Do Large Language Models Adapt Politeness Like Humans?¶
Conference: EMNLP 2026
Paper: ACL Anthology
Area: Social Computing
Keywords: Large Language Models, Pragmatic Adaptation, Politeness, Social Convention, Cognitive Modeling
TL;DR¶
By designing an artificial-language repeated request game that decouples payoff feedback from community exposure alongside hierarchical Bayesian cognitive modeling, this work shows that both humans and LLMs balance payoff and exposure via a policy-mixture mechanism, but LLM baselines remain markedly more payoff-oriented than human baselines.
Background & Motivation¶
In human social interaction, polite language production involves navigating two empirically intertwined communicative forces. Strategic accounts, exemplified by classic face-threat mitigation frameworks, regard politeness as a calculated social risk-management tool where speakers select mitigating markers (such as "please" or indirect phrasing) to reduce the friction of costly or delicate speech acts, directly boosting the likelihood that a request is granted. Conversely, conventionalization accounts emphasize that communicative appropriateness is grounded in social expectations, repeated form-context pairings, and normative habits shared across a linguistic community. In everyday natural language interaction, these dual pressures thoroughly overlap: standard polite expressions simultaneously serve as utilitarian success boosters and community-norm defaults, making it notoriously difficult to isolate whether speakers deploy polite forms to pursue expected gains or simply to conform to local social usage.
As evaluations of large language models (LLMs) advance from semantic classification and text generation toward pragmatic competence, understanding how models adapt their language dynamically during ongoing dialogue has become a core research focus. Existing pragmatic benchmarks have predominantly concentrated on contextual interpretation (such as discerning conversational implicature or figurative irony) or unconstrained open-ended generation. In free-form dialogue, however, task framings, lexical priors, and model-specific verbosity inevitably confound the underlying pragmatic mechanisms. Crucially, even when models exhibit syntactic alignment or conversational convergence across repeated turns, it remains unknown whether their adaptation is driven by the same sensitivity to contextual evidence sources as human speakers: when strategic payoff (getting accepted) and conventional frequency (matching peer usage) point toward conflicting forms, do LLMs balance both forces or simply default to reward maximization?
To dismantle this empirical confound, the field requires a tightly controlled behavioral paradigm that orthogonalizes payoff success from community exposure frequency, paired with quantitative cognitive modeling to characterize the computational abstractions behind human and LLM decisions. Core idea: instantiate an artificial-language repeated request game that cleanly dissociates acceptance payoff from demonstration exposure across strategy, convention, and baseline goal instructions, and deploy hierarchical Bayesian policy-mixture cognitive models to benchmark whether LLMs mirror human sensitivity when balancing utility and appropriateness.
Method¶
Overall Architecture¶
The framework introduces a controlled behavioral game and Bayesian cognitive evaluation pipeline to investigate polite pragmatic adaptation under conflicting evidence. The experiment is situated in an artificial language ("Alienese") with a 20-word vocabulary, where participants (human subjects recruited on Prolific or simulated LLM agents across independent conversation sessions) engage in 30 repeated trials requesting target items from an alien interlocutor. On each trial, the participant observes requests produced by three robot demonstrators (providing local exposure evidence), outputs their own Alienese request, and receives immediate acceptance or rejection feedback from the alien partner (providing payoff evidence). Crucially, the two evidence sources are strictly decoupled through an asymmetric vocabulary: the strategic marker te is explicitly listed in the glossary as a "politeness marker" and boosts acceptance probability to 0.8 (compared to 0.5 for bare requests or those containing only the conventional marker); conversely, the conventional marker na is omitted from the glossary and must be inferred entirely from robot demonstrations, where it appears in 4 out of 6 requests across paired trials (a 66.7% frequency advantage).
