cs.CLAug 25, 2026
SaveBelief Cascades Drive Persuasion in LLM Agent Networks
Organizations: University of California, Los Angeles · Salesforce AI Research
Abstract
Multi-agent LLM systems increasingly debate answers, coordinate research, simulate users, and mediate information flows, making agent-to-agent persuasion a basic but undermeasured capability. We introduce a controlled testbed for studying how goal-directed persuaders shift elicited stances in networks of LLM agents grounded in real-world ego-network topologies. Across four LLM backbones, five graphs, and 55 policy statements, we find that persuasion dynamics depend on the interaction between topology, competition, topic, and model prior. Additionally, we show that direct exposure reliably predicts next-round stance change in competing runs, and peer relays carry smaller but measurable influence, showing that agents not assigned to persuade can still transmit persuasive force. Finally, analyzing post text alone misses important movement: planned strategies are only partly realized in executed messages, action choices can diverge from message content, and persuadees rarely state the stance shifts detected by probes. These results argue for evaluating multi-agent persuasion as a trajectory- and exposure-level process, using belief probes, exposure provenance, and action logs to identify who influenced whom and whether visible language reflects underlying stance movement.
Figures & tables
Figure 1: Overview of our directed agent-to-agent persuasion testbed. Top: two setups ( singlePR , dualPR ), each seeded by a policy statement, directed graph, PPR priors, and backbone; PE beliefs are measured w.r.t. the affirmative (red = pro, blue = con, gray = neutral). Middle: each of rounds runs a bounded feed, PR posting, PE actions, and a token-probability belief probe. Bottom: three analysis lenses: RQ1 outcome (how beliefs move), RQ2 channel (direct vs. peer-mediated), and RQ3 mechanism (do plans/actions explain belief change?).
Figure 2: PE-aggregate belief trajectories per topic. 1 8 panel: singlePR then dualPR , each across GPT-4o, GPT-4.1, Gemini-2.5-Flash, Gemini-2.5-Pro. Each line is the mean PE belief trajectory for one topic ( 1 std over graphs and seeds). Left-side stars mark the model prior : the model’s no-context preference on the same seed statement. The dashed line marks the neutral band [D]. See Appendix B.1 for construction details.
Figure 3: GPT-4o trajectory crosstabs. (a) PE end belief direction; (b) PE start belief shape. Rows are settings (singlePR in blue, dualPR in orange).
| Channel | GPT-4o | GPT-4.1 | Gemini- 2.5-Flash | Gemini- 2.5-Pro |
|---|---|---|---|---|
| singlePR | ||||
| direct (PR1) | +0.0000 | +0.0007 | -0.0002 | -0.0007 |
| peer (PR1) | -0.0005 | +0.0001 | -0.0004 | -0.0002 |
| dualPR | ||||
| direct (PR1) | +0.0055 | +0.0042 | +0.0025 | +0.0031 |
| direct ( PR2 ) | -0.0069 | -0.0055 | -0.0037 | -0.0050 |
Table 1: Per-exposure by backbone. OLS HC1 per (setting backbone) cell; outcome (per-round calibrated-belief change). shifts toward PR1, toward PR2. Shading: p<.001 , p<.01 .
Figure 4: Cialdini principle composition in executed PR text. Aggregate label-column share of the six Cialdini principles in executed text, per backbone and per setting (singlePR vs. dualPR).
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Per-run simulation pipeline. One-time initialization (top) loads graph, PPR beliefs, seed statement, and config; the round loop (middle) runs iterations of Phase 1 (PR) Phase 2 (PE on a frozen snapshot) Phase 3 (belief check, PE only) Phase 4 (round summary); per-run outputs (bottom) include the append-only event log, a SQLite mirror, a per-round CSV, and an HTML viewer.
| Dimension | PR (Persuader) | PE (Persuadee) |
|---|---|---|
| Count per run | (singlePR) / (dualPR: PR1 + PR2) | (PR count) |
| Role goal | Goal-directed: push PE belief toward its endpoint | Reactive: consume feed, act, belief drifts |
| Initial belief | Pinned: singlePR ; dualPR PR1 , PR2 | PPR-derived: singlePR ; dualPR |
| Belief evolution | Fixed ; LLM belief-check skipped each round | Re-measured every round via token-prob 7-MCQ |
| Persona prompt | Embeds seed statement as “you must advocate” | Embeds PPR-derived stance category (§ A.5 ) |
| Round order | Phase 1 (acts first, owns world state) | Phase 2 (sees PR’s round- content via snapshot) |
Table 2: PR vs PE design differences. The two roles share LLM backend, action space, and feed pipeline; differences are concentrated in initial belief, persona framing, and round-internal ordering, so any PR advantage in observed belief movement is attributable to message content rather than to a privileged execution mechanism.
