Which persuasion strategies, if any, are associated with donation compliance? Answering this requires fine-grained strategy labels across a full corpus and statistical tests corrected for multiple comparisons. We annotate all 10,600 persuader turns in the 1,017-dialogue PersuasionForGood corpus (Wang et al., 2019), where donation outcomes are directly observable, with a taxonomy of 41 strategies in 11 categories, using three open-source large language models (LLMs; Qwen3:30b, Mistral-Small-3.2, Phi-4). Strategy categories alone explain little variance in donation outcome (pseudo R2≈0.015, consistent across all three annotators). Guilt Induction is the only strategy significantly associated with lower donation rates (Δ≈−23 percentage points), an effect that replicates across all three models despite only moderate inter-model agreement. Reciprocity is the most robust positive correlate. Target sentiment and interest predict whether a donation occurs but show at most a weak correlation with donation amount. These findings suggest that strategy identification alone is insufficient to explain persuasion effectiveness, and that guilt-based appeals may be counterproductive in prosocial settings. We release the fully annotated corpus as a public resource.
Figures & tables
Figure 1: Distribution of 11 persuasion strategy categories (Qwen3:30b). Each bar shows the number of persuader turns assigned to the category; percentages indicate the category’s share of all N=10,600 persuader turns. Conversation Management turns (44.0%) are omitted.
Pres.
Abs.
pB
pFDR
Strategy–donation (22 tests, χ2 )
Guilt
32.7%
56.0%
<.001
<.001
Reciprocity
72.2%
52.2%
.034
.016
Commit./Cons.
63.5%
51.2%
.049
.016
Target response–donation (MWU)
Don.
No don.
p
Table 1: Bivariate associations with donation outcome (Qwen3:30b). Top : donation rates (%) in dialogues where the strategy is present (Pres.) vs. absent (Abs.); pB : Bonferroni, pFDR : Benjamini-Hochberg correction over 22 tests. Effect sizes are small ( ϕ=0.14 for Guilt, 0.10 for Reciprocity). Middle : mean target sentiment ( −1 to +1 ) and interest ( 0 – 2 ) in donated (Don.) vs. not-donated (No don.) dialogues, Mann-Whitney U test. Bottom : Pearson r between target response and donation amount (donors only). Cross-model robustness is reported in Table 4 .
Figure 2: (a) Donation rates in dialogues containing Guilt Induction ( n=104 ) vs. dialogues without ( n=913 ); difference of −23.3 pp ( χ2=19.4 , p<0.001 ). This effect replicates across all three annotators (Table 4 ). (b) Donation rate by predominant target sentiment across a dialogue’s target turns ( χ2(2)=44.3 , V=0.21 , p<0.001 ).
Predictor
OR
95% CI
p
Model 3 (parsimonious): pseudo R2=0.080 , AUC =0.67
Sentiment
3.94
[2.30, 6.74]
<.001
Interest
2.77
[1.64, 4.67]
<.001
Reciprocity
2.41
[1.38, 4.20]
.002
Guilt Ind.
0.60
[0.37, 0.95]
.029
Table 2: Logistic regression predicting donation (binary), parsimonious Model 3. OR: odds ratio ( >1 = higher donation probability). Sentiment: mean target sentiment per dialogue ( −1 to +1 ). Interest: mean target interest per dialogue ( 0 – 2 ).
Strategy
n
Mean sent.
% neg.
Bottom 3 (lowest mean sentiment)
Guilt Induction
114
+0.02
27.2%
Fear Appeal
69
+0.19
21.7%
Unity
43
+0.23
23.3%
Top 3 (highest mean sentiment)
Commit. & Cons.
182
+0.54
7.1%
Table 3: Mean target sentiment ( −1 to +1 ) in the turn immediately following each persuader strategy ( n=5,387 persuader–target pairs). Only persuasion strategies with n≥20 pairs shown; Conversation Management labels excluded. % neg.: proportion of negative target responses.
Finding
Qwen
Mistral
Phi-4
Repl.
Guilt Δ (pp)
−23.3 ***
−23.0 *
−23.9 **
3/3
Recip. Δ (pp)
+20.1 **
+5.8
+10.4 *
2/3
Categories R2
.015
.011
.016
3/3
+ Sent./Int. R2
.082
.079
.077
3/3
Table 4: Cross-model robustness. Δ : difference in donation rate (pp) between dialogues with vs. without the strategy. R2 : McFadden pseudo R2 . Significance: ∗p<.05 , ∗∗p<.01 , ∗∗∗p<.001 (uncorrected praw ; multiple-comparison correction is applied only to the primary Qwen model in Table 1 ). Repl.: models where the finding reaches significance ( praw ) or matches direction (for R2 ).
