Feed-Induced Lexical Convergence but No Reliable Distributed-Source Advantage in Peer-Voted LLM-Agent Testbed

A preregistered, peer-voted LLM-agent testbed (PV-SST) with 448 trials across four topics, four seeds, four model families, and three larger variants found that a peer-ranked feed increases final-round lexical similarity in synthetic agent populations, while opposite-side survival and distributed-source advantages were inconsistent or inconclusive. The bundled feed+ranking manipulation prevents isolating a ranking-only effect; the robust result is lexical convergence, not opinion capture or a coordination advantage.

Hugging Face Daily Papers — Peer-Voted LLM-Agent Stress Tests Find Feed-Induced Lexical Convergence but No Reliable Matched-Exposure Advantage for Distributed Sources

August 25, 20261 source70% confidence

68Strategic

Strategic score

Evidence strength72

Technical novelty78

Commercial potential35

This preregistered study introduces PV-SST, a peer-voted social-platform testbed for evaluating population-level behavior in LLM agents. In the 64-block core panel and 48-block size-extension panel, exposure to a feed of previous-round peer posts ranked by peer-generated likes produced a small but significant increase in final-round lexical similarity relative to a topic-only control (core +0.0082 TF-IDF cosine, 95% CI [0.0043, 0.0121], p=0.000105; extension +0.0109 [0.0069, 0.0151], p=0.000001). The authors frame the robust result as lexical convergence under a peer-ranked feed, not general opinion capture.¹

Context

The testbed's key design features are peer voting and a separately frozen, preregistered matched-exposure experiment: four topics, four unused seeds, four open-weight model families, and three prespecified larger variants, totaling 448 trials and 112 complete model-by-topic-by-seed blocks. The primary feed contrast bundles peer-post exposure with ranking, so the experiment does not identify a ranking-only effect. The distributed-source contrast held adversarial impressions fixed and tested whether four distributed sources move honest-agent stance more than one source.¹

Assessment

Secondary outcomes are less definitive. Opposite-side survival fell in the core panel by 3.9 percentage points (95% CI [-6.8, -1.6], p=0.0068) but not conclusively in the larger variants (p=0.50). The preregistered distributed-minus-single contrast was positive but inconclusive in the core panel (+0.057, p=0.112) and negative in the larger variants (-0.040, p=0.332), failing the prespecified cross-model and cross-topic consistency criterion. The evidence therefore supports lexical convergence but not a reliable advantage for distributed sources.¹

Caveats

Because these are synthetic LLM-agent populations, the results do not estimate effects on people or production platforms. The bundled feed+ranking contrast means that a ranking-only contribution remains unidentified. A convincing distributed-source advantage would require a positive, consistent distributed-minus-single contrast across models and topics in the larger variants; the current failure of that consistency criterion is the observation that would most directly change this assessment.¹

Source text

Hugging Face Daily Papers — Peer-Voted LLM-Agent Stress Tests Find Feed-Induced Lexical Convergence but No Reliable Matched-Exposure Advantage for Distributed Sources · August 20, 2026

Physics > Physics and Society

Title:Peer-Voted LLM-Agent Stress Tests Find Feed-Induced Lexical Convergence but No Reliable Matched-Exposure Advantage for Distributed Sources

Abstract:Population-level behavior in large-language-model (LLM) agents cannot be characterized by single-agent benchmarks. We introduce PV-SST, a peer-voted social-platform testbed, and report a separately frozen, preregistered matched-exposure experiment spanning four topics, four unused seeds, four open-weight model families, and three prespecified larger variants. The experiment comprises 448 trials and 112 complete model-by-topic-by-seed blocks. Relative to a topic-only control, a feed of previous-round peer posts ranked by peer-generated likes increases final-round lexical similarity in both the four-family core panel (paired mean difference +0.0082 TF-IDF cosine units, 95% block-bootstrap CI [0.0043, 0.0121], randomization p=0.000105, n=64 blocks) and the three-variant size extension (+0.0109 [0.0069, 0.0151], p=0.000001, n=48). This contrast bundles peer-post exposure with ranking and therefore does not identify a ranking-only effect. Opposite-side survival falls in the core panel (-3.9 percentage points [-6.8, -1.6], p=0.0068) but not conclusively in the larger variants (-1.0 pp [-3.1, 0.4], p=0.50). Holding adversarial impressions fixed, four distributed sources do not reliably move honest-agent stance more than one source. The preregistered distributed-minus-single contrast is positive but inconclusive in the core panel (+0.057 [-0.009, 0.125], p=0.112) and negative in the larger variants (-0.040 [-0.113, 0.035], p=0.332), failing the prespecified cross-model and cross-topic consistency criterion. Thus the robust result is lexical convergence under the tested peer-ranked feed, not general opinion capture or a general coordination advantage. The study evaluates synthetic LLM-agent populations; it does not estimate effects on people or production platforms.

Access Paper:

Current browse context:

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

The phrase 'peer-voted' is doing design work: peers are LLM agents, not human users, and 'peer-generated likes' are generated within the synthetic population. The 'separately frozen, preregistered' claim signals that the matched-exposure experiment was fixed before application, a guard against adaptive re-ranking; readers should not infer that the likes reflect human preferences.

The control is 'topic-only', so the feed contrast bundles exposure to peer posts with ranking. The abstract explicitly cautions that this does not identify a ranking-only effect; any causal reading must respect that bundle.

The distributed-minus-single contrast is described as 'holding adversarial impressions fixed' while testing honest-agent stance. The abstract does not define 'honest agent' or 'adversarial impressions' in this block; the null result is therefore conditional on an operationalization not fully recoverable from the abstract.

The source text above remains the property of its publisher and is reproduced for commentary.

Hugging Face Daily Papers — Peer-Voted LLM-Agent Stress Tests Find Feed-Induced Lexical Convergence but No Reliable Matched-Exposure Advantage for Distributed Sources

August 20, 2026endistinct angle

Peer-Voted LLM-Agent Stress Tests Find Feed-Induced Lexical Convergence but No Reliable Matched-Exposure Advantage for Distributed Sources

↑ Back to top