Language as Trait, Politeness as Strategy

Large language model (LLM) agents coordinate through conversation, and the same message can be phrased in simpler or more complex language, bluntly or politely. We test whether such differences change how other agents judge and act, using the LURK Werewolf framework, which separates each utterance's speech-act predicate from its natural-language realization. We rewrite only the realization, check that the game content is preserved, and replay each game day from saved checkpoints against a re-sampled control. We vary linguistic range consistently across a speaker's messages and politeness strategy per utterance. Across 30 games and 4,335 replays, neither manipulation changes the trust other agents place in the speaker, and no effect on votes or night actions survives correction for multiple comparisons; the largest uncorrected effect is that shifting range away from a neutral paraphrase changes about 4 percentage points more of the other players' votes. In natural play, 88% of face-threatening speech is bald on-record regardless of role.

Research questions

Change the wording, keep the move

In LURK Werewolf, every utterance starts as a structured speech-act predicate, such as estimate, question, or recommend_action, with its speaker, targets, and referenced statements. The predicate is realized as natural language, and listeners only ever see the realization. That separation lets us intervene on form alone.

The gate checks player numbers, role labels, statement IDs, polarity, and action/target, then asks the judge whether the action and the reasons are unchanged.

Counterfactual replays

Thirty baseline games are played once with deepseek-chat, checkpointing the start of each day. Each living player takes a turn as the designated speaker whose messages are rewritten for the whole replayed day. Because the model is non-deterministic even at temperature 0, an unrewritten control replay measures the noise floor, and every treatment is compared with it. A branching cache shares all model calls up to the speaker's first rewritten message.

Outcomes, measured against the baseline day: vote and skill (share of other players' votes / night targets that change) and trust (mean absolute change in 1–10 reliability scores). Games are the unit: 95% intervals resample whole games, p-values come from a game-level sign-flip test, Holm-corrected per table.

Linguistic range: a trait held across a speaker's messages

Following the CEFR's notion of range and computational proficiency assessment, the rewriter produces three relative levels. Low uses frequent, repeated words and a narrower syntactic repertoire, while staying fully grammatical; Neutral paraphrases the untreated (Natural) text at the same range; High uses rarer, more varied vocabulary and broader syntax.

The manipulation is perceptible

Against Neutral, Low passages use more frequent words (SUBTLEX-US +0.26), lower lexical diversity (MATTR −0.015), and fewer subordinate clauses, without adding grammar errors. Both LLM listeners (GPT Luna 5.6, Qwen3.5-27B) order Low < Neutral < High in every passage and both presentation orders. Two researchers, blind to condition, agree with the expected relation in 82% of pairs (κ = 0.74).

Blind human validation of the range rewriter (% matching the expected relation; “no meaningful difference” for Neutral vs. Natural). Consensus: a third annotator, blind to condition, adjudicated the 11 disagreements.

…but it does not move trust

Paraphrasing at the original range (Neutral) leaves decisions where the control leaves them. Moving away from Neutral in either direction changes about 4 pp more of the other players' votes (High − Neutral +3.9 pp [+1.2, +6.5]; Low − Neutral +4.2 pp [+1.3, +7.3]), but part of this contrast is Neutral sitting slightly below the control, and nothing survives correction. Reliability scores do not move at any level: every interval lies within ±0.03 points against a control drift of 0.62.

Politeness: a strategy chosen per utterance

We rewrite nine face-threatening speech acts into Brown & Levinson's strategies: bald on-record, positive politeness (solidarity, in-group framing), negative politeness (minimizing imposition), and off-record (indirect). Two researchers recognize the requested strategy in 84% of delivered rewrites (κ = 0.77), and the LLM judge matches their agreed labels on 91% (κ = 0.88).

Blind researcher validation of the politeness rewriter: both researchers chose the requested strategy / both judged the meaning preserved. Exact 95% intervals.

How agents speak on their own (RQ2A)

We labeled all 1,429 eligible utterances in 30 separate untreated games. Natural play is overwhelmingly direct, and almost identically so for every role: roles do not differ detectably in how often they leave bald on-record (p = .73, permuting roles within games). The speech act matters more than the role: 18% of recommend_action utterances are non-bald, against 1% of question.

Does politeness change how others act?

We drop off-record (its rewrites lose meaning) and negative politeness (it confounds strategy with hedged certainty), and apply four settings to every rewritable utterance of the designated speaker: all_bald, all_positive, trust_positive (positive to trusted listeners, bald otherwise, which strengthens the natural pattern), and trust_bald (the inverse). If agents read politeness as a relational signal, the congruent and incongruent settings should differ.

Positive and bald rewrites produce effects of similar size on all three measures, whether applied uniformly or matched to the speaker's trust in the listener. Vote intervals all include zero, and no setting adds more than 0.01 points to the control's 0.62-point trust drift.

Takeaways

Limitations

• All game agents use one model (DeepSeek Flash) in one game with a fixed seven-player setup; other model families may behave differently.
• The range manipulation covers lexical and syntactic range, not the full construct of proficiency. Human-validated range rewrites come from a pilot with Qwen3.5-27B; in-game DeepSeek rewrites are checked automatically only.
• Researchers agreed poorly on whether politeness rewrites preserve meaning (κ = 0.14), and failed rewrites fall back to the original text, diluting treatment (11% for all_bald).
• Replays estimate within-day effects, 30 games limit precision for small effects, and the politeness judge has low off-record recall.

BibTeX

@misc{li2026language,
  title  = {Language as Trait, Politeness as Strategy: Controlled Tests
            of Linguistic Effects in {LLM} Social-Deduction Agents},
  author = {Li, Yuchen and Liu, Lisa and Zou, Runqi and Kuang, Shuxuan
            and Zhang, Qingquan and Togelius, Julian},
  year   = {2026},
  note   = {Preprint}
}

This work builds on LURK Werewolf: Evaluating Linguistic Bias in Large Language Models (Liu et al., IEEE CoG 2026). Code, prompts, game logs, and annotations will be released upon publication.