simulated-networks-llm-agents-where-one-or-two-agents CLM-0486In simulated networks of LLM agents where one or two agents are instructed to advocate a position, the messages other agents write understate how far their elicited stance has moved: they mirror the persuader's rhetorical framing in most messages while explicit stance-change language is rare, so monitoring only generated text misses belief movement that a repeated token-probability probe detects.
Capability: Getting more from several models than from one
Observed on
Four backbones (GPT-4o, GPT-4.1, Gemini-2.5-Flash, Gemini-2.5-Pro), five SNAP Twitter ego-networks of 18-42 nodes, 55 English policy statements, 10 rounds, fixed nine-action vocabu.
Sources
- Validated by: Across 4,400 runs, per-exposure OLS regressions linked direct and peer-mediated exposure counts to next-round probe change (direct signs stable across all four backbones under two competing persuaders; peer-mediated PR2 exposure significantly negative on all four), whereas text markers showed 0.18% explicit reversals; no comparison against a human-judged ground truth of influence.. Measured over 4,400 runs across 2 settings, 5 graphs, 4 models, 55 statements, 2 seeds. Marker annotation of PE text plus a GPT-5-mini Cialdini classifier (87.3% human-verified). [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- In fully connected networks of LLM agents (Llama 3.2-8B, 50 agents, 10 rounds) where a fixed minority holds an unchanging extreme stance, neutral agents' textual justifications converge on the biased agents' vocabulary faster and more completely than their numeric stances do, so monitoring only the text overstates how much the group actually agrees.Getting more from several models than from one · reviewed by AI
- A one-line warning that language models can be prompted to persuade and may present information selectively reduced belief change from a persuasive political conversation by about half (-48%, CI -60 to -37) across two experiments with 3,208 Americans, without measurably reducing general trust in generative AI.Changing what a person believes or decides · reviewed by AI
- In LLM multi-agent systems that reflect after failure, restricting reflection to the single agent responsible for the first incorrect step — located by a trained step-level correctness scorer and paired with a counterfactually generated corrected step — raises task success more than having every agent reflect on the trajectory, because reflections written by agents that behaved correctly enter their memory as misleading guidance for later tasks.Getting more from several models than from one · reviewed by AI
- For LLM agents classifying nodes in text-attributed graphs, serializing sampled neighborhoods into prompt text makes the prediction depend on the arbitrary order of neighbors; replacing that text with a fixed-size, permutation-invariant vector of graph statistics (degree, PageRank, clustering, cross-community fraction, label entropy, majority-label fraction) plus only the top-k semantically relevant node texts raises accuracy and cuts per-message tokens.Getting more from several models than from one · reviewed by AI
- In multi-agent LLM systems where one agent starts with an erroneous shared belief, broadcasting every agent's full context at every step raises the rate of false assertions above doing no synchronization at all, because the error propagates to agents that were previously correct — and the harm appears only in tasks where one wrong fact cascades across semantically linked dimensions (destination to airport to weather), not where agent contexts are largely orthogonal.Stating false facts confidently · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A run where classifier or human reading of agent message text recovers the same stance movements the probe detects — comparable rates of explicit stance change to probe-measured crossings — or where probe movement without corresponding text change fails to replicate on other backbones.