Peer-Voted LLM-Agent Stress Tests Find Feed-Induced Lexical Convergence but No Reliable Matched-Exposure Advantage for Distributed Sources
The clear finding is that peer-ranked feeds made LLM agents sound more alike, while stronger opinion-shift claims did not hold up.
The paper introduces PV-SST, a peer-voted social-platform testbed for synthetic LLM-agent populations. Across 448 trials, prior peer posts ranked by peer-generated likes increased final-round lexical similarity versus a topic-only control. The authors say that result combines exposure and ranking, so it does not isolate ranking alone. They did not find reliable evidence that four distributed adversarial sources moved honest-agent stance more than one source. HF Daily Papers' note
The paper introduces PV-SST, a peer-voted social-platform testbed for synthetic LLM-agent populations. Across 448 trials, prior peer posts ranked by peer-generated likes increased final-round lexical similarity versus a topic-only control. The authors say that result combines exposure and ranking, so it does not isolate ranking alone. They did not find reliable evidence that four distributed adversarial sources moved honest-agent stance more than one source. HF Daily Papers' note
score 4