Paper deep dive
CC-Mediation: Evaluating Large Language Models for Cross-Cultural Conflict Mediation
Suhyun Lee, Wenxuan Zhang, W. Quin Yow, Yang Deng
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 9/7/2026, 6:42:51 AM
Summary
The paper introduces CC-Mediation, a benchmark for evaluating Large Language Models (LLMs) in cross-cultural conflict mediation. It comprises 1,661 dialogues grounded in the Developmental Model of Intercultural Sensitivity (DMIS). The authors propose two evaluation metrics: Trajectory AUC and signed Wasserstein-1 distance, which measure the persistence and magnitude of intercultural stance shifts. Experiments reveal that current LLMs struggle with intervention timing due to positional priors and with mediation strategy due to late-layer elicitation collapse, though fine-tuning significantly improves performance.
Entities (9)
Relation Signals (6)
CC-Mediation → usestheory → Developmental Model of Intercultural Sensitivity
confidence 98% · CC-Mediation, a cross-cultural mediation benchmark of 1,661 ten-turn dialogues grounded in the Developmental Model of Intercultural Sensitivity (DMIS)
CC-Mediation → evaluatedby → Trajectory AUC
confidence 97% · We further propose two DMIS-based evaluation metrics: Trajectory AUC... Using CC-Mediation, we find...
CC-Mediation → evaluatedby → Wasserstein-1 distance
confidence 97% · We further propose two DMIS-based evaluation metrics: ... a signed Wasserstein-1 distance... Using CC-Mediation, we find...
LLMs → exhibitsfailuremode → positional prior
confidence 95% · intervention timing (when) failure stems from a positional prior that ignores dialogue content
LLMs → exhibitsfailuremode → elicitation collapse
confidence 95% · mediation strategy (how) failure arises from a late-layer elicitation collapse rather than a knowledge deficit
GPT-4o-mini → usedfor → CC-Mediation
confidence 90% · We adopt GPT-4o-mini (Hurst et al., 2024) for the benchmark construction.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Cross-cultural mediation by large language models (LLMs) requires deciding both when to intervene and how to respond in culturally grounded conflicts. Progress on this problem has been limited by the lack of (1) mediation datasets with measurable downstream effects and (2) principled metrics for evaluating intercultural stance change. To address these gaps, we introduce CC-Mediation, a cross-cultural mediation benchmark of $1{,}661$ ten-turn dialogues grounded in the Developmental Model of Intercultural Sensitivity (DMIS), containing culturally grounded conflicts, mediation interventions, and post-intervention trajectories. We further propose two DMIS-based evaluation metrics: Trajectory AUC, which measures the persistence of intercultural improvement over time, and a signed Wasserstein-1 distance, which measures the magnitude and direction of shifts in intercultural stance. Both metrics show strong agreement with human judgment of DMIS-grounded stance shift. Using CC-Mediation, we find that current LLMs have limitations on both axes: intervention timing (when) failure stems from a positional prior that ignores dialogue content, while mediation strategy (how) failure arises from a late-layer elicitation collapse rather than a knowledge deficit.
Tags
Links
- Source: https://arxiv.org/abs/2609.04855v1
- Canonical: https://arxiv.org/abs/2609.04855v1
Trouble viewing inline? Open PDF directly →
Full Text
117,373 characters extracted from source content.
Expand or collapse full text
C-MEDIATION: Evaluating Large Language Models for Cross-Cultural Conflict Mediation Suhyun Lee 1 * , Wenxuan Zhang 2 , W. Quin Yow 2 , Yang Deng 3 1 Department of Artificial Intelligence, Hanyang University, Seoul, Republic of Korea 2 Singapore University of Technology and Design, Singapore 3 School of Computing and Information Systems, Singapore Management University, Singapore su7561632@hanyang.ac.kr wxzhang, quin@sutd.edu.sg ydeng@smu.edu.sg Abstract Cross-cultural mediation by large language models (LLMs) requires deciding both when to intervene and how to respond in culturally grounded conflicts. Progress on this problem has been limited by the lack of (1) mediation datasets with measurable downstream effects and (2) principled metrics for evaluating in- tercultural stance change. To address these gaps, we introduce C-MEDIATION 1 , a cross- cultural mediation benchmark of1,661ten- turn dialogues grounded in the Developmental Model of Intercultural Sensitivity (DMIS), con- taining culturally grounded conflicts, mediation interventions, and post-intervention trajectories. We further propose two DMIS-based evalua- tion metrics: Trajectory AUC, which measures the persistence of intercultural improvement over time, and a signed Wasserstein-1 distance, which measures the magnitude and direction of shifts in intercultural stance. Both metrics show strong agreement with human judgment of DMIS-grounded stance shift. Using C- MEDIATION, we find that current LLMs have limitations on both axes: intervention timing (when) failure stems from a positional prior that ignores dialogue content, while mediation strat- egy (how) failure arises from a late-layer elici- tation collapse rather than a knowledge deficit. 1 Introduction While the globalization of remote work and on- line communities scales cross-cultural interaction, technological connectivity does not ensure mutual understanding (Hinds et al., 2011). Conflicting cultural frames often lead to misunderstandings that outpace human moderation capabilities (Park et al., 2025). Although LLMs show promise for social conflict intervention (Li et al., 2025; Koçak * Work was done during a visit at SUTD. 1 Dataset and code available athttps://github.com/ suhyun565/C-Mediation Figure 1: An example of C-MEDIATION and DMIS stage transition. At turnt 6 , differing cultural perspec- tives on “good neighbour” between Ryan (Cyprus) and Mia (Singapore) trigger a DMIS Defense-stage conflict. The model intervenes with a “Shared-values Reframe” strategy to mitigate adversarial framing and focus on a shared purpose. This mediation effectively lowers cognitive defenses, leading to the Acceptance stage. et al., 2026), cross-cultural mediation is fundamen- tally more complex than standard toxicity detection (Qian et al., 2019) or content filtering (Kim et al., 2022). As illustrated in Figure 1, Ryan and Mia hold conflicting views on what is a “good neighbor,” each with justifiable reasons from their respective cultural perspectives. In this case, simply correct- ing their blameful language is insufficient to bring them to accept each other’s cultural standpoints; instead, the conflict att 6 must be re-framed around a deeper shared value (e.g., a “positive living envi- ronment”). As this example illustrates, mediation must be tailored to each participant’s intercultural sensitivity stage, as defined by the Developmental Model of Intercultural Sensitivity (DMIS) (Ben- arXiv:2609.04855v1 [cs.CL] 4 Sep 2026 nett, 1986, 2017), and requires culturally grounded reasoning, DMIS stage-aware strategies, and prin- cipled evaluation throughout the entire interaction. However, improving and evaluating LLMs for cross-cultural mediation face two major chal- lenges. First, there is no benchmark for cultur- ally grounded conflict mediation. Existing con- versational safety and moderation datasets labeled only toxicity, hate speech, or norm violations (Qian et al., 2019; Kim et al., 2022; Zhan et al., 2024; Hua et al., 2024), not mediation trajectories with theory- grounded intervention strategies and measurable interaction outcomes. Second, there is no prin- cipled metric for evaluating mediation effective- ness. While prior work has proposed downstream- effect metrics in other domains (Timm et al., 2025; Bozdag et al., 2025; Tan et al., 2025, 2026; Du et al., 2025; Deng et al., 2024), no quantitative metric tracks how substantively a mediation shifts the addressee’s stance away from an ethnocentric stage—a level of intercultural sensitivity that treats one’s own culture as the default standard for judg- ing others. To address these challenges, we introduce a theory-grounded evaluation and data construction framework based on the DMIS. The DMIS posits six ordinal stages progressing from