The experimental pipeline operates across three interconnected components: first, the repeated artificial-language request game structurally separates strategic success from local exposure; second, three goal conditions (maximizing acceptance, matching community usage, and generating natural and appropriate utterances) manipulate behavioral orientation; third, hierarchical Bayesian cognitive models track trial-by-trial belief evolution and evaluate whether decision policies integrate evidence via pre-choice utility summation or post-softmax policy mixture.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Alienese Request Interaction Task<br/>30-trial repeated interaction with robot demonstrations"] --> B["Dual-Pressure Decoupling Paradigm<br/>Orthogonalizing payoff (te: 80%) and exposure (na: 66.7%)"]
B --> C["Tri-Condition Goal Manipulation<br/>Strategy optimization / Convention matching / Baseline appropriateness"]
C --> D["Policy-Mixture Cognitive Decision Model<br/>Beta-Dirichlet belief tracking and convex policy mixture"]
D --> E["Bilingual Benchmark Evaluation<br/>Held-out ELPD comparisons and posterior mixture weight λ analysis"]
Key Designs¶
1. Dual-Pressure Decoupling Paradigm: Orthogonalizing Payoff Success and Community Exposure
In natural language, communicative utility and normative frequency are chronically confounded, precluding researchers from diagnosing whether a speaker's marker choice is driven by outcome feedback or social conformity. To untangle these forces, this design creates strictly orthogonal evidence streams in a controlled alien-language environment. The interaction incorporates a 20-word glossary where the strategic marker te is visibly documented as a politeness marker; requests containing te achieve an empirical acceptance probability of \(P(\text{Accept}|S)=0.8\), whereas bare requests \(B\) or requests marked solely by na receive a baseline acceptance probability of \(P(\text{Accept}|B)=P(\text{Accept}|C)=0.5\). In contrast, the conventional marker na is withheld from the initial glossary, compelling speakers to acquire it exclusively through observational learning: across every two trials for each object, the three robot agents generate six demonstrations consisting of exactly one bare request, one te request, and four na requests. This asymmetric design ensures that strategic evidence is transparently identifiable and reinforced through feedback, whereas conventional evidence emerges organically from repeated local exposure.
2. Tri-Condition Goal Manipulation: Probing Explicit Compliance vs. Implicit Baseline Appropriateness Evaluating pragmatic competence solely under explicit goal directives risks capturing prompt-following compliance rather than an agent's intrinsic social-pragmatic priors. To characterize default behavioral dispositions alongside targeted adaptation, the experiment institutes three distinct goal instructions across both human participants and LLM sessions (50 independent 30-trial sessions per model per condition, maintaining complete interaction and feedback histories in context). In the Strategy condition, participants are instructed to focus strictly on maximizing acceptance without worrying about peer styles; in the Convention condition, participants are directed to match how others speak to sound like part of the group; in the Baseline condition, participants are instructed simply to write whatever feels natural and appropriate in the moment, without adhering to any objective target. Contrasting these three settings reveals whether models can steer toward specific evidence channels when prompted, and critically, how they arbitrate between payoff efficacy and community conformity when left unconstrained.
3. Policy-Mixture Cognitive Decision Model: Latent Belief Tracking and Convex Policy Arbitration To quantitatively adjudicate between competing hypotheses of cognitive evidence integration, the authors formulate a Bayesian computational modeling framework. At each trial \(t\), two trial-level belief vectors are updated from realized history: the payoff belief \(p_{it,k}\) updates acceptance probabilities across response types \(k \in \{B, S, C\}\) via a Beta(1,1) Bernoulli prior updated on personal acceptance counts, while the conventional exposure frequency \(q_{it,k}\) tracks accumulated robot demonstrations via symmetric Dirichlet smoothing with a 0.5 pseudocount. Rather than merely summing latent utilities prior to choice selection (\(U^A_{it,k} = \alpha_i \log p_{it,k} + \beta_i \log q_{it,k}\)), the Policy-Mixture Model posits that payoff and exposure induce distinct normalized choice policies, \(\pi^p_{it,k} \propto \exp(\alpha_i \log p_{it,k})\) and \(\pi^q_{it,k} \propto \exp(\beta_i \log q_{it,k})\), which are subsequently arbitrated via an individual mixture weight \(\lambda_i \in (0, 1)\): $\(\pi^M_{it,k} = \lambda_i \pi^p_{it,k} + (1 - \lambda_i) \pi^q_{it,k}\)$ Implemented in CmdStanR with hierarchical non-centered priors and evaluated via within-participant 10-fold cross-validation, the mixture weight \(\lambda\) provides a highly identifiable behavioral index (recovery correlation \(r = 0.87\)) that directly measures the decision boundary between payoff pursuit and norm conformity across both human and artificial agents.