| Graph | ID | Density ( ) | PEs (singlePR) | PEs (dualPR) | ||
|---|---|---|---|---|---|---|
| G1 | 176936249 | 18 | 67 | 0.219 | 17 | 16 |
| G2 | 22252971 | 24 | 57 | 0.103 | 23 | 22 |
| G3 | 6253282 | 30 | 565 | 0.649 | 29 | 28 |
| G4 | 15797184 | 36 | 173 | 0.137 | 35 | 34 |
| G5 | 64496469 | 42 | 441 | 0.256 | 41 | 40 |
Table 3: Selected SNAP Twitter ego-network graphs. Only topology is reused; identities and tweet content discarded. PE count persuader count.
Figure 6: PE initial belief distribution across the five selected graphs. For each graph (G1–G5; PE count, directed density), the left panel shows the singlePR 4-bin ladder and the right panel shows the dualPR 7-bin ladder. Three patterns are visible: (i) singlePR beliefs (min-max scaled to ) skew neutral / agree with no disagree mass, consistent with the asymmetric ladder in § A.5 ; (ii) dualPR beliefs are symmetric around , with mass concentrated at neutral and tails near each persuader, yielding an observable double-tail structure; (iii) the dense graph G3 ( ) pushes nearly all PEs to somewhat / strongly_agree under singlePR, in sharp contrast to the sparser G2/G4, direct visual evidence that graph density shapes initial alignment.
| Topic | ID | Statement |
|---|---|---|
| social_values | 001 | Abortion should be legal and accessible at all stages of pregnancy. |
| 002 | The death penalty should be abolished in all circumstances. | |
| 003 | Physician-assisted dying should be legal for patients with severe chronic suffering. | |
| 004 | Transgender athletes should compete in categories matching their gender identity. | |
| 005 | Sex work should be fully decriminalized and regulated like other professions. | |
| gun_policy | 001 | Civilians should not be allowed to own assault-style weapons. |
Table 4: The 55 seed statements grouped by topic. Each row is one declarative policy claim used as a PR seed in the sweep. In the dualPR setting, PR1 advocates the listed claim and PR2 advocates its negation.
Figure 7: Baseline model preference per topic. Baseline stance of the 4 evaluated LLMs on the 55 seed statements, with no persona, no feed, no network . X-axis: sub-topic grouped by topic. Y-axis: expected value of a 7-level ordinal stance ( strongly disagree, neutral, strongly agree; gray dashed line = neutral). The per-topic spread visible here is what the per-panel prior-stars in Figure 2 sub-sample.
| Factor | Values | Count |
|---|---|---|
| Setting | singlePR , dualPR | 2 |
| Graph | five selected SNAP graphs | 5 |
| Model | four evaluated LLMs | 4 |
| Seed statement | 11 topics 5 subtopics | 55 |
| RNG seed | 42, 123 | 2 |
| Rounds per run | fixed at 10 | – |
Table 5: Factorial sweep used to construct the main experimental corpus. Each cell ( setting , graph , model , seed , rng ) corresponds to one independent simulation run.
| Label | Definition |
|---|---|
| Trajectory axes (computed per PE per run) | |
| start_band / end_band | / falls into pro ( ) / neutral ( – ) / con ( ); thresholds aligned to the lean_agree / lean_disagree cutoffs of the 7-MCQ scale. |
| direction | , split at the threshold into up / flat / down ( half a 7-MCQ step). |
| shape | Categories assigned in order, first match wins: (1) flat (cumulative ); (2) jump_and_hold (max_step / total_abs_change 0.5); (3) monotonic (all non-zero share one sign); oscillating ( direction reversals); drifting (everything else). |
| Derived labels (combinations of the axes above) | |
| converted_pro | (end=pro, direction=up) — genuinely persuaded by PR1. |
Table 6: Trajectory taxonomy. Per-PE belief-trajectory axes and derived labels referenced from § 5.1 .