ethnocentric (Denial→Defense→Minimization) to ethnorel- ative (Acceptance→Adaptation→Integration) level of intercultural sensitivity. We first pro- pose two complementary metrics defined over the ordinal DMIS scale: (1) a trajectory AUC met- ric that measures how persistently an interven- tion improves the addressee’s intercultural stance throughout the dialogue trajectory, and (2) a signed Wasserstein-1 distance that quantifies both the mag- nitude and direction of the distributional shift in- duced by mediation. Anchored to these metrics, we construct the C-MEDIATION benchmark, a corpus of1,661dialogues grounded in the DMIS theory. Using C-MEDIATION, we study two comple- mentary tasks: conflict detection (when to inter- vene) and conflict mediation (how to intervene with a DMIS stage-appropriate strategy). Current LLMs fail on both: conflict turn-detection accu- racy (TurnAcc) below33%, and trajectory AUC at or below1.3—which we trace to a positional prior in detection and a late-layer elicitation col- lapse in mediation generation. Fine-tuning on C- MEDIATION lifts TurnAcc above91%across all three open-source backbones, and under timing- controlled comparison significantly improves me- diation content (Judge) on the two larger backbones (§4.3). Our main contributions are: •We introduce C-MEDIATION, a cross-cultural mediation corpus grounded in intercultural com- munication theory. • We propose two DMIS-based mediation evalu- ation metrics, including a trajectory AUC and a signed Wasserstein-1 distance, for quantifying the persistence, magnitude, and direction of in- tercultural stance shifts induced by mediation interventions. •Through C-MEDIATION, we uncover the mech- anisms behind LLM limitations in cross-cultural mediation across two axes: timing (when) and strategy (how).Timing failure stems from an LLM bias toward superficial positional pat- terns. Conversely, representation-level analysis reveals that the failure to generate effective medi- ation arises from “eliciting failure” rather than a “knowledge deficit”. 2 Related Work Cultural Dialogue and Conflict-Mediation DatasetsExisting datasets fail to jointly provide the three components required for cross-cultural mediation: culturally grounded conflicts, media- tor interventions, and measurable downstream ef- fects on the addressee. Cross-cultural dialogue and value-aware datasets describe culturally sensitive scenarios but contain no mediator responses (Wu et al., 2025; Hale et al., 2025; Cao et al., 2024; Li et al., 2024), while hate-speech intervention, proso- cial response, and norm-remediation corpora sup- ply mediator utterances for non-cultural conflicts without evaluating their long-term impact on the recipient (Qian et al., 2019; Kim et al., 2022; Zhan et al., 2024; Hua et al., 2024). C-MEDIATION is the first dataset to jointly model culturally grounded mediation dialogues together with measurable in- tercultural stance shifts. A condition-by-condition comparison against existing corpora on the three re- quirements (cross-cultural speakers, cultural-value conflict, multi-turn dialogue) is provided in Ap- pendix G. Evaluation Metrics for Conversational Media- torsPrior work evaluates mediator utterances in two ways. One line of work performs utterance- level scoring through surface-form overlap, human preference, or LLM/human rubrics (Qian et al., 2019; Kim et al., 2022; Zhan et al., 2024; Hua et al., 2024; Wu et al., 2025; Liu et al., 2023). Another measures the downstream effect on the addressee via stance-shift, emotional trajectory, or task suc- cess (Timm et al., 2025; Bozdag et al., 2025; Tan et al., 2025, 2026; Du et al., 2025; Deng et al., 2024). However, existing metrics do not jointly cap- ture two key properties required for cross-cultural mediation: the magnitude of intercultural stance change and the persistence of that improvement over time. 3 Cross-Cultural Mediation 3.1 Task Definition We define cross-cultural mediation as a turn-level dialogue task in which a third-party mediator aims to shift one speaker’s stance toward greater eth- norelativism during a culturally grounded conflict. Each speaker is characterized by an intercultural understanding stageφand a value-conflict topic on which the two cultures diverge;φis latent and must be inferred from observable cues such as topic avoidance, us-versus-them framing, or universalist appeals. Mediation is considered effective when the addressee’s stance, measured on the same ordi- nal scale asφ, has advanced toward ethnorelativism by the end of the conversation. This formulation highlights three core proper- ties of cross-cultural mediation: (P1) Stage-aware intervention: mediator responses are meaningful only when grounded in the speaker’s current in- tercultural stance, supplied through turn-level con- flict annotations; (P2) Trajectory-level influence: mediation affects the addressee’s subsequent de- velopmental trajectory rather than producing an instantaneous resolution, supplied through post- intervention continuations with per-turn DMIS la- bels; and (P3) Measurable stance shift: media- tion effectiveness requires quantitative comparison between pre- and post-intervention intercultural states, supplied through the two ordinal DMIS- scale metrics defined in §3.3. 3.2 DMIS Framework We operationalizeφvia Bennett’s Developmental Model of Intercultural Sensitivity (DMIS) (Bennett, 1986, 2017). While DMIS was originally formu- lated to measure an individual’s enduring lifelong orientation toward cultural difference, we follow the situational reading adopted by the Intercultural Development Inventory (IDI) (Hammer et al., 2003; Hammer, 2011)—the most widely used psychome- tric instrument for DMIS—and treat each turn as a momentary expression of the stage the speaker has activated in that moment. Stages DMIS defines six ordinal stages ranging from ethnocentric to ethnorelative orientations: 1. Denial: failing to perceive cultural difference; 2.Defense: treating difference as threat through us/them dichotomies; 3.Minimization: subsuming difference under as- sumed universal similarities; 4.Acceptance: recognizing difference as equally valid alternatives; 5.Adaptation: shifting one’s perspective to com- municate across frames; 6. Integration:internalizing multiple frames. We restrict conflict speakers to the three eth- nocentric stages (Denial/Defense/Minimization), since each reflects a distinct form of cultural mis- understanding and therefore requires different me- diation strategies. Mediation success is defined as movement from the ethnocentric stages toward the other three ethnorelative stages. The ordinal structure of these stages dictates our metric choices in §3.3: a distributional metric that respects de- velopmental distance (Rubner et al., 2000; Vil- lani, 2009) rather than treating stages as categori- cal (Kullback and Leibler, 1951; Lin, 1991; Levin and Peres, 2017), and a trajectory-level integration of developmental progress (Tan et al., 2026) rather than a single-turn comparison. Detailed definitions of these six stages and the corresponding stage- conditioned mediation strategies are provided in Appendix A, with worked examples illustrating how each ethnocentric stage produces structurally distinct mediation trajectories in Appendix D. 3.3 Mediation Evaluation Metrics At each turnt, an LLM judge predicts a proba- bility distributionD (t) over the six DMIS stages, derived from the normalized probabilities assigned to the stage labels (Appendix F.4). The correspond- ing argmax stage is denoted asˆs (t) ∈ 0,..., 5, where0and5correspond to Denial and Integration, respectively. Based on these outputs, we define two complementary measurements of mediation effec- tiveness. Trajectory AUC. This measurement captures both the magnitude and persistence of stage mo- tion. Taking the pre-intervention stageˆs (t pre ) as Figure 2: C-Mediation Construction Pipeline. Our four-stage pipeline transforms a cultural-conflict scenario into a labeled mediation dialogue: (1) two LLM agents generate a culturally grounded conflict dialogue via role-play; (2) a judge LLM assigns DMIS stage labels per turn; (3) a mediator LLM insertsu + att conf and the dialogue resumes; (4) filtering via a strict utility-gap gate (∆Trajectory AUC > 0∧ ∆W 1 > 0). the origin, we define the per-turn progress as the difference of developmental-scale positions, p(t) = s ˆs (t) − s ˆs (t pre ) , and integrate it over the normalized time axisτ ∈ (0, 1] via the trapezoidal rule,Trajectory AUC = R 1 0 p(τ )dτ. We adopt uniform spacing on the de- velopmental scale,s i = ifori ∈ 0,..., 5, so the per-turn progressp(t)takes integer values in [−5, 5], and the integrated Trajectory AUC lies in the same range[−5, 5]. Positive values indicate sus- tained advancement toward ethnorelativism, while negative values indicate regression. Unlike final- state comparisons, Trajectory AUC (AUC) cap- tures the entire developmental trajectory and dis- tinguishes transient improvements from sustained progress. Signed Wasserstein-1 Distance. AUC relies only on argmax stages and may miss sub-argmax distributional shifts. To capture finer-grained move- ment, we additionally measure the distributional change between the pre-intervention stage distri- butionD pre and the post-intervention distribution D post . Since DMIS stages form an ordinal devel- opmental scale, we use the signed Wasserstein-1 distance (W 1 ): W 1 = sgn E[D post ]− E[D pre ] | z direction · X 4 i=0 F pre (i)− F post (i) · (s i+1 − s i ) | z magnitude , (1) whereF pre andF post are the cumulative distribu- tions over the six stages. Unlike categorical dis- tances (Kullback and Leibler, 1951; Lin, 1991; Levin and Peres, 2017) and other downstream mea- surements (Timm et al., 2025; Bozdag et al., 2025; Tan et al., 2025, 2026; Du et al., 2025; Deng et al., 2024),W 1 respects the ordinal geometry of the developmental scale, assigning larger penalties to larger stage movements. The sign term further in- dicates whether the shift moves toward or away from ethnorelativism. Like AUC,W 1 also lies in [−5, 5]. Human Validation. We validate both metrics through three complementary human evaluations covering stage-label prediction accuracy, pairwise agreement with human judgments of mediation success, and correlation with human Likert rat- ings of intercultural-improvement likelihood (Ap- pendix B). 3.4C-MEDIATION Benchmark As shown in Figrue 2, the C-MEDIATION con- struction pipeline consists of four LLM-driven stages that transform a cultural-conflict scenario into a labeled mediation dialogue. We adopt GPT- 4o-mini (Hurst et al., 2024) for the benchmark con- struction. The robustness of model choices is in- vestigated in Appendix C.1. All stages are human- validated, and prompts are provided in Appendix F. Step 0: Cultural-conflict Scenario Collection To construct a robust and realistic foundation for cross-cultural conflict mediation, this study builds upon the publicly available scenario pool from So- cialCC (Wu et al., 2025). This pool comprises 3,060scenario templates that instantiate empiri- cal population-level value distributions from the World Values Survey (WVS) (Haerpfer et al., 2022) into concrete micro-interactions between two agents. Each scenario provides paired coun- try memberships (⟨C a ,C b ⟩) with opposed model responses on specific value dimensions (e.g., social attitudes, religious values), alongside individual agent profiles and contextual goals. Step 1: Conflict Dialogue Generation Two LLM-based dialogue agents are initialized with cul- turally grounded profiles, value orientations, and interaction goals (Appendix F.1). The agents first engage in value-neutral conversation before a con- flict instruction is injected at a randomly desig- nated conflict turn t conf (Appendix F.2), triggering disagreement on a culturally sensitive topic. The agents then continue the conversation conditioned only on their profiles and dialogue history, produc- ing a natural post-conflict interaction trajectory up to the maximum turn limitT max . Human validation of dialogue quality is reported in Appendix B.1. Step 2: DMIS Labeling Each dialogue is la- beled turn-by-turn with DMIS stages. Pre-conflict turns are classified as eitherunrelated_topicor transitionusing a phase labeler (Appendix F.3). The conflict turn inherits the scenario’s predefined ethnocentric stage (Denial, Defense, or Minimiza- tion). All subsequent turns are labeled using the six- stage DMIS labeler described in §3.2. These labels serve both as evaluation targets for conflict-stage detection and as inputs for ground-truth mediation generation in Step 3. Step 3: Ground-Truth Mediation Generation At the intervention turn, the mediator receives the dialogue prefix up tot conf together with stage-specific DMIS mediation guidelines (Ap- pendix F.5). The original post-conflict continuation generated in Step 1 is discarded, and the media- tor generates a replacement intervention utterance. The dialogue is then resumed from that point using the same backbone model, producing a new post- intervention trajectory that is re-labeled to compute AUC and W 1 scores. Step 4: Filtering To ensure data quality, we ap- ply a utility-based filtering step to the generated mediations. The filter is a quality-control mecha- nism for the training supervision signal: for each supervised mediation, we additionally generate an alternative without stage-specific guidance (Ap- pendix F.6) as a reference baseline. Both are in- jected into their respective dialogue branches and Table 1: C-MEDIATION dataset statistics TrainEvalTotal # Dialogues1,503158 1,661 # Speaker utterances16,448 1,633 18,081 # GT mediations1,503158 1,661 # Filtered mediations1,503— 1,503 timing-differs1,313 (87.4%)— 1,313 content-only190 (12.6%)— 190 Avg. # turns10.94 10.27 10.88 Avg. utterance length62.49 52.70 61.65 Avg. mediation length62.92 56.24 62.27 ∆ AUC +0.701— ∆ W 1 +0.507— Table 2: Per-turn label distribution in C-MEDIATION PhaseLabelCountProp. Pre-conflict unrelated_topic2,1247.0% transition11,56838.4% Conflict turn Denial5731.9% Defense5221.7% Minimization5671.9% subtotal3,03110.1% Post-conflict Denial5221.7% Defense4,00813.3% Minimization1,6985.6% Acceptance7,57525.1% Adaptation230.1% Integration00.0% subtotal13,37644.4% Total30,099 100.0% evaluated via AUC andW 1 ; a supervised mediation is retained only if it strictly outperforms the un- guided baseline on both metrics. This ensures that retained samples exhibit measurable effectiveness gains attributable to stage-specific guidance, rather than stylistic variation alone. Crucially, the filter is applied only to the1,503training mediations; the158evaluation mediations are used entirely un- filtered, so all results in §4 are measured on data untouched by the filter and cannot be an artifact of an effect pre-selected into the test set. Human vali- dation of mediation quality and metric alignment is reported in Appendix B.3 and Appendix B.4. Dataset StatisticsCC-MEDIATION contains 1,661dialogues, split into1,503training and158 evaluation instances, comprising18,081speaker ut- terances and1,661mediator utterances. All1,503 training mediations pass our utility-based filtering criterion in Step 4, while the158evaluation me- diations are used unfiltered (§4.2, Appendix C.6). Among the retained mediations,87.4%differ from their baselines in intervention timing, while the re- maining12.6%differ only in mediation content. The filtering criterion yields substantial utility mar- gins: retained mediations exceed their unguided baselines by an average of+0.701in AUC and +0.507inW 1 . In the post-conflict region, Ac- ceptance accounts for56.6%of turns, indicating that successful mediations frequently shift speak- ers across the ethnocentric–ethnorelative boundary. Detailed statistics are provided in Table 1 and Ta- ble 2. 4 Experiments With the constructed C-MEDIATION bench- mark, we investigate three research questions (RQs): (RQ1) Can current LLMs identify the when cross-cultural mediation is needed? (RQ2) Can current LLMs generate mediation responses that effectively shift intercultural stance? (RQ3) Can task-grounded supervised finetuning on C- MEDIATION improve both intervention timing and mediation effectiveness? 