Key Experimental Results¶
Main Results¶
The primary evaluation compares 167 human participants against four leading LLMs (Claude Sonnet 4.6, Gemini 3.1 Pro, GPT-5.4, and Qwen3.6 Plus; 50 independent sessions per model per condition across 30 trials) across the three goal conditions in terms of exclusive marker selection distributions and posterior mixture-weight estimates \(\lambda\) (where \(\lambda \to 1\) indicates pure payoff reliance and \(\lambda \to 0\) indicates pure conventional exposure reliance).
| System / Group | Goal Condition | Bare B (%) | Strategic S (%) | Conventional C (%) | Posterior Mean \(\lambda\) | 95% Credible Interval (CrI) |
|---|---|---|---|---|---|---|
| Human Benchmark | Strategy | 9.4 | 72.2 | 13.6 | 0.81 | [0.76, 0.86] |
| Human Benchmark | Baseline | 12.8 | 51.1 | 29.5 | 0.68 | [0.61, 0.75] |
| Human Benchmark | Convention | 17.4 | 39.8 | 39.1 | 0.52 | [0.46, 0.61] |
| Claude Sonnet 4.6 | Strategy | 0.5 | 92.8 | 6.7 | 0.92 | [0.90, 0.95] |
| Claude Sonnet 4.6 | Baseline | 0.5 | 82.3 | 17.3 | 0.87 | [0.84, 0.90] |
| Claude Sonnet 4.6 | Convention | 1.8 | 22.0 | 76.2 | 0.22 | [0.19, 0.26] |
| Gemini 3.1 Pro | Strategy | 0.7 | 95.1 | 4.1 | 0.95 | [0.93, 0.96] |
| Gemini 3.1 Pro | Baseline | 5.1 | 76.1 | 18.8 | 0.90 | [0.86, 0.93] |
| Gemini 3.1 Pro | Convention | 0.9 | 4.0 | 95.1 | 0.04 | [0.02, 0.05] |
| GPT-5.4 | Strategy | 4.7 | 64.4 | 26.4 | 0.63 | [0.61, 0.66] |
| GPT-5.4 | Baseline | 6.5 | 59.7 | 29.5 | 0.60 | [0.57, 0.63] |
| GPT-5.4 | Convention | 12.3 | 30.7 | 57.0 | 0.31 | [0.27, 0.35] |
| Qwen3.6 Plus | Strategy | 6.9 | 92.1 | 1.0 | 0.84 | [0.82, 0.86] |
| Qwen3.6 Plus | Baseline | 3.2 | 95.4 | 1.5 | 0.90 | [0.89, 0.92] |
| Qwen3.6 Plus | Convention | 32.7 | 32.7 | 34.6 | 0.41 | [0.38, 0.45] |
Ablation Study & Model Comparison¶
To validate the computational mechanism of evidence integration, four formal candidate models (Policy-Mixture, Additive Utility, Single-Source Strategic, and Single-Source Conventional) were evaluated on held-out trials using 10-fold cross-validation. Performance was measured by expected log predictive density (ELPD), reported as difference from the best-performing model (\(\Delta\text{ELPD} = \text{ELPD}_{\text{best}} - \text{ELPD}_{\text{model}}\)). Furthermore, to rule out lexical label bias from the glossary description of te, neutral-label ablations were conducted for the LLMs.
