Multi-agent LLM information networks influence collective behavior
Start here
Researchers treat the information network of "who observes whom" among LLM agents as an independent design object, examining the effect of observation structure on collective welfare in repeated group decision-making experiments with ten identical LLM agents, covering 1080 groups and six types of games. Results show that information disclosure creates behavioral coupling, whose value depends on whether individually favorable behavior aligns with the collective goal: under free-riding or congestion externalities it amplifies harmful behavior, and in public goods games the group performs best when information is not shared at all. Under a fixed budget, who observes whom matters mainly when coupling is harmful, and a modular "barbell" structure that lets subgroups solidify harmful behavior is consistently the worst. The authors describe this regularity as a cross-game regularity within a single controlled LLM population, not a model-invariant law, and on this basis argue against adopting full transparency by default for agent collectives.
Generated by AI from reports · Updated 2 days ago
Attention and developments
Attention, hourly
Now 31, -1% in 24 hours·Peak 34 (Nov 10 22:00)
Related research
This event's paper
- Steering Emergent Behavior in Multi-Agent LLM Systems via Information Networks
arXiv · 2026-11-09 · Deep reading
Progress timeline
- arXiv: Jailbreaks, Prompt Injection, Poisoning & DefensesGuiding emergent behavior in multi-agent LLM systems through information networks
The researchers treat the inter-agent information network of "who observes whom" as a design object independent of the interaction network, and study how observation structure affects collective welfare in repeated group decision-making experiments with ten identical LLM agents. The experiments cover 1,080 groups and six types of games; results show that information disclosure creates behavioral coupling, whose value depends on whether individually favorable behavior aligns with the collective goal, amplifying harmful behavior under free-riding or congestion externalities, and in public goods games the group performs best when no information is shared at all. Under a fixed budget, who observes whom mainly matters when coupling is harmful, and a modular "barbell" structure that locks subgroups into harmful behavior is consistently the worst. The authors state that this sign separation is a cross-game regularity within a single controlled LLM population rather than a model-invariant law, and on this basis argue against defaulting to full transparency for agent collectives.
Related discussion
Attention trend
New sources per day
- Nov 10: 1 new sources (1 outlets, 0 accounts)
- Nov 11: 0 new sources (0 outlets, 0 accounts)
- Nov 12: 0 new sources (0 outlets, 0 accounts)
- Nov 13: 0 new sources (0 outlets, 0 accounts)