Researchers show LLM agents mimic human network effects with simple randomization trick
Adding a one-sentence random instruction boosts group payoff 3x in LLM agent experiments.
A new study accepted at ASONAM 2026 explores how groups of large language model (LLM) agents collaborate on spatial search problems. Researchers Hao He, Chris Kuhlman, and Xinwei Deng recreated the Mason-Watts experiment, which originally showed that human groups connected via shorter-path networks outperform those in longer-path networks. They deployed 16 LLM agents across all eight Mason-Watts network topologies, asking them to solve a two-dimensional search task that requires balancing exploration (finding new solutions) and exploitation (refining known ones).
The key finding: LLM agents only replicated the human network-efficiency effect when their first-round choices were randomized. Without randomization, the effect disappeared. Adding a single sentence instructing agents to randomize initial choices improved collective payoff by more than three times the estimated payoff difference across the eight topologies—a dramatic gain from a trivial tweak. However, mechanistic Bayesian optimization agents consistently achieved higher payoffs than the LLM agents, and the paper also analyzes exploration-exploitation behavior, copying dynamics, and spatial diversity among agents.
- 16 LLM agents tested on 8 network topologies in a Mason-Watts spatial search task.
- Network-efficiency effect (shorter paths outperform longer) only emerges when first-round choices are randomized.
- A single randomization instruction boosts collective payoff >3x, but Bayesian optimization agents still beat LLMs.
Why It Matters
Simple prompt engineering can dramatically improve multi-LLM collaboration, with direct implications for AI team design and collective intelligence.