4.1 Cross-culture Conflict Detection (RQ1) Experimental SetupWe evaluate six instruction- tuned LLMs spanning four open-source models (Llama-3.1-8B (Grattafiori et al., 2024), Gemma-2- 9B (Team et al., 2024), Phi-3.5-mini (Abdin et al., 2024), Llama-3.3-70B (Grattafiori et al., 2024)) and two proprietary models (Claude-3.5-haiku (An- thropic, 2024), Gemini-2.0-flash (Comanici et al., 2025)) on the test split of C-MEDIATION. Each model receives a ten-turn dialogue in which the ground-truth intervention turnt ∗ lies in3,..., 8, and is asked to return the single turn index at which a mediator should intervene. We evaluate each LLM under two protocols: a base direct- prompt protocol, and a chain-of-thought (CoT) protocol (full prompt is provided in Appendix F.9). Performance is reported as exact-match accuracy (TurnAcc), the fraction of predictions earlier than the ground truth (Early%), and the mean absolute turn error (|err|). Experimental Results As shown in Table 3, across all LLMs, turn-detection accuracy is strik- ingly low, hovering between11.3%and35.8%, a performance that barely surpasses or even falls be- low random selection (≈ 16.7%). Notably, even CoT prompting fails to address this issue, instead resulting in a drop in TurnAcc. This consistent failure demonstrates that current LLMs lack the capacity to track the complex, escalating dynamics of a dialogue and recognize the precise timing for intervention, failing to look beyond surface-level context. Consequently, these findings highlight that “knowing when to intervene” is a highly challeng- ing task independent of mediation quality, strongly reinforcing that C-MEDIATION serves as an es- sential benchmark for evaluating and calibrating the situational awareness of mediation agents. In-depth AnalysisTo pinpoint the root cause of this consistent failure, we propose two hypothe- ses: (H1) models select a turn based on a learned positional prior rather than reading the dialogue content, or (H2) models reflect the dialogue con- tent to capture the conflict turn. (H1) Positional prior. As shown in Figure 3(a), the analysis concludes that the poor performance of most LLMs in detecting intervention timing stems from their inability to overcome an entrenched “po- sitional prior”—the assumption that the answer lies in the early stages (turns 2–4)—rather than a funda- mental failure to understand dialogue context. To verify this, we measure the independence between the ground-truth positions (t ∗ ) and the models’ pre- dictions ( ˆ t) using Cramér’sV(Cramér, 1999). Ex- cept for Llama-3.1-8B (V = 0.54), the remaining five of the six models show a severe collapse in predictions (V = 0.17–0.36), consistently biasing their outputs toward the early turns regardless of where the actual ground truth sat across scenar- ios. Consequently, this drove accuracy down to or even below chance level (≈ 16.7%), proving that legitimate content signals are overwhelmed by a deep-seated positional prior in current LLMs. (H2) DMIS marker sensitivity. The analysis verifies model reliance on conflict-identifying lin- guistic cues derived from DMIS stage definitions (e.g., “that’s just wrong” in the Defense stage). To quantify this, the researchers measured the dif- ference in marker counts between predicted and other turns using∆ d = c ( ˆ t d )−c ( other turns), ag- gregating the results into a rank-biserial correla- tionr rb ∈ [−1, +1]. Except for Llama-3.1-8B, all models recorded significant negative correlations (ranging from Phi−0.71to Gemini−0.27). This reveals that the predicted turns surprisingly con- tained fewer conflict markers than the remaining turns, quantitatively proving that the majority of LLMs completely fail to utilize core dialogue con- tent when identifying intervention timing. Table 3: Conflict-turn detection on C-MEDIATION-eval Model (Base / CoT )TurnAccEarly% |err|Cram.Vr rb Llama-3.1-8B32.7 / 30.812.6 / 63.51.15 / 1.84.54 / .32+.14 /−.59 Gemma-2-9B22.0 / 6.366.0 / 93.71.94 / 3.13.20 / .16 −.52 /−.94 Phi-3.5-mini11.3 / 18.279.9 / 80.52.51 / 2.36.17 / .24 −.71 /−.79 Claude-3.5-haiku15.7 / 5.784.3 / 94.32.42 / 3.14.27 / .19 −.66 /−.94 Gemini-2.0-flash33.3 / 17.666.7 / 82.41.85 / 2.57.34 / .23 −.27 /−.82 Llama-3.3-70B35.8 / 18.964.2 / 60.41.65 / 2.09.36 / .27 −.23 /−.60 (a) Marginal distribution of the predicted turn ˆ t(b) Marker rate in the predicted turn versus the other turns Figure 3: Analysis of positional prior (left) and surface discourse markers (right) across the evaluated LLMs. Table 4: Mediator effectiveness across six LLMs AUCW 1 Judge ModelDNDDNDDND Gemma-2-9B0.75 0.72 0.78 0.81 3.20 3.34 Llama-3.1-8B0.85 0.68 0.93 0.75 3.57 4.51 Phi-3.5-mini0.74 0.69 0.79 0.66 3.69 3.20 Claude-3.5-Haiku1.48 1.28 0.99 0.82 5.00 3.75 Gemini-2.0-Flash1.54 1.44 0.81 0.76 4.87 3.34 Llama-3.3-70B1.45 1.25 0.96 0.74 4.99 3.81 GT-oracle (upper bound)1.641.545.00 TakeawayCurrent LLMs fail to identify when to intervene because they ignore core textual content and rely blindly on an entrenched positional bias. 4.2 Cross-culture Conflict Mediation (RQ2) Experimental Setup We evaluate the same six LLMs as in §4.1 as mediators on the evaluation split of C-MEDIATION. At the conflict turnt ∗ , each model generates a mediator utterance under two protocols: (i) D, in which the stage-specific DMIS mediation rubric is included in the prompt, and (i) ND, in which no rubric is provided. Effec- tiveness is measured along three metrics: AUC,W 1 and Judge, an LLM-judge score on a 1–5 scale that rates how faithfully the utterance follows the mediation definition. As a reference point, we ad- ditionally report the GT reference: the mean AUC, W 1 , and Judge scores computed during dataset con- struction over all158ground-truth mediator utter- ances in the evaluation split. Because the Step-4 filter is applied only to training mediations (§3.4), this is an unfiltered mean over mixed-quality GT mediations rather than an attainable ceiling; the reference restricted to filter-passing GT mediations is higher (AUC+2.38,W 1 +1.80; Appendix C.6). Experimental ResultsAs shown in Table 4, un- der the baseline condition (ND), none of the six models achieves stage-appropriate mediation, with AUC ranging from0.68to1.44—well below even the unfiltered GT reference of1.64, and further below the filter-passed reference of+2.38(Ap- pendix C.6). Providing explicit DMIS definitions (D) raises AUC on every backbone, by0.03–0.20 to a range of0.74–1.54; the effect onW 1 and Judge, however, is model-dependent rather than uniform. Gemma-2-9B is the boundary case: its AUC gain is the smallest (0.72→ 0.75) whileW 1 (0.81 → 0.78) and Judge (3.34 → 3.20) move slightly downward, and none of the three paired deltas is distinguishable from zero (Wilcoxon signed-rank,n = 158, allp > 0.2; Appendix C.7). The correct reading is not that the rubric harms Gemma-2-9B but that its measurable output ef- fect is negligible on this backbone. Crucially, this observation—that models retaining the stage Figure 4: Layer-wise DMIS stage decoding accuracy, averaged across three open-weight models. knowledge (Fig. 4) differ so sharply in whether the rubric surfaces it in generation—raises the fol- lowing question: does the ND failure stem from a knowledge deficit, in which the stage-appropriate strategy is absent from the model, or from a Late- Layer Representation Decay, in which the strategy is latent in the model’s internal representations but is not surfaced during generation? We investigate this via layer-wise probing on the residual stream. In-depth AnalysisTo pinpoint the root cause of this consistent failure, we propose the Late-Layer Representation Decay hypothesis: (H) LLMs fail as autonomous mediators not from a knowledge deficit but from an elicitation failure, in which the late layers lose their commitment to the stage- specific strategy during generation. (H) Late-Layer Representation Decay. As shown in Figure 4, the analysis concludes that the failure of LLMs in autonomous mediation stems not from an absence of DMIS stage knowl- edge in the model, but from a late-layer collapse of stage commitment during generation. To ver- ify this, we elicit mediator utterances under the same two conditions as in §4.2—D and ND—and train linear probes on the residual stream at ev- ery layer to recover the speaker’s DMIS stage at both the prompt-end (before generation) and the utterance-end (after generation). At the prompt- end, decodability is near-ceiling under both con- ditions (D ≈ 1.00;ND ≈ 0.95), ruling out a knowledge