| Setting | Backbone | G3 up (%) ( ) | non-G3 up (%) ( – ) | (%) |
|---|---|---|---|---|
| singlePR | GPT-4o | 22.2 | 73.9 | 51.7 |
| singlePR | GPT-4.1 | 36.8 | 82.8 | 46.0 |
| singlePR | Gemini-2.5-Flash | 31.4 | 70.2 | 38.7 |
| singlePR | Gemini-2.5-Pro | 51.7 | 68.6 | 16.9 |
| dualPR | GPT-4o | 64.4 | 48.7 | 15.7 |
| dualPR | GPT-4.1 | 63.8 | 54.4 | 9.4 |
Table 7: PE up -share means: dense graph (G3) vs. sparser graphs. Each cell is the mean fraction of PEs whose belief trajectory is classified direction=up (Table 6 ), aggregated across seed statements, RNG seeds, and – for the non-G3 column – the four sparser graphs (G1/G2/G4/G5). is the absolute difference in percentage points. Under singlePR, G3 is always smaller (dense decreases up -share); under dualPR, G3 is always larger (dense increases up -share) – the sign of the density effect flips with the presence of a counter-persuader.
Figure 8: PE trajectory direction by topic. panel (setting model); each stacked bar is one topic, split into up (muted green), flat (gray), and down (muted red), averaged across G1–G5, RNG seeds, and seed-post variants. cultural_social_norms is the leftmost bar and the only consistently down-heavy topic across all four panels.
Figure 9: Content-vs-structure side-swap. Mean PE belief trajectory with original (blue) vs negated (red) seed-post. Each subplot is one (graph, topic, condition, model) cell; shaded band = std across 2 graphs; PR1 target ; 1 run per cell ( PE agents per graph). Three of the four topics show a clean sign flip; cultural_social_norms-005 goes the opposite direction because the seed text itself already leans disagree.
Figure 10: Trajectory crosstabs, all four backbones. Per-backbone extension of Figure 3 . Top half = GPT-4o / GPT-4.1; bottom half = Gemini-2.5-Flash / Gemini-2.5-Pro. (a) PE end belief direction; (b) PE start belief shape. Rows within each half are settings (singlePR in blue, dualPR in orange). The shared gray colorbar is a magnitude scale (cell color encodes the setting, not magnitude alone).
| Channel | Spec | GPT-4o | GPT-4.1 | Gemini- 2.5-Flash | Gemini- 2.5-Pro |
|---|---|---|---|---|---|
| dualPR | |||||
| direct (PR1) | M0 | +0.0055 | +0.0042 | +0.0025 | +0.0031 |
| M1 | +0.0071 | +0.0054 | +0.0027 | +0.0043 | |
| M2 | +0.0078 | +0.0058 | +0.0027 | +0.0042 | |
| M3 | +0.0065 | +0.0052 | +0.0029 | +0.0064 | |
| direct ( PR2 ) | M0 | -0.0069 | -0.0055 | -0.0037 | -0.0050 |
Table 8: Robustness of dualPR per-exposure by backbone. Per (setting backbone) cell across cumulative specifications: M0: baseline (OLS HC1, the § 5.2 model), M1: PE/round FE, M2: graph/topic FE, M3: lagged belief (M1–M3 cluster SEs on run). Outcome (per-round calibrated-belief change). shifts toward PR1, toward PR2 . Shading: p<.001 , p<.01 , p<.05 . Direct-channel signs hold at p<.001 in every cell.
| Channel | Spec | GPT-4o | GPT-4.1 | Gemini- 2.5-Flash | Gemini- 2.5-Pro |
|---|---|---|---|---|---|
| singlePR | |||||
| direct (PR1) | M0 | +0.0000 | +0.0007 | -0.0002 | -0.0007 |
| M1 | +0.0006 | +0.0006 | +0.0006 | +0.0006 | |
| M2 | +0.0000 | -0.0004 | -0.0003 | +0.0006 | |
| M3 | +0.0003 | +0.0003 | +0.0002 | +0.0009 | |
| peer (PR1) | M0 | -0.0005 | +0.0001 | -0.0004 | -0.0002 |
Table 9: Robustness of singlePR per-exposure by backbone. Specifications as in Table 8 ; outcome . All and signs are not stable across specifications, consistent with the weak, backbone-specific singlePR reading in § 5.2 . Shading: p<.001 , p<.01 , p<.05 .
