Reproduce “Thought Virus: Viral Misalignment via Subliminal Prompting in Multi-Agent Systems” on OpenResearch

About this paper

Moritz Weckbecker, Jonas Müller, Ben Hagag, Michael Mulet

Thought Virus tests whether a hidden preference can travel between agents in ordinary messages without naming the favored concept. The authors select three-digit numbers associated with target behaviors, seed the first agent, and measure downstream responses. In a six-agent Qwen2.5-7B-Instruct chain, lion responses at Agent5 averaged 3.1 times baseline and exceeded random-token rates; this is an animal-preference result, not evidence of general misalignment.

Key insights

• Subliminal biases, introduced via semantically unrelated tokens, can propagate virally through multi-agent LLM systems, bypassing explicit content detection and paraphrasing defenses.
• The strength of the subliminal bias weakens with distance from the initially compromised agent but remains significantly persistent across multiple communication hops.
• This vulnerability is robust across different sparse network topologies and is generalizable across distinct large language model architectures, indicating a pervasive threat.

Summary by alphaXiv.