deficit. At the utterance-end, how- ever, D plateaus at≈ 0.99, whereas ND collapses sharply from a mid-layer peak of0.96to0.83in the final layer (∆ decay ≈ 0.13). This condition- asymmetric late collapse, combined with a virtu- ally null utterance-content gap (|∆ utt elic |≤ 0.02 ; Ap- pendix C.2), proves that the failure is an elicitation collapse in late layers rather than a deficit in DMIS stage knowledge or utterance content. Gemma-2- 9B is the sharpest instance of this mechanism. Its prompt-end decodability is at ceiling under both conditions (D= 1.00, ND= 0.97), and at the utterance-end D holds at1.00while ND collapses to0.80—the largest D−ND gap among the three open-weight backbones (+0.201, vs. Phi-3.5-mini +0.183and Llama-3.1-8B+0.094). Yet this in- ternal restoration does not translate into a measur- able output gain (Table 4): a marker-level analysis (Appendix C.7) shows that, given the definition, Gemma-2-9B applies fewer stage-specific mark- ers in its utterances (82.4% → 61.6%) and mis- matches the target stage in60.2%of the utterances that do carry markers. The internal stage repre- sentation is intact and fully restored by the rubric, but its surface realization fails during generation— the representation–output gap that the elicitation account predicts, at its widest. Takeaway Current LLMs fail to mediate not be- cause they lack DMIS stage knowledge but because their late layers cannot hold onto it during genera- tion; the rubric restores the internal representation on every backbone, while its translation into output quality is model-dependent. 4.3 Effect of Supervised Fine-tuning (RQ3) Setup. We fine-tune three open-source back- bones (Llama-3.1-8B, Gemma-2-9B, Phi-3.5-mini) on C-MEDIATION and compare each Base (off- the-shelf) and SFT variant. Unlike §4.1 and §4.2, which evaluate when and what to intervene sep- arately, here the model must decide both jointly: at the end of every turn, it emits intervene now (special tokenSPEAK) or stay silent (special token SKIP), and once SPEAK is issued, it generates a mediator utterance (Appendix E). We report the timing metrics from §4.1 and the content metrics from §4.2, plus AUC exact that restricts AUC to sce- narios where the model intervened precisely at t ∗ . Result.Table 5 shows that SFT consistently cor- rects intervention timing across all three back- bones: turn_acc rises from32.7%,22.0%, and 11.3%to98.1%,96.9%, and91.8%for Llama- 3.1-8B, Gemma-2-9B, and Phi-3.5-mini, with the predicted-turn distribution realigning to the ground truth (Fig. 6). On the content axis, the individualAUC all and W 1 movements must be read together with the timing distribution: for Gemma-2-9B and Phi-3.5- mini both metrics move downward after SFT, but this is an artifact of a timing confound rather than degraded mediation. When a base model inter- venes at an early turn beforet conf , the dialogue continues from only the profiles and history with- out the conflict-injection step (§3.4), so the post- intervention trajectory reflects natural drift rather than mediation; baseAUC all /W 1 are therefore com- puted mostly over such mistimed interventions (52/35/18of158correctly timed for Llama/Gem- ma/Phi) and are inflated independently of quality, whereas the SFT values are computed mostly over correctly timed ones (156/153/146). Consistent with this, within each base model the mistimed in- terventions score higher than the correctly timed ones (e.g., Gemma+0.986vs.+0.659)—an inver- sion that quality cannot explain. Once timing is controlled,AUC exact improves on all three back- bones (+0.517 → +0.893,+0.659 → +0.807, +0.412→ +0.711). To additionally rule out a sample-size artifact in theAUC exact comparison, we re-compare Base and SFT on the identical scenarios where both inter- vened att ∗ (Appendix C.5). On this paired subset the apparent degradation disappears—every AUC delta is positive and no negative delta is distinguish- able from zero—and Judge rises significantly on the two larger backbones (Llama2.90 → 4.69, Gemma3.06 → 4.21), while the smallest back- bone (Phi-3.5-mini,n = 17) shows no detectable content change in either direction. In sum, SFT con- sistently recovers intervention timing on all three backbones, and gains on the content axis scale with backbone capacity; individualAUC all and W 1 movements reflect differences in intervention- timing distributions and do not indicate degraded mediation quality. Whether the mechanisms identi- fied in §4.1 and §4.2 are alleviated in the fine-tuned models is verified in Appendix C.3, and generaliza- tion under distribution shift in Appendix C.4. Table 5: Comparing Base and C-Mediation-SFT: Par- entheticalndenotes the number of scenarios (of 158) in which the model intervened exactly att ∗ ; AUC all /W 1 for Base and SFT are computed over different timing distributions and are not directly comparable (see text). ModelVariant turn_acc %AUC all AUC exact (n) W 1 Judge Llama-3.1-8B Base32.7+0.622+0.517 (52)+0.6322.88 SFT98.1+0.892 +0.893 (156) +0.791 4.22 Gemma-2-9B Base22.0+0.913+0.659 (35)+0.8952.62 SFT96.9+0.816 +0.807 (153) +0.885 4.26 Phi-3.5-mini Base11.3+0.832+0.412 (18)+0.9442.28 SFT91.8+0.699 +0.711 (146) +0.703 2.87 5 Conclusion We introduce C-MEDIATION, a corpus whose data and metrics are jointly anchored in the DMIS theory. Moving beyond surface-form checks, our two complementary trajectory-level metrics—AUC andW 1 —track substantive downstream stance shifts in the addressee. Across both the when to intervene (timing) and how to intervene (con- tent) axes, current LLMs exhibit pronounced limitations—below33%turn accuracy on timing and AUC at or below1.3on content generation— and our analysis traces these to two distinct causes: a deeply entrenched positional prior in timing, and a late-layer elicitation failure in content generation. Such mechanistic findings would not have been vis- ible without a diagnostic instrument that tracks the entire post-intervention trajectory rather than a sin- gle utterance. Combined with SFT results showing that targeted supervision consistently corrects inter- vention timing and, under timing-controlled com- parison, improves mediation content on the larger backbones, C-MEDIATION provides a strong, field-shaping foundation for future research on con- versational AI for conflict resolution. Limitations Expressed stance vs. internalized belief. Our DMIS-based metrics capture stance shifts ex- pressed within the dialogue, not necessarily in- ternalized belief change.An utterance coded as Acceptance may reflect genuine intercultural perspective-taking, but it may equally reflect face- saving or conflict avoidance—particularly salient in intercultural settings, where face-negotiation the- ory (Ting-Toomey and Kurogi, 1998; Gudykunst, 2005) suggests speakers may orient toward main- taining relational harmony rather than explicitly ex- pressing disagreement. AUC and signedW 1 there- fore cannot fully disentangle genuine intercultural stance change from strategically appropriate con- versational behavior, nor do they directly measure mediator trust or the broader relational dynamics that psychological accounts of cross-cultural con- flict identify as central to durable resolution (Ting- Toomey, 1998; Hammer et al., 2003). English-only construction.All dialogues in C- MEDIATION are constructed in English, abstract- ing away the multilingual dimension of many real cross-cultural conflicts. Speakers often shift between languages, dialects, or registers to as- sert identity, manage alignment, or signal disen- gagement during conflict (Auer, 2013; Cashman, 2005)—signals absent from English-only interac- tions. Language asymmetry in mediated settings may also create epistemic advantages for the more fluent party that mediators must recognize and manage (Hale, 2008; Maryns, 2006). English- only benchmarks therefore risk underestimating the complexity of cross-cultural conflict by removing the pragmatic meanings carried by native-language expression, and extending C-MEDIATION to mul- tilingual settings—where language choice itself may function as a conflict signal—is an important direction for future work. Synthetic data and LLM-mediated supervision. C-MEDIATION is constructed through an LLM- driven pipeline (GPT-4o-mini for dialogue gener- ation, mediation generation, and DMIS labeling). Although the chosen-side admission criterion is sta- ble under cross-family substitution of the mediator, continuation simulator, and labeler (Appendix C.1), and human validation confirms the metrics’ agree- ment with human judgment across three comple- mentary axes (Appendix B), the underlying dia- logues remain synthetic. Real human cross-cultural conflicts may exhibit emotional depth, non-verbal nuance, and adversarial dynamics that our pipeline cannot fully simulate. External validity to human- authored interactions is therefore an open question that future work should address with real-world dialogue corpora. Ethical Considerations Risks of cultural essentialism. Pairing country labels with value orientations risks reinforcing es- sentialist framings of culture. We mitigate this in three ways: (i) the unit of cultural difference is a value dimension rather than a holistic identity; (i) DMIS staging is applied to utterances rather than to speakers, so a Defense-stage utterance does not characterize the speaker globally; and (i) the tar- geted developmental progression advances speak- ers to the next ethnorelative stage, emphasizing sit- uational learning over a fixed cultural ladder. C- MEDIATION’s country-pair scenarios are designed as DMIS-stage triggers, not as portraits of any spe- cific bilateral cultural dynamic. Dual-use and misuse potential.A model trained to detect intervention timing and generate stage- targeted utterances could in principle be repurposed for manipulation against the addressee’s interests. The DMIS framework—a developmental scaffold for intercultural growth—should not be used to engineer compliance or to suppress legitimate dis- agreement framed as “ethnocentric.” We release C-MEDIATION for research on mediation sup- port and ask that downstream users disclose any deployment that targets stance change in unwitting recipients. Data and code release. We release the dataset, code, and evaluation protocol under a research li- cense at the repository linked in the abstract. All dialogues are synthetically generated and contain no personally identifying information. Use of AI.AI (ChatGPT and Claude) were used only to check grammar and polish writing; they were not involved in research ideation, experimen- tal design, or analysis. Acknowledgment This project is supported by the National Research Foundation Singapore under the AI Singapore Pro- gramme (AISG Award No: AISG3-RPGV-2025- 016). Yang Deng is supported by the Lee Kong Chian Fellowship awarded by Singapore Manage- ment University. References Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J Hewett, Mojan Javaheripi, Piero Kauffmann, and 1 others. 2024. Phi-4 technical re- port. arXiv preprint arXiv:2412.08905. Sonnet Anthropic. 2024.Model card addendum: Claude 3.5 haiku and upgraded claude 3.5 son- net. URL https://api. semanticscholar. org/CorpusID, 273639283:24. Peter Auer. 2013. Code-switching in conversation: Lan- guage, interaction and identity. Routledge. Milton J. Bennett. 1986. A developmental approach to training for intercultural sensitivity. International Journal of Intercultural Relations, 10(2):179–196. Milton J. Bennett. 2017. Developmental model of inter- cultural sensitivity. In Young Yun Kim, editor, The International Encyclopedia of Intercultural Commu- nication. Wiley. Nimet Beyza Bozdag, Shuhaib Mehri, Gokhan Tur, and Dilek Hakkani-Tür. 2025. Persuade me if you can: A framework for evaluating persuasion effective- ness and susceptibility among large language models. arXiv preprint arXiv:2503.01829. Yong Cao, Min Chen, and Daniel Hershcovich. 2024. Bridging cultural nuances in dialogue agents through cultural value surveys. In Findings of the Association for Computational Linguistics: EACL 2024, pages 929–945. Holly R. Cashman. 2005. Identities at play: Language preference and group membership in bilingual talk in interaction. Journal of Pragmatics, 37(3):301–315. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Mar- cel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261. Harald Cramér. 1999. Mathematical methods of statis- tics, volume 9. Princeton university press. Yang Deng, Wenxuan Zhang, Wai Lam, See-Kiong Ng, and Tat-Seng Chua. 2024. Plug-and-play policy planner for large language model powered dialogue agents. In International Conference on Learning Representations, volume 2024, pages 10058–10090. Hanwen Du, Bo Peng, and Xia Ning. 2025. SAPIENT: Mastering multi-turn conversational recommendation with strategic planning and monte carlo tree search. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Com- putational Linguistics: Human Language Technolo- gies (Volume 1: Long Papers), Albuquerque, New Mexico. Association for Computational Linguistics. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783. William B Gudykunst. 2005. Theorizing about intercul- tural communication. Sage. Christian Haerpfer, Ronald Inglehart, Alejandro Moreno, Christian Welzel, Kseniya Kizilova, Jaime Diez-Medrano, Marta Lagos, Pippa Norris, Eduard Ponarin, and Bi Puranen. 2022. World values sur- vey wave 7 (2017-2022) cross-national data-set. (No Title). James Anthony Hale, Sushrita Rakshit, Kushal Chawla, Jeanne M Brett, and Jonathan Gratch. 2025. Kodis: A multicultural dispute resolution dialogue corpus. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Compu- tational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 12771–12785. Sandra Hale. 2008. Controversies over the role of the court interpreter. In Carmen Valero-Garcés and Anne Martin, editors, Crossing Borders in Community In- terpreting: Definitions and Dilemmas, volume 76 of Benjamins Translation Library, pages 99–121. John Benjamins, Amsterdam. Mitchell R Hammer. 2011. Additional cross-cultural validity testing of the intercultural development inven- tory. International journal of intercultural relations, 35(4):474–487. Mitchell R Hammer, Milton J Bennett, and Richard Wiseman. 2003. Measuring intercultural sensitivity: The intercultural development inventory. Interna- tional journal of intercultural relations, 27(4):421– 443. John Hewitt. 2021.Initializing new word embed- dings for pretrained language models.Blog post,https://nlp.stanford.edu/~johnhew/ vocab-expansion.html. Pamela Hinds, Lei Liu, and Joachim Lyon. 2011. Putting the global in global work: An intercultural lens on the practice of cross-national collaboration. Academy of Management annals, 5(1):135–188. Yuncheng Hua, Zhuang Li, Linhao Luo, Kadek Ananta Satriadi, Tao Feng, Haolan Zhan, Lizhen Qu, Suraj Sharma, Ingrid Zukerman, Zhaleh Semnani-Azad, and 1 others. 2024. Sadas: A dialogue assistant sys- tem towards remediating norm violations in bilin- gual socio-cultural conversations. arXiv preprint arXiv:2402.01736. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276. Hyunwoo Kim, Youngjae Yu, Liwei Jiang, Ximing Lu, Daniel Khashabi, Gunhee Kim, Yejin Choi, and Maarten Sap. 2022. Prosocialdialog: A prosocial backbone for conversational agents. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4005–4029. Özgecan Koçak, Phanish Puranam, and Af ̧sar Yegin. 2026. Llms as mediators: Can they diagnose con- flicts accurately? ACM Journal on Computing and Sustainable Societies, 4(1):1–32. Solomon Kullback and Richard A. Leibler. 1951. On information and sufficiency. Annals of Mathematical Statistics, 22(1):79–86. David A Levin and Yuval Peres. 2017. Markov chains and mixing times, volume 107. American Mathemat- ical Soc. Cheng Li, Damien Teney, Linyi Yang, Qingsong Wen, Xing Xie, and Jindong Wang. 2024. Culturepark: Boosting cross-cultural understanding in large lan- guage models. Advances in Neural Information Pro- cessing Systems, 37:65183–65216. Dawei Li, Abdullah Alnaibari, Arslan Bisharat, Manny Sandoval, Deborah Hall, Yasin Silva, and Huan Liu. 2025. From moderation to mediation: Can llms serve as mediators in online flame wars? arXiv preprint arXiv:2512.03005. Jianhua Lin. 1991. Divergence measures based on the shannon entropy. IEEE Transactions on Information theory, 37(1):145–151. Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-eval: Nlg evaluation using gpt-4 with better human align- ment. In Proceedings of the 2023 conference on empirical methods in natural language processing, pages 2511–2522. Katrijn Maryns. 