TL;DR
This paper demonstrates that in multi-agent LLM systems, more transparency is not always better. By treating the "information network" (who observes whom) as a design variable, the authors show that while more information helps agents coordinate on shared goals, it can actually decrease collective welfare in social dilemmas by spreading "contagious" harmful behaviors like free-riding.
What Problem Does This Paper Address?
As LLM agents are increasingly deployed in groups to negotiate, allocate resources, and compete, a central governance challenge emerges: how do we ensure that decentralized, self-interested agents produce good outcomes for the system as a whole?
Existing governance levers often focus on changing agent prompts, roles, or the "interaction network" (who transacts with whom). However, interaction networks are often fixed by the task or environment (e.g., a specific supply chain). This paper proposes a different lever: the information network. By controlling what each agent is permitted to observe about others' past actions, an operator can steer emergent behavior without changing the underlying game or the agents' incentives.
Methods
The researchers conducted 1,080 experimental episodes using groups of 10 identical LLM agents (powered by qwen2.5:7b). The agents played six different repeated group-decision games designed to test various strategic tensions:
- Aligned Games: Coordination (Sum-guessing, Stag-hunt) and Discovery (Bandit).
- Misaligned Games: Social dilemmas (Public Goods Game, Threshold PGG) and Anti-coordination (El Farol).
The study manipulated two primary variables:
- Information Amount ($K$): The number of directed observation links added to the group, ranging from zero (total anonymity) to 90 (full transparency).
- Information Architecture: At a fixed budget of 40 links, they compared four topologies: a degree-heterogeneous Hub, a Regular graph, a well-mixed Expander, and a modular Barbell.
The authors modeled the system as a "behavioral coupling" process, where observation acts as a contagion channel.
Datasets
The study does not use a standard pre-existing dataset. Instead, it introduces a controlled experimental framework across six classic game-theoretic environments adapted for LLM agents. The full prompts and game specifications are provided in the paper's supplementary information to serve as a benchmark for future multi-agent LLM research.
Key Insights and Ideas
- The Value of Information is "Alignment-Contingent": The most important finding is that the "sign" of information's value flips depending on the game. In aligned games, more information helps. In misaligned games (like the Public Goods Game), more information actually erodes welfare because agents "imitate each other downward" into free-riding.
- Information as Contagion: Observation creates behavioral coupling. Agents don't just learn; they catch behaviors. If the attractive behavior is harmful to the group, the information network becomes a vector for that harm.
- The "Barbell Penalty": In human evolutionary game theory, modular/clustered structures often protect cooperation. Surprisingly, the authors found the opposite for LLMs. In misaligned games, the modular "barbell" was reliably the worst architecture because it allowed harmful behaviors to entrench themselves in local sub-populations without being corrected by a global signal.
- Less is More: In the standard Public Goods Game, the group achieved its highest welfare with zero peer information.
Key Results
The researchers found a consistent "sign separation" across the games:
- Aligned Games: Welfare generally increased with the information budget. For example, in the Bandit discovery game, more information led to significantly higher discounted welfare ($r=+0.89$). Welfare increases with information in the Bandit game
- Misaligned Games: Welfare decreased as transparency increased. In the Public Goods Game, the correlation was strongly negative ($r=-0.60$). Welfare decreases as information increases in the Public Goods dilemma
- Architecture Matters: In misaligned regimes, the modular barbell architecture cost the group 13–33% of its welfare compared to the best-performing "Hub" architecture. In aligned regimes, the specific wiring of the network was statistically irrelevant.
Limitations and Future Work
- Single Model: The study used only one model family (
qwen2.5:7b). While the authors argue the results represent an empirical regularity, they acknowledge it may not be a "model-invariant law." - Scale: The experiments were limited to groups of $N=10$. Emergent behaviors might shift in much larger populations.
- Static Networks: The study looked at fixed information networks. Future work could explore adaptive networks that open or close observation links based on the agents' current behavior.
- Communication: The agents were prohibited from sending messages to one another. Combining information network design with structured communication (like "multi-agent debate") is a promising next step.
Summary updated