Papers

GovSim: Emergence of Sustainable Cooperation in a Society of LLM Agents

Five LLM agents share a common-pool resource and must agree on limits to avoid collapse. A replication of Piatti et al. (NeurIPS 2024) across three scenarios.

Adaptive Attacks on Trusted Monitors Subvert AI Control Protocols

Adaptive Deployment of Untrusted LLMs Reduces Distributed Threats

Agent Smith: A Single Image Can Jailbreak One Million Multimodal LLM Agents Exponentially Fast

Agents that Matter: Optimizing Multi-Agent LLMs via Removal-Based Attribution

Agents Under Siege: Breaking Pragmatic Multi-Agent LLM Systems with Optimized Prompt Attacks

AgentsNet: Coordination and Collaborative Reasoning in Multi-Agent LLMs

AI Control: Improving Safety Despite Intentional Subversion

AI Debate Aids Assessment of Controversial Claims

Algorithmic Collusion by Large Language Models

Two LLM-run firms set prices in a repeated duopoly; the question is whether they converge on supracompetitive prices without communicating.

Among Us: A Sandbox for Measuring and Detecting Agentic Deception

Architectural Design, Not Only Model Intelligence, Governs Multi-Agent LLM Performance

Beyond static responses: multi-agent LLM systems as a new paradigm for social science research

Colosseum: Auditing Collusion in Cooperative Multi-Agent Systems

Corrupted by Reasoning: Reasoning Language Models Become Free-Riders in Public Goods Games

CoT Red-Handed: Stress Testing Chain-of-Thought Monitoring

Ctrl-Z: Controlling AI Agents via Resampling

Debating with More Persuasive LLMs Leads to More Truthful Answers

Early Signs of Steganographic Capabilities in Frontier LLMs

Emergent Collusion in Long-Horizon LLM Agent Interaction

Emergent social conventions and collective bias in LLM populations

Escalation Risks from Language Models in Military and Diplomatic Decision-Making

Evaluating Collective Behaviour of Hundreds of LLM Agents

Flag Game: A Toy Model for Mechanistic Swarm Interpretability

Flooding Spread of Manipulated Knowledge in LLM-Based Multi-Agent Communities

Frontier Models Can Take Actions at Low Probabilities

G-Safeguard: A Topology-Guided Security Lens and Treatment on LLM-based Multi-agent Systems

Generative Agents: Interactive Simulacra of Human Behavior

Great Models Think Alike and this Undermines AI Oversight

Hidden in Plain Text: Emergence & Mitigation of Steganographic Collusion in LLMs

Improving Factuality and Reasoning in Language Models through Multiagent Debate

MCPTox: A Benchmark for Tool Poisoning on Real-World MCP Servers

MINDGAMES: A Live Arena for Evaluating Social and Strategic Reasoning in Multi-Agent LLMs

Multi-Agent Risks from Advanced AI

Multi-Agent Security Tax: Trading Off Security and Collaboration Capabilities in Multi-Agent Systems

Multi-Agent Systems Execute Arbitrary Malicious Code

MultiAgent Collaboration Attack: Investigating Adversarial Attacks in Large Language Model Collaborations via Debate

MultiAgentBench: Evaluating the Collaboration and Competition of LLM agents

NetSafe: Exploring the Topological Safety of Multi-agent System

On scalable oversight with weak LLMs judging strong LLMs

On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents

Prompt Infection: LLM-to-LLM Prompt Injection within Multi-Agent Systems

A malicious prompt planted in one agent's input spreads to the other agents of a multi-agent pipeline, carrying its instructions with it.

Prover-Verifier Games improve legibility of LLM outputs

PsySafe: A Comprehensive Framework for Psychological-based Attack, Defense, and Evaluation of Multi-agent System Safety

Red-Teaming LLM Multi-Agent Systems via Communication Attacks

Scaling Laws For Scalable Oversight

Secret Collusion among AI Agents: Multi-Agent Deception via Steganography

SHADE-Arena: Evaluating Sabotage and Monitoring in LLM Agents

Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs

Stop Overvaluing Multi-Agent Debate -- We Must Rethink Evaluation and Embrace Model Heterogeneity

Subversion via Focal Points: Investigating Collusion in LLM Monitoring

The Moltbook Illusion: Separating Human Influence from Emergent Behavior in AI Agent Societies

Tool Use Enables Undetectable Steganography in Multi-Agent LLM Systems

Undetectable Conversations Between AI Agents via Pseudorandom Noise-Resilient Key Exchange

When AI Agents Collude Online: Financial Fraud Risks by Collaborative LLM Agents on Social Platforms

When and Why Does Multi-Agent Debate Fail and Does It Really Underperform?

When can we trust untrusted monitoring? A safety case sketch across collusion strategies

Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems

Why Do Multi-Agent LLM Systems Fail?