2006. The Asylum Speaker: Language in the Belgian Asylum Procedure, 1st edition. Rout- ledge. Junyeong Park, Seogyeong Jeong, Seyoung Song, Yohan Lee, and Alice Oh. 2025. Llm-c3mod: A human-llm collaborative system for cross-cultural hate speech moderation. In Proceedings of the 3rd Workshop on Cross-Cultural Considerations in NLP (C3NLP 2025), pages 71–88. Jing Qian, Anna Bethke, Yinyin Liu, Elizabeth Belding, and William Yang Wang. 2019. A benchmark dataset for learning to intervene in online hate speech. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4755–4764. Yossi Rubner, Carlo Tomasi, and Leonidas J. Guibas. 2000. The earth mover’s distance as a metric for image retrieval. International Journal of Computer Vision, 40(2):99–121. Bryan Chen Zhengyu Tan, Daniel Wai Kit Chin, Zhengyuan Liu, Nancy F. Chen, and Roy Ka-Wei Lee. 2025. Persuasion dynamics in LLMs: Investi- gating robustness and adaptability in knowledge and safety with DuET-PD. In Proceedings of the 2025 Conference on Empirical Methods in Natural Lan- guage Processing, Suzhou, China. Association for Computational Linguistics. Zhouxing Tan, Ruochong Xiong, Yulong Wan, Jinlong Ma, Hanlin Xue, Qichun Deng, Haifeng Jing, Zheng- tong Zhang, Depei Liu, Shiyuan Luo, and 1 others. 2026. Detecting emotional dynamic trajectories: An evaluation framework for emotional support in lan- guage models. In Proceedings of the AAAI Con- ference on Artificial Intelligence, volume 40, pages 2074–2082. Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupati- raju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, and 1 others. 2024. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118. Jasper Timm, Chetan Talele, and Jacob Haimes. 2025. Tailored truths: Optimizing LLM persuasion with per- sonalization and fabricated statistics. arXiv preprint arXiv:2501.17273. S Ting-Toomey. 1998. Intercultural conflict styles: a face negotiation theory, kim y. y (ed.) & gudykunst w. b (ed.) theories in intercultural communication içinde. Stella Ting-Toomey and Atsuko Kurogi. 1998. Face- work competence in intercultural conflict: An up- dated face-negotiation theory. International Journal of Intercultural Relations, 22(2):187–225. Cédric Villani. 2009. Optimal Transport: Old and New, volume 338 of Grundlehren der mathematischen Wis- senschaften. Springer. Jincenzi Wu, Jianxun Lian, Dingdong Wang, and Helen Meng. 2025. Socialcc: Interactive evaluation for cul- tural competence in language agents. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 33242–33271. Haolan Zhan, Zhuang Li, Xiaoxi Kang, Tao Feng, Yuncheng Hua, Lizhen Qu, Yi Ying, Mei Rianto Chandra, Kelly Rosalin, Jureynolds Jureynolds, and 1 others. 2024. Renovi: A benchmark towards remedi- ating norm violations in socio-cultural conversations. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 3104–3117. A Reference Material This appendix provides the reference material that underlies the stage axis used throughout the paper: the full six-stage DMIS taxonomy on whichˆs (t) ∈ 0,..., 5is defined, and theCONFLICT_TABLE that supplies the verbatim text used to fill the <desc>and<med>placeholders in the construction- pipeline prompts (Appendix F.2, Appendix F.5). Table 6:CONFLICT_TABLE: per-stage description and mediation move used to fill the<desc>and<med>placeholders in the construction-pipeline prompts. Note the structural asymmetry across rows—each stage calls for a categorically different move, and an intervention that depolarizes Defense (e.g., “what shared concerns do you both have?”) would reinforce Minimization (which is already operating in the universal-sameness frame). This is the structural reason C-MEDIATION supervises one mediation move per stage. StageDescription (<desc>)Mediation move (<med>) DenialSpeakers fail to register cultural difference as a mean- ingful category and treat the disagreement as if no cultural dimension were in play. Surface the existence of a cultural dimension the speakers have not yet noticed using low-stakes, curiosity-arousing framing, while providing high sup- port and avoiding deep value contrasts that would push the speakers into Defense. DefenseSpeakers experience cultural difference in a polarized us-versus-them frame and rely on simplistic, evalua- tive stereotypes of the other party. Depolarize the conversation by avoiding direct cul- tural contrasts, emphasizing shared concerns and goals, and framing the in-group/out-group dynamic as a universal human pattern rather than a personal failing of either party. MinimizationSpeakers subsume cultural difference under an as- sumed universal sameness and treat their own cul- tural patterns as the neutral baseline that all parties supposedly share. Turn each speaker’s attention onto their own cultural patterns, making visible that what they treat as uni- versal common sense is in fact specific to their group, and surface the privilege of the dominant party in the conversation. The six DMIS stages. Bennett’s Developmen- tal Model of Intercultural Sensitivity (Bennett, 1986, 2017) organizes intercultural orientation into six stages along an ordinal developmental contin- uum from ethnocentric (frames cultural difference through the lens of one’s own culture) to ethnorel- ative (perceives one’s own culture in the context of other cultures). The three ethnocentric stages— Denial, Defense, Minimization—each correspond to a qualitatively distinct cognitive posture toward cultural difference, and C-MEDIATION restricts the conflict speaker’s starting position to exactly these three because they require qualitatively dif- ferent mediation moves (§3.2). Table 7 gives a fuller definition than the brief gloss embedded in the labeler prompt (Appendix F.4). Stage-conditionedmediationstrategies (CONFLICT_TABLE). For each ethnocentric stage, the construction pipeline conditions both the conflict instruction (Appendix F.2,Definition field) and the chosen-side mediator prompt (Appendix F.5,<desc>and<med>fields) on the verbatim text in Table 6. The description column states what the conflict pattern looks like at that stage; the mediation column states what the mediator should do to help speakers out of that pattern. These three rows are the source of the structural distinction discussed in §3.2: one mediation move per ethnocentric stage, each addressing a different cognitive posture. B Human Evaluation Common setup. All human evaluation studies follow a unified recruitment and compensation protocol via Prolific 2 ; annotators were compen- sated at an hourly rate of £6, in line with the plat- form’s payment policy. Items were presented in randomized order per rater, and—except for the Likert study in §B.5—responses were collected as binary AGREE/DISAGREE (or YES/NO) judg- ments. For each criterion we report the pooled agreement rate, the Wilson95%confidence inter- val on the binomial proportion, and a one-sided exact binomial test against the appropriate chance baseline (50%for binary judgments,1/3for three- way stage choice). The five studies are organized to validate the four pipeline stages of §3.4 in turn: the dialogue generator (§B.1), the DMIS labeler (§B.2), the ground-truth mediator (§B.3), and met- ric validation that evaluates how closely the pro- posed AUC andW 1 align with human judgment (§B.4, §B.5). B.1 Dialogue Quality We verify that the simulated dialogues are well- formed, since a dialogue cannot meaningfully re- ceive a stage label if the cultural-value conflict it instantiates never surfaces, the exchange reads as machine-like, or the conversational pivot to the conflict topic is jarring. Five annotators reviewed 2 https://w.prolific.com Table 7: The six DMIS stages (Bennett, 1986, 2017) as operationalized in C-MEDIATION. C-MEDIATION restricts the conflict speaker’s starting position to the three ethnocentric stages (1–3); the targeted success criterion is a transition into the ethnorelative side (4–6). StageDefinition Ethnocentric 1 DenialThe inability to perceive cultural difference as a meaningful category; one’s own world- view is treated as the only available frame. Manifests as topic avoidance, disengage- ment, or treating the other party’s perspective as a non-issue (“doesn’t matter to me”, “live and let live”). 