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Category
Strategy
n
%
Norms / Morality / Values
Appeal to Values
543
5.1
Moral Appeal
521
4.9
Guilt Induction
129
1.2
Self-feeling Appeal
138
1.3
Rational / Impact Appeal
Rational Appeal
928
8.8
Logical Appeal
45
0.4
Appendix
Table 5: Complete taxonomy: 41 persuasion strategies in 11 categories plus 9 conversation management labels (Qwen3:30b annotations). Columns show absolute turn counts and each label’s percentage of all N=10,600 persuader turns (percentages sum to 100%). Five of the 41 taxonomy strategies (Punishing Activity, Overloading, Confusion Induction, Promise, and Activation of Impersonal Commitment) received zero assignments and are not shown.
Large language models (LLMs) still struggle with strategic persuasion, largely because existing approaches either neglect information asymmetry or rely on unrealistic pre-commitment assumptions. We introduce a type-induced commitment-communication mechanism that grounds Bayesian Persuasion (BP) in natural language dialogue without pre-commitment: the persuader narrates their potential types (e.g., honest vs. dishonest) to dynamically construct an information schema, enabling the persuadee to perform Bayesian belief updates within the conversation itself. We implement two variants: Semi-Formal-Natural-Language (SFNL) and Fully-Natural-Language (FNL), evaluating them against strong baselines across multiple LLMs and human judges. BP strategies consistently outperform baselines: SFNL excels in logical credibility, while FNL shows superior robustness and emotional resonance. We verify that gains stem from genuine Bayesian reasoning rather than superficial formatting, and we further show that supervised fine-tuning enables small models to match the persuasive performance of much larger ones.
Buwei He, Yang Liu, Zhaowei Zhang +6
Beijing University of Posts and Telecommunications · Beijing Institute for General Artificial Intelligence · Peking University +1
Large Language Model-powered conversational agents (CAs) are increasingly capable of projecting sophisticated personalities through language, but how these projections affect users is unclear. We thus examine how CA personalities expressed linguistically affect user decisions and perceptions in the context of charitable giving. In a crowdsourced study, 360 participants interacted with one of eight CAs, each projecting a personality composed of three linguistic aspects: attitude (optimistic/pessimistic), authority (authoritative/submissive), and reasoning (emotional/rational). While the CA's composite personality did not affect participants' decisions, it did affect their perceptions and emotional responses. Particularly, participants interacting with pessimistic CAs felt lower emotional state and lower affinity towards the cause, perceived the CA as less trustworthy and less competent, and yet tended to donate more toward the charity. Perceptions of trust, competence, and situational empathy significantly predicted donation decisions. Our findings emphasize the risks CAs pose as instruments of manipulation, subtly influencing user perceptions and decisions.
Uğur Genç, Heng Gu, Chadha Degachi +3
Delft University of Technology, Delft, Netherlands
Training large language models (LLMs) for multi-turn strategic persuasion faces two coupled obstacles: (i) realistic simulation of psychologically grounded persuadees, and (ii) dense, faithful supervision for credit assignment in strategic dialogues. Outcome-only rewards are sparse and uninformative about which conversational moves drove success, while LLM-judge process rewards are expensive and decoupled from the underlying interaction dynamics. We address both challenges with a unified framework. First, we introduce PersuArena, a persuasion simulation environment that produces realistic, non-monotonic trajectories. Second, building directly on this environment, we propose Process Rewards via Online Belief Elicitation for Relative Policy Optimization (Probe-RPO), an online RL framework that converts the persuadee's internal state into dense process rewards. At each turn, three introspective probes are read from the frozen persuadee's logits, yielding a multi-dimensional reward signal that is faithfully aligned with persuasion dynamics. With extensive experiments on PersuArena spanning healthcare, financial advising, workplace management, education, and personal relationships, we demonstrate that our Probe-RPO achieves the full acceptance rate of 52.5%, compared with the 35.8% for the strongest GRPO ablation and 25.8% for the best off-the-shelf LLM. We also conduct comprehensive analyses on the generalization across persuadee models and the persuasion dynamics across conversations. Our work offers a promising paradigm that can provide effective process supervision for strategic dialogue.