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.
Figures & tables
Comparison
n
Both matched
Consensus
Low vs. High
14
100 [77, 100]
100 [77, 100]
Low vs. Neutral
15
67 [38, 88]
87 [60, 98]
Neutral vs. High
14
100 [77, 100]
100 [77, 100]
Neutral vs. Natural
19
68 [43, 87]
84 [60, 97]
All
62
82 [70, 91]
92 [82, 97]
Table 1: Blind validation of the linguistic-range rewriter, % matching the expected range relation ( no meaningful difference for Neutral vs. Natural). Both matched: both researchers chose it independently ( κ=0.74 ). Consensus: after a third annotator, blind to condition, adjudicated the 11 disagreements.
Skill
Vote
Trust
(pp)
(pp)
(points)
control drift
9.3%
14.2%
0.62
Low − control
+1.7
+2.1
+0.01
[ − 1.0, +4.4]
[ − 0.7, +4.8]
[ − 0.01, +0.02]
Neutral − control
− 1.0
− 2.1
− 0.00
[ − 3.6, +1.7]
[ − 4.5, +0.2]
[ − 0.02, +0.01]
Table 2: Linguistic range over 30 seeds (289 player-days). Skill and vote measured in percentage of decisions changed relative to the baseline; trust measured by mean absolute change in reliability scores (1–10 scale). 95% seed-clustered bootstrap intervals. ∗p<.05 (two-sided, uncorrected); no result remains significant after correcting for the 15 tests.
Requested
Recognised
Meaning kept
bald ( n=10 )
100 [69, 100]
70 [35, 93]
positive ( n=13 )
92 [64, 100]
85 [55, 98]
negative ( n=13 )
77 [46, 95]
92 [64, 100]
off-record ( n=13 )
69 [39, 91]
8 [0, 36]
all rewrites ( n=49 )
84 [70, 93]
63 [48, 77]
Table 3: Blind researcher validation of the politeness rewriter on the 49 delivered rewrites. Recognised: both researchers chose the requested strategy; meaning kept: both judged the meaning preserved. Exact 95% confidence intervals are provided. Cohen’s κ between researchers was 0.77 for strategy.
Dataset ( n )
Accuracy
κ
Politeness rewrites (43)
91 [78, 97]
0.88
Zhao and Hawkins (2025) (135)
70 [62, 78]
0.59
Table 4: Agreement of the LLM judge with human labels. Politeness rewrites are the items on which both researchers agreed in Table 3 .
n
Bald
Pos.
Neg.
Off
Werewolf
473
87
10
2
1
Villager
552
88
6
2
4
Seer
162
86
9
1
4
Guard
242
89
5
1
5
All
1,429
88
8
2
3
Table 5: Politeness strategies in natural play by speaker role (% of utterances), labeled by the politeness judge.
Skill
Vote
Trust
(%)
(%)
(points)
control drift
9.9%
16.0%
0.62
all_bald
+2.3 ∗
+1.0
+0.00
[+0.4, +4.0]
[ − 0.9, +2.8]
[ − 0.00, +0.01]
all_positive
+1.5
− 1.4
− 0.00
[ − 1.3, +4.5]
[ − 3.2, +0.4]
[ − 0.01, +0.01]
Table 6: Skill and vote measured in percentage of decisions changed. Trust measured by mean absolute difference in trust scores (1–10 scale). 95% confidence intervals provided. ∗p<.05 (two-sided, uncorrected); no result remains significant after correcting for the 12 tests.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Predicate
Meaning
claim_role (P, R)
claim a role (may be false)
claim_divine (P, T, A)
claim a seer result (may be false)
claim_guard (P, T)
claim a guard protection
estimate (P, T, R) †
state a read on T’s role
question (P, T, s) †
challenge T about statement s
answer (P, s)
answer question s
Appendix
Table 7: Speech predicates (phase argument omitted). † The nine face-threatening acts rewritten in the politeness experiment. The range experiment rewrites every speech act of the designated speaker.
Experiment
Condition
Rewrites
Original (%)
Range
Low
3,615
1.2
Neutral
0.7
High
1.0
Politeness
all_bald
7,785
11.0
all_positive
1.3
trust_positive
6.0
Appendix
Table 8: Share of treated utterances for which both attempts failed the gate and the original wording was delivered. Totals are pooled over conditions. In the range experiment, the highest fallback rate on any seed and level was 8.7%.
Table 9: Rewrite pipelines in validation and in the replays.
Measure
Low − Neu.
High − Neu.
Neu. − Nat.
( n=15 )
( n=14 )
( n=19 )
SUBTLEX Lg10CD
+0.26‡
−0.50
−0.03
MATTR
−0.015‡
+0.019
−0.002
Subord. clauses / 100w
−1.0‡
−0.34
−0.23
MTLD
−7.3
+13
–
Repetition
+0.060
–
−0.018
Appendix
Table 10: Range pilot measures, as passage-level differences between conditions. ‡ Primary measure whose 95% game-clustered interval excludes zero. ≈0 : within noise. –: not reported.
Requested
n
Res. 1
Res. 2
Judge
Bald on-record
13
12
13
13
Positive
13
12
13
11
Negative
13
12
11
13
Off-record
13
9
13
8
All
52
45
50
45
Appendix
Table 11: Items on which each rater chose the requested strategy, over all 52 items (including the 3 bald items where the original was shown).
Human \ Judge
Bald
Pos.
Neg.
Off
Other
Researcher-agreed rewrites ( n=43 )
Bald
12
0
0
0
0
Positive
0
11
1
0
0
Negative
0
0
10
0
0
Off
1
1
1
6
0
Zhao and Hawkins (2025) ( n=135 )
Appendix
Table 12: Politeness judge against human labels. Other: failed manipulation or no majority across passes.
Scores moved
Moved 2+
Score for speaker
(pp)
(pp)
(village, pts)
Low − control
+0.24
+0.70
+0.03
[ − 0.59, +0.99]
[+0.01, +1.43]
[ − 0.02, +0.09]
Neutral − control
− 0.00
+0.06
+0.04
[ − 0.87, +0.85]
[ − 0.60, +0.76]
[ − 0.02, +0.08]
High − control
+0.07
+0.73
+0.02
Appendix
Table 13: Additional reliability outcomes for linguistic range (30 seeds, 289 player-days), with 95% seed-clustered intervals.
Scores moved
Moved 2+
Score for speaker
(pp)
(pp)
(village, pts)
control level
30.08%
10.03%
− 0.11
all_bald
+0.42
+0.02
− 0.00
[ − 0.04, +0.91]
[ − 0.45, +0.52]
[ − 0.05, +0.05]
all_positive
− 0.29
+0.19
− 0.08
[ − 0.87, +0.28]
[ − 0.32, +0.70]
[ − 0.14, − 0.02]
Appendix
Table 14: Additional reliability outcomes for politeness, each setting minus the control (30 seeds, 289 player-days).
Listener
n
Non-bald
Village-side speakers: score for listener
Low (1–3)
44
2 (5%)
Mid (4–7)
481
53 (11%)
High (8–10)
12
4 (33%)
Werewolf speakers
Teammate (score 10)
26
2 (8%)
Appendix
Table 15: Non-bald utterances in natural play by the speaker’s reliability score for the listener (881 addressed, rated utterances).
Model
Role
DeepSeek Flash ( deepseek-chat )
game agents; in-game range and politeness rewriter; politeness validation rewriter
GPT Luna 5.6 ( gpt-5.6-luna )
in-game gate questions; politeness judge; second range listener
Qwen3.5-27B
range pilot rewriter, fidelity judge, inverse parser, first listener, CEFR rater