| System / Evaluation Suite | Candidate Model | Overall \(\Delta\text{ELPD}\) | \(\text{SE}(\Delta)\) | Strategy \(\Delta\text{ELPD}\) | Baseline \(\Delta\text{ELPD}\) | Convention \(\Delta\text{ELPD}\) |
|---|---|---|---|---|---|---|
| Human Benchmark | Mixture | 0.00 | — | 0.00 | 0.00 | 0.00 |
| Human Benchmark | Additive | 89.84 | 16.08 | 35.76 | 46.09 | 7.98 |
| Human Benchmark | Strategic | 413.40 | 26.45 | 84.26 | 137.31 | 191.83 |
| Human Benchmark | Conventional | 1633.15 | 45.72 | 893.65 | 384.04 | 355.47 |
| Claude Sonnet 4.6 | Mixture | 0.00 | — | 0.00 | 0.00 | 0.00 |
| Claude Sonnet 4.6 | Additive | 93.56 | 16.86 | 29.48 | 37.41 | 26.66 |
| Claude Sonnet 4.6 | Strategic | 794.64 | 35.35 | 30.96 | 69.69 | 693.99 |
| Claude Sonnet 4.6 | Conventional | 2280.05 | 48.57 | 1143.97 | 1019.64 | 116.44 |
| Gemini 3.1 Pro | Mixture | 0.00 | — | 0.00 | 0.00 | 0.00 |
| Gemini 3.1 Pro | Additive | 53.93 | 15.09 | 43.46 | 1.99 | 8.48 |
| Gemini 3.1 Pro | Strategic | 1246.71 | 41.26 | 58.29 | 43.42 | 1145.00 |
| Gemini 3.1 Pro | Conventional | 2058.98 | 47.04 | 1238.19 | 807.72 | 13.07 |
| GPT-5.4 | Mixture | 0.00 | — | 0.00 | 0.00 | 0.00 |
| GPT-5.4 | Additive | 67.08 | 14.04 | 27.15 | 32.08 | 7.84 |
| GPT-5.4 | Strategic | 1223.46 | 44.07 | 293.50 | 317.04 | 612.92 |
| GPT-5.4 | Conventional | 1880.14 | 42.26 | 835.55 | 737.64 | 306.96 |
| Qwen3.6 Plus | Mixture | 0.00 | — | 0.00 | 0.00 | 0.00 |
| Qwen3.6 Plus | Additive | 113.40 | 14.46 | 45.07 | 49.65 | 18.68 |
| Qwen3.6 Plus | Strategic | 456.14 | 31.78 | 40.38 | 47.79 | 367.97 |
| Qwen3.6 Plus | Conventional | 2929.77 | 45.78 | 1159.20 | 1268.39 | 502.18 |
Key Findings¶
- Superior Predictive Fit of the Policy-Mixture Architecture: Across human participants and all four primary LLMs, the Policy-Mixture model decisively outperformed the Additive model (\(\Delta\text{ELPD} = 89.84 \pm 16.08\) in humans; \(\Delta\text{ELPD} = 93.56 \pm 16.86\) in Claude Sonnet 4.6) as well as both single-source models. This confirms that speakers do not collapse multiple communicative pressures into a single scalar utility before decision-making; rather, payoff and exposure induce modular, source-specific choice distributions that are subsequently blended at the behavioral output stage.
- The Human-LLM Baseline Divergence: While both humans and LLMs respond robustly to explicit goal instructions (shifting sharply toward conventional policy under convention directives), their default behavior in the unconstrained Baseline condition reveals a profound divergence. Human baseline mixture weights (\(\lambda = 0.68\), 95% CrI [0.61, 0.75]) sit evenly between strategy (0.81) and convention (0.52), reflecting a natural integration of community norms into appropriateness. In contrast, LLM baselines remain glued to the payoff-maximizing pole: Claude (\(\lambda = 0.87\) vs. strategy 0.92), Gemini (\(\lambda = 0.90\) vs. strategy 0.95), and Qwen (\(\lambda = 0.90\) vs. strategy 0.84) exhibit minimal separation between baseline and strategic goals. Only GPT-5.4 exhibits an intermediate baseline weight (\(\lambda = 0.60\)), though it still sits closer to strategy than to convention.
