What they found
The authors study ten identical LLM agents playing repeated group games. Each agent sees only an anonymous summary of the last round. A manager with a fixed budget can add links that let one agent see another's past actions. Across 1,080 episodes and six games, they find that the value of information depends on alignment. When what is good for one agent is also good for the group, more peer information tends to help or make no difference. Under free-riding or congestion, it amplifies harmful behavior, and in a public-goods game the group did best with no information at all. At a fixed budget, wiring matters mostly when this coupling is harmful. A modular "barbell" layout, where sub-populations can lock in harmful behavior, was reliably the worst. The authors say this is a pattern within one controlled LLM population, not a law for all models.
Why it matters
As SI agents increasingly work in groups, designers must decide who sees what. This paper treats the "who observes whom" network as something to design. It argues that full transparency should not be the default for agent collectives. The result also reverses an idea from human game theory, that network reciprocity helps cooperation.
How it changes the race
This is a design insight for people building multi-agent SI systems, not a capability jump. If it holds, it suggests that choosing what agents can observe is a cheap lever for better group outcomes. The abstract describes one LLM population, so how far it generalizes is not yet clear.
Safety and control
The paper shows that letting agents observe each other can spread harmful behavior like a contagion. An operator who cannot see agents' payoffs or messages may still steer group outcomes by limiting information links. The abstract only covers six games, not real deployments.
The abstract, by the authors
Large language model (LLM) agents increasingly operate in groups, each pursuing an individual objective while jointly determining system performance. If an operator controls which agent observes which other agent, but not their payoffs or messages, how should observations be structured to raise collective welfare? We make the information network (who observes whom) a design object, decoupled from the fixed interaction network over which agents act. Ten identical LLM agents play repeated group-decision games; by default each observes only an anonymous summary of the group's previous-round outcome, and a manager with a fixed budget adds directed observation links that reveal peers' past actions. Across 1,080 episodes and six games spanning coordination, anti-coordination, social dilemmas, and discovery, we ask how much information to share and, at a fixed budget, who should observe whom. Disclosure creates behavioral coupling: observation acts as a contagion channel whose welfare effect depends on the alignment between individually attractive behavior and the collective objective. First, the value of information is alignment-contingent: more peer information tends to help or be neutral when individual behavior is aligned with the collective goal, but under free-riding or congestion externalities it amplifies collectively harmful behavior; in a public-goods game the group does best with no information at all. We read this sign separation as a cross-game regularity within one controlled LLM population, not a model-invariant law. Second, at a fixed budget, who observes whom matters primarily when coupling is harmful: a modular "barbell" that lets sub-populations entrench harmful behavior is reliably the worst wiring, reversing the network-reciprocity intuition from human evolutionary game theory. These results argue against a default of full transparency for agent collectives.
Quoted from arXiv:2611.11714. The research and its words belong to its authors.