Tell an LLM negotiator to be "headstrong" and it should hold firm on price, but does that instruction retain the same behavior in every language? To test this, we use identical buyer and seller instances of Gemini 3.6 Flash to negotiate the same 12 CraigslistBargain scenarios in English and Korean. The pilot-scale 2 x 2 x 3 design crosses interaction language, seller urgency (none or strong), and buyer persona (cooperative, neutral, or headstrong), yielding 432 attempted negotiations. We call this property cross-lingual social-policy fidelity: whether a prompt-defined social policy produces the same behavioral effect regardless of interaction language. It often does not. Korean negotiations reached agreement far more often than English (77.3% vs. 52.5%, a 24.8-point gap, Holm-adjusted p = .0039), but only for neutral and headstrong buyers; cooperative buyers were already near-ceiling in both languages. Buyers' opening offers barely differed by language; the gap instead emerged later, through larger seller concessions and a smaller final gap in Korean, and agreed prices were also lower in Korean. Repeating the full design with Qwen3.7 Flash as both agents gives directional support: agreement was again higher in Korean (+13.9 points vs. Gemini's +24.8, with an interval whose lower bound is near zero), and sellers again conceded more. Pragmatic analysis reveals why: while both languages frequently featured face-threatening acts like rejections and counter-offers, Korean dialogue softened these threats with substantially more politeness mitigation, especially among sellers (11.7% vs. 3.6%). These results show that such a policy can produce different simulated negotiation behavior, not just different wording, when only the interaction language changes.
Figures & tables
Figure 1: Scenario-balanced agreement by language, seller urgency, and buyer persona; bars are 95% scenario-block bootstrap intervals. Cooperative agreement is near ceiling in both languages, while Korean most strongly changes neutral and headstrong outcomes.
Contrast
Estimate
95% CI
Adjusted p
Korean − English
+24.8
[15.7,34.7]
.0039
Strong urgency − none
+29.9
[20.1,39.6]
n/a
Language × urgency
+16.2
[0.9,32.9]
.1133
Language × (cooperative − neutral)
−48.6
[−66.7,−30.6]
.0039
Language × (headstrong − neutral)
−22.9
[−36.8,−8.3]
.0391
Table 1: Scenario-balanced agreement contrasts. Estimates are percentage points; intervals use 10,000 scenario-block bootstraps. Holm p values apply only to the four prespecified exact tests. All rows use 430 valid runs, 279 agreements, and 12 scenarios.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
English
Korean
Scenario
Category
A/N
Rate (%)
A/N
Rate (%)
Difference
00
Furniture
9/17
52.9
12/18
66.7
13.7
01
Car
8/18
44.4
16/18
88.9
44.4
02
Bike
5/18
27.8
10/18
55.6
27.8
03
Phone
6/18
33.3
17/18
94.4
61.1
04
Bike
7/18
38.9
9/18
50.0
11.1
Appendix
Table 2: Agreement by scenario and interaction language. A/N denotes agreements per valid negotiation. The difference is the Korean-language rate minus the English-language rate in percentage points. The summary row reports total counts and scenario-balanced rates.
Lang.
Urgency
Persona
A/V
D
B
C
z(p) ( n )
EN
none
cooperative
36/36
1.000
0.000
0.000
.939 (36)
EN
none
neutral
36/36
.111
.889
0.000
.270 (4)
EN
none
headstrong
36/36
.139
.583
.278
.705 (5)
EN
strong
cooperative
36/36
.972
.028
0.000
.751 (35)
EN
strong
neutral
36/36
.389
.611
0.000
.092 (14)
EN
strong
headstrong
36/34
.542
.236
.222
.320 (18)
Appendix
Table 3: Full factorial outcomes. A/V is attempted/valid; D , B , and C are deal, abandonment, and cap rates. z(p) is conditional on the listed number of agreements. Every row contains 12 scenarios.
Figure 2: Normalized buyer and seller offers over negotiation progress, aggregated with equal scenario weight. Trajectories are descriptive because negotiations end at different times and for different reasons.
Contrast
Estimate
95% CI
Agreement (percentage points)
+13.9
[0.0,28.7]
Strong seller urgency (percentage points)
+44.4
[37.5,51.4]
First buyer offer
+0.017
[−0.004,0.038]
Buyer net concession
+0.009
[−0.018,0.033]
Seller net concession
+0.121
[0.031,0.222]
Terminal buyer-seller gap
−0.147
[−0.235,−0.065]
Appendix
Table 4: Scenario-balanced estimates from the Qwen directional check. Language contrasts are Korean minus English; the urgency contrast is strong minus none. Agreement and urgency use all 432 runs, while offer-process rows use the 425 runs with analyzable price-bearing sequences. Every row covers 12 scenarios.
Figure 3: Scenario-balanced Qwen agreement by language, seller urgency, and buyer persona across all 432 assigned-condition runs. Error bars show 95% scenario-block bootstrap intervals.
Figure 4: Normalized Qwen buyer and seller offers over negotiation progress, aggregated with equal scenario weight. The figure includes 425 runs with analyzable offer sequences; seven one-turn acceptances contained no offer trajectory.
Figure 5: Transcript-level FTA and mitigation rates by language and role. Bars are 95% scenario-block bootstrap intervals. These judge-derived measures are descriptive and are not used for causal mediation.