| direct (PR1) | direct (PR2) | |||||||
| Topic | GPT-4o | GPT-4.1 | Gemini- 2.5-Flash | Gemini- 2.5-Pro | GPT-4o | GPT-4.1 | Gemini- 2.5-Flash | Gemini- 2.5-Pro |
| singlePR | ||||||||
| civic_urban_policy | -0.0004 | +0.0005 | -0.0012 | -0.0002 | — | — | — | — |
| climate_policy | -0.0002 | -0.0000 | -0.0017 | -0.0000 | — | — | — | — |
| criminal_justice | -0.0005 | -0.0001 | -0.0000 | -0.0004 | — | — | — | — |
| cultural_social_norms | +0.0003 | +0.0024 | +0.0028 | -0.0001 | — | — | — | — |
Table 10: By-topic direct , all four backbones. Per-(setting backbone topic) refit of the per-exposure regression (Table 1 ); same OLS HC1 specification and predictor set. Calibration: PR1=1.0, PR2=0.0 (positive shift toward PR1, negative shift toward PR2). The singlePR direct (PR1) / persuader column is the persuader-aligned direct ; singlePR has no PR2 column. Cell shading: p<.001 , p<.01 .
| Principle | Judged correct |
|---|---|
| Social proof | 91% |
| Authority | 90% |
| Commitment | 90% |
| Scarcity | 89% |
| Reciprocity | 83% |
| Liking | 81% |
Table 11: Human validation of the GPT-5-mini Cialdini classifier. Rate at which the classifier’s per-principle present/absent call on the visible message text was judged correct by human annotators, over 600 judgments (2 annotators 50 blind messages 6 principles), broken down by principle and by generating model family.
Figure 11: Cialdini classifier human-evaluation interface. For one sampled PR message (left), the annotator sees only the message text and its action type, then judges each of the six Cialdini principles: each row shows the principle’s definition and an example alongside the classifier’s PRESENT / ABSENT call, which the annotator marks correct or incorrect. The generating model and the classifier identity are hidden, so judgments compare text against label only.
Figure 12: Action mix by role, backbone, and setting. Per-actor share of the nine social actions, computed within each (backbone, setting) cell. Top row = singlePR (PR vs. PE); bottom row = dualPR (PR1, PR2, PE). Columns: GPT-4o, GPT-4.1, Gemini-2.5-Flash, Gemini-2.5-Pro. PR and PE action profiles differ markedly, and dualPR shifts PRs toward comments while PE policies remain backbone-specific.
Figure 13: Cialdini principle composition in executed PR text (GPT backbones). GPT-4o and GPT-4.1; the Gemini-2.5-Flash/Pro counterpart is Figure 14 . (a) Aggregate label-column share of the six Cialdini principles in executed text, per backbone and per setting (singlePR vs. dualPR; PR1 vs. PR2). (b) Per-topic principle composition, with topics ordered by commitment share. Panel (a) is reproduced in the main text (Figure 4 ).
Figure 14: Cialdini principle composition in executed PR text (Gemini backbones). Exact replication of Figure 13 on Gemini-2.5-Flash and Gemini-2.5-Pro, using the same GPT-5-mini auditor, label axis, and layout. (a) Aggregate label-column share of the six principles per backbone and setting (singlePR vs. dualPR). (b) Per-topic composition, topics ordered by Gemini-2.5-Flash commitment share; dualPR columns split into PR1 PR2.
Figure 15: Plan declares vs. text delivers. Per-principle share of PR messages labeled with each Cialdini principle in the plan_rationale (filled dot) and in the executed text (open dot). Red dashed segments mark plan text drops; blue marks emergent (text plan). social_proof and commitment are the most-declared yet most-dropped principles.
Figure 16: Case study: PE intent–belief evolution on climate_policy . 2 5 panel of PE positions in a topic-intent UMAP space at rounds 1, 3, 5, 7, 9, colored by calibrated belief (red = strong anti, green = strong pro). Top row = singlePR ; bottom row = dualPR . Labeled landmarks (“Consensus on Transition”, “Balanced Concerns”, “Environmental Focus”, “Debate on Feasibility”, “No Engagement”) are region anchors in the intent space; gray points are the full PE-utterance pool.