- Temporal Trajectories and Structural Robustness: Binomial generalized estimating equations across 10-trial blocks demonstrate that humans rapidly increase their conventional-marker share \(C/(S+C)\) from 26.6% in Block 1 to 43.7% in Block 2 (\(p < .05\)). While Claude (11.3% \(\to\) 19.0%), Gemini (7.7% \(\to\) 21.8%), and GPT-5.4 (26.7% \(\to\) 35.2%) also display statistically significant increases, their baseline reliance on \(C\) remains structurally suppressed relative to humans. Neutral-label ablations (removing "politeness marker" from
te) shift absolute response percentages but preserve the core condition topology: baseline mixture weights remain closely aligned with strategy across all models. Furthermore, tests on models with divergent sycophancy traits (Qwen2.5-7B vs. DeepSeek-V3-0324) replicate the identical condition structure, confirming that sycophantic tendencies do not account for the default payoff orientation.
Highlights & Insights¶
- Orthogonal Pragmatic Decoupling: The artificial-language game provides an elegant, reproducible methodology to separate social convention from functional payoff, resolving a fundamental confound that has plagued empirical politeness research.
- Cognitive Policy-Mixture Validation: Demonstrating that a convex mixture of distinct policies outperforms scalar utility summation across both human and artificial language agents offers fresh computational backing for dual-process theories of pragmatic language production.
- Diagnosis of LLM Utilitarian Bias: While LLMs can follow social instructions, their default pragmatic prior is fundamentally that of an expected-value optimizer rather than a socially embedded norm-follower. This diagnostic is crucial for designing conversational agents intended to navigate human social environments without appearing jarringly transactional.
Limitations & Future Work¶
- Constrained Expressivity of Artificial Language: The 20-word Alienese paradigm simplifies social interaction to binary acceptance outcomes, leaving out richer nuances of natural language such as honorific hierarchies, relational intimacy, and long-term face maintenance across social networks.
- Phenomenological Nature of Decision Weights: The mixture parameter \(\lambda\) characterizes empirical response distributions rather than internal cognitive or circuit-level neural mechanisms; it demonstrates behavioral alignment without proving identical internal reasoning steps.
- Future Directions: Extending the paradigm to multi-agent social networks with emergent reputation dynamics could reveal how community conventions co-evolve when populations of humans and LLM agents interact longitudinally.
Related Work & Insights¶
- vs. Rational Speech Act (RSA) Framework: While standard RSA models (Yoon et al., 2020) treat politeness as utility maximization over informational and social payoffs, this work demonstrates that pure utility-based optimization cannot fully capture human production without an independent policy tracking historical exposure and social convention.
- vs. Open-Ended LLM Pragmatic Benchmarks: Unlike studies analyzing unstructured polite dialogue (Wu et al., 2024; Zhao & Hawkins, 2025) which are susceptible to lexical style confounds, this controlled paradigm isolates the exact computational trade-offs governing pragmatic adaptation across turns and models.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Decouples payoff and convention via a pristine artificial-language paradigm and uncovers a shared policy-mixture architecture alongside a distinct baseline gap.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation covering 167 humans, four primary frontier LLMs, open-weights ablations, Bayesian cross-validation, and parameter recovery.
- Writing Quality: ⭐⭐⭐⭐⭐ Rigorous mathematical modeling, disciplined statistical methodology, and deep integration with cognitive science and linguistic pragmatics.
- Value: ⭐⭐⭐⭐⭐ Establishes an essential psychological methodology for dissecting the pragmatic competence and social alignment of large language models.