overview
The study investigates the dynamics of convention formation in populations of open-weight language models (LMs), ranging from 1.1 billion to 32 billion parameters. It specifically examines how the runtime interaction graph, often considered an implementation detail in multi-agent LM systems, influences consensus and clique formation. The research employs a naming-game protocol, utilizing restricted first-token scores over tokenizer-safe labels to measure prompt-conditioned score-state distributions. This approach allows for the construction of state-similarity graphs and the differentiation between sampled-label agreement and latent state-space consensus.
Research Context
Multi-agent language-model systems increasingly rely on local interactions. The runtime interaction graph, which dictates these interactions, is frequently treated as a secondary aspect of system design. This research positions the interaction graph as a central variable in understanding how LMs embedded in multi-agent configurations develop shared conventions or fragment into disparate groups. The naming-game protocol provides a structured environment to observe these phenomena, particularly in open-weight models, which are accessible for such experimentation.
Approach
The researchers used a naming-game protocol to study convention formation. This involved:
- Restricting first-token scores to tokenizer-safe labels.
- Measuring prompt-conditioned score-state distributions.
- Constructing state-similarity graphs based on these distributions.
- Separating sampled-label agreement from latent state-space consensus.
- Homophilous threshold-similarity routing: This mechanism involved partners interacting based on a certain similarity threshold.
- Bridge-seeking routing: This mechanism actively sought to connect disparate parts of the interaction graph.
The study also considered the role of memory, specifically `retained-memory` runs, and examined both homogeneous and mixed model populations. Models tested included Qwen2.5-32B, among others, within a varied three-seed mixed four-model grid setup.
Findings
The research yielded several findings regarding the influence of graph feedback on LM population dynamics:
- Retained partner-label evidence was found to be necessary but not sufficient for achieving consensus. This implies that while memory of past interactions is important, the way interactions are routed critically determines the outcome.
- Homophilous threshold-similarity routing tended to delete cross-basin exposure and amplify fragmentation. In 189 `setting-seed` runs, this routing type produced no final behavioral or state consensus in a three-seed mixed four-model grid.
- Bridge-seeking routing often repaired fragmentation when memory was available. In the same three-seed mixed four-model grid, `state-component` and `label-disagreement` bridges recovered final behavioral consensus in 14 out of 18 `retained-memory` runs.
- Across homogeneous model populations, a retained history generally shifted fragmented dynamics towards consensus.
- Qwen2.5-32B, a specific homogeneous model population, demonstrated stable behavioral and final state consensus in all 18 `retained-history well-mixed settings`. This contrasts with threshold-similarity, which reached neither form of consensus in 189 settings for Qwen2.5-32B.
- The qualitative ordering of these findings was preserved across variations in state thresholds, population size, and vocabulary size, suggesting robustness.
- Early-window graph-energy features provided useful within-grid diagnostics, indicating their potential for monitoring the consensus formation process.
Why This Matters
The findings illustrate the critical role of runtime interaction graphs in determining whether multi-agent language model populations achieve consensus or fragment. Understanding these mechanisms is crucial for designing more effective and coherent multi-agent LM systems. The distinction between necessary conditions (retained evidence) and sufficient conditions (appropriate routing) provides a framework for guiding the development of robust collaborative AI environments.