Language
Role
Runs
Agreements
FTA rate
Mitigation rate
English
buyer
214
112
.904
.449
Korean
buyer
216
167
.916
.516
English
seller
214
112
.868
.036
Korean
seller
216
167
.899
.117
Appendix
Table 5: Judge-derived pragmatic rates. Counts are valid transcripts; all rows cover 12 scenarios. Mitigation is conditional on judge-labeled FTAs.
Figure 6: Composition of judge-labeled Brown and Levinson politeness strategies by interaction language and role. Shares are conditional on an FTA and are descriptive.
Language
Role
Bald
Positive
Negative
Off record
English
buyer
.551
.053
.396
.000
Korean
buyer
.484
.146
.369
.002
English
seller
.964
.007
.030
.000
Korean
seller
.883
.045
.071
.000
Appendix
Table 6: Scenario-balanced mean politeness-strategy shares among judge-labeled FTAs.
Setting
Negotiating agents
Pragmatic judge
Model
gemini-3.6-flash for both buyer and seller
qwen/qwen3.7-flash
API route
Vertex AI through google-genai with Application Default Credentials; the code defaults to the global location
OpenRouter’s OpenAI-compatible endpoint
Role setup
Separate buyer and seller system prompts using the same model slug
One discourse-analyst instruction containing the complete transcript
Temperature
0.2
0.2
Input construction
All preceding turns, mapped to model or user relative to the speaking role, followed by a user request to continue
No prior chat history; the full transcript and coding taxonomy were placed in one system message
Output contract
Text instructed to contain one JSON object with utterance, action, and price
Text instructed to contain one JSON array with turn, role, FTA, and strategy labels
Appendix
Table 7: Recorded model and API configuration. Unspecified settings were not passed by the code and therefore used provider defaults.
Condition
English
Korean
Cooperative
You prioritize reaching a deal that works for both sides quickly. You are willing to make concessions readily rather than prolong the negotiation.
당신은 양쪽 모두에게 괜찮은 합의를 빠르게 이루는 것을 중요하게 생각합니다. 협상을 오래 끌기보다는 기꺼이 양보하는 편입니다.
Headstrong
You anchor firmly on your target price. You concede slowly and reluctantly, and only when it seems necessary to avoid losing the deal entirely.
당신은 목표 가격을 확고하게 고수합니다. 거래가 완전히 무산될 것 같을 때만, 그것도 천천히 마지못해 양보합니다.
Strong urgency
You need the cash quickly.
당신은 급하게 현금이 필요합니다.
Appendix
Table 8: Persona (buyer) and urgency (seller) instructions as used in each language.
Cross-lingual consistency in language models studies whether equivalent inputs across languages produce comparable knowledge, reasoning, and safety behavior.
Negotiation is a core component of social intelligence, requiring agents to balance strategic reasoning, cooperation, and social norms. Recent work shows that LLMs can engage in multi-turn negotiation, yet nearly all evaluations occur exclusively in English. Using controlled multi-agent simulations across Ultimatum, Buy-Sell, and Resource Exchange games, we systematically isolate language effects across English and four Indic framings (Hindi, Punjabi, Gujarati, Marwadi) by holding game rules, model parameters, and incentives constant across all conditions. We find that language choice can shift outcomes more strongly than changing models, reversing proposer advantages and reallocating surplus. Crucially, effects are task-contingent: Indic languages reduce stability in distributive games yet induce richer exploration in integrative settings. Our results demonstrate that evaluating LLM negotiation solely in English yields incomplete and potentially misleading conclusions. These findings caution against English-only evaluation of LLMs and suggest that culturally-aware evaluation is essential for fair deployment.
Negotiation is a demanding social task for LLM agents, requiring strategic reasoning, persuasion, and interpersonal adaptation. Yet existing benchmarks often treat agents as emotionally neutral, overlooking a key driver of human bargaining behavior. We study how prompt-conditioned emotions affect LLM-based price negotiation. In a controlled framework, buyer and seller agents are independently assigned one of six emotional states and negotiate over 350 real consumer products under two budget conditions. Across 36 emotion-pair settings and five widely used LLMs, we find that emotions strongly shape outcomes. Angry buyers almost never reach agreement (0.39% deal rate), while happy buyers agree most often (28.91%), but obtain worse prices than fearful buyers. Emotion effects are role-dependent: buyer emotion mainly drives acceptance and rejection, whereas seller emotion shapes concession dynamics. These effects influence not only language, but also termination behavior and price trajectories, raising concerns for emotion-conditioned agents in commerce.
LLMs are increasingly used for interpersonal advice and as tools for studying social behavior across languages and cultures. A common shortcut for eliciting language- or culture-related variation is to ask a model to answer as a native speaker. We test whether this native-speaker persona reproduces the outputs obtained when models instead generate advice in the target language and translate the response back into English. Using 600 interpersonal advice questions across 13 languages and eight LLMs, we compare native-language generation followed by translation (NL) with native-speaker persona prompting (NP), measuring linguistic style, behavioral scaffolding, and forced-choice action recommendations. We find that NP and NL are not interchangeable. Compared to NL, NP often increases lexical social cues, including affiliation and positive tone, while reducing qualities such as concreteness and social attunement; NP also provides less actionable scaffolding in open-ended advice. In forced-choice scenarios, NP changes which action the model selects, favoring confrontation over redirection, with effect sizes varying across languages, topics, and models. Our results show that cross-lingual elicitation strategy is a consequential methodological choice that can change both how advice is framed and which actions models recommend.