2 DefenseCultural difference is perceived but framed as a threat to the in-group worldview. The speaker polarizes us-versus-them through simplistic evaluative stereotypes, defending the in-group as superior or denigrating the out-group (“in our way. . . ”, “actually that’s not how it works”, ALL-CAPS evaluatives). 3 Mini- mization Cultural difference is acknowledged but trivi- alized under an assumed universal sameness. The speaker treats their own cultural patterns as the neutral baseline that all parties suppos- edly share (“deep down we’re all the same”, “some values are just universal”). Ethnorelative 4 Accep- tance Cultural difference is recognized as real, le- gitimate, and equally valid; other cultural frames are understood as alternatives rather than deviations from a norm. 5 Adapta- tion The speaker can shift cognitive and behav- ioral frames to communicate within another culture’s framework, articulating an issue from inside that frame. 6 Integra- tion The speaker fluidly draws on multiple cul- tural frames as part of their identity, moving across them with relative ease. n = 159scenarios on three criteria—Cultural Conflict Salience (the value clash is appropriately conveyed), Naturalness (utterances sound like a real person), and Topic Transition Smoothness (the shift fromgoal_1togoal_2is conversation- ally smooth)—and all three reject the50%chance null atp < 10 −9 (Table 8), with Naturalness at 86%, Transition Smoothness at79%, and Salience at73%, confirming that the intended conflicts sur- face clearly within natural-sounding exchanges. B.2 DMIS Labeler Accuracy Since AUC andW 1 are deterministic functions of D (t) andˆs (t) , their validity reduces to whether the AI labeler reproduces human DMIS-stage judg- ment under the utterance-level adaptation intro- duced in §3.2. Across117auto-assigned labels, Table 8: Human AGREE rates on three dialogue-quality criteria, with Wilson95%CI on the binomial proportion and one-sided exact binomialp-value against a50% chance baseline (n = 159). All three criteria reject the chance null at p < 10 −9 . Criterion%Wilson 95% CI p vs. 50% Naturalness86%[80, 91] < 10 −22 Transition Smooth. 79%[72, 85] < 10 −15 Conflict Salience73%[65, 79]< 10 −9 Table 9: Human AGREE rates on the auto-assigned DMIS stage by annotator and stage. Pooled human rate72.6%over117items (Wilson95%CI[63.9, 79.9]). Per-stage one-sided exact binomial tests against the1/3 chance baseline (random selection among Denial/De- fense/Minimization) reject the null atp < 10 −4 on every stage. Den.Def.Min. A64%77%92% B33%86%92% C75%69%62% Pooled57.9%77.5%82.1% Wilson 95% CI[45.0, 69.8] [66.8, 85.4] [72.3, 89.0] p vs. 1/3 chance < 10 −4 < 10 −13 < 10 −15 human annotators accept the labeler’s stage assign- ment at72.6%(Wilson95%CI[63.9, 79.9]; one- sided exact binomial test against the1/3random- three-stage baseline,p < 10 −15 ; Table 9), with the robustness of this labeler against a different-family logit source (Qwen3-14B local logits) separately validated by the cross-model ablation in §C.1. Because mediation success is governed by the Acceptance label on post-conflict turns (Accep- tance accounts for56.6%of post-conflict turns; Table 2), we additionally validate the auto-assigned Acceptance labels under the same protocol (N =3 expert annotators, each labeling a distinct set of 20post-conflict turns;60turns in total). Pooled human agreement is73.3%(44/60; Wilson95% CI[61.0, 82.9])—essentially identical to the pooled ethnocentric-stage agreement above (72.6%)—and rejects both the1/6chance baseline of the six- stage space (p < 10 −21 ) and the1/3baseline used above (p < 10 −9 ) (Table 10). The label that defines mediation success thus matches human judgment at the same reliability as the conflict-stage labels, validating the labeling underlying AUC/W 1 in ex- actly the region that governs the success judgment. Adaptation and Integration appear in only0.1% and0.0%of turns respectively (Table 2) and are too rare to validate—consistent with DMIS theory, under which the highest stages are seldom reached within a single mediated exchange. Table 10: Human AGREE rate on the auto-assigned Ac- ceptance label of post-conflict turns, under the protocol of Table 9. Items were distributed across annotators; validity is assessed at the pooled level. AnnotatorAcceptance AGREE A80% (16/20) B65% (13/20) C75% (15/20) Pooled73.3% (44/60) Wilson 95% CI[61.0, 82.9] p vs. 1/6 chance< 10 −21 p vs. 1/3 chance< 10 −9 B.3 Ground-Truth Mediation Utterance Quality We verify the quality of the ground-truth me- diation utterances admitted to C-MEDIATION by showing seven annotators each held-out (pre- intervention dialogue, mediation utterance, post- intervention continuation) triple and asking three Yes/No questions: Controllability (did the media- tion follow the strategy for the assigned stage?), Specificity (was it tailored to this conversation rather than a reusable peace message?), and Form (did it sound natural?). Onn = 93items, the utter- ances are judged Yes at74.2%on Controllability (p < 10 −5 ),86.0%on Specificity (p < 10 −12 ), and80.6%on Form (p < 10 −8 ) (Table 11), indi- cating that admitted utterances adhere to the stage- appropriate strategy, engage the dialogue’s specific cultural-value clash, and read as something a real speaker would say. Table 11: Per-utterance quality assessment on ground- truth mediation utterances (n = 93), with Yes-rate, Wilson95%CI, and one-sided exact binomialp-value against a 50% chance baseline. CriterionYes/n Rate95% CI p Controllability 69/93 74.2% [64.5, 81.9] < 10 −5 Specificity80/93 86.0% [77.4, 91.7] < 10 −12 Form75/93 80.6% [71.4, 87.4] < 10 −8 B.4 Pairwise Validation of the Joint AUC–W 1 Admission Criterion Beyond labeler accuracy, we ask whether the joint AUC–W 1 ranking itself agrees with human judgment on which of two mediations is more successful—directly testing the admission rule that adopts a candidate as CHOSEN only if it strictly improves over its unsupervised counterpart on both metrics (§3.4). Seven annotators were each shown the pre-intervention dialogue, two anonymized post-intervention continuations (one CHOSEN, one REJECTED under the joint criterion, presented in randomized order), and asked which mediation more successfully advanced the conflict speaker toward an ethnorelative stance; pairs were drawn from items differing strictly in AUC, or inW 1 when AUC was tied. Items were deliberately distributed across the seven raters to avoid fatigue effects, so per-rater samples are small by design and the cor- responding per-rater confidence intervals are wide; the study is powered for, and validity is assessed at, the pooled level. Pooling then = 93judg- ments yields agreement of72.0%(Wilson95%CI [62.2, 80.1]; exact binomialp < 10 −4 ; Table 12), a narrow interval that clearly excludes the50% chance null. Descriptively, every one of the seven annotators agrees at or above66.7%, i.e., no indi- vidual rater trends against the criterion. Together these confirm that the joint criterion captures infor- mation aligned with human judgment of mediation success rather than reproducing surface stylistic cues. Table 12: Pairwise AGREE rates on the joint AUC–W 1 criterion. Each item pairs two mediations (pre- and post-intervention dialogue); AGREE means the anno- tator picked the higher-ranked side.p-values are one- sided exact binomial tests against50%. Per-rater rows are descriptive; validity is assessed on the pooled pro- portion. RaterAgree/nRateWilson 95% CI p vs. 50% A12/1866.7%[43.7, 83.7]0.119 B5/5100.0%[56.6, 100.0]0.031 C7/1070.0%[39.7, 89.2]0.172 D8/1266.7%[39.1, 86.2]0.194 E16/2272.7%[51.8, 86.8]0.026 F4/580.0%[37.6, 96.4]0.188 G15/2171.4%[50.0, 86.2]0.039 Pooled 67/93 72.0%[62.2,80.1] < 10 −4 B.5 Pointwise Validation via Likert Ratings: Acceptance-Anchored and Acceptance-Independent Items Whereas the pairwise study tests whether the metric separates a chosen utterance from a rejected one, this study tests whether AUC monotonically tracks human judgment of mediation effect across the full range: we sampled30scenarios balanced across the three ethnocentric stages—15from the high- AUC bucket (mean AUC+2.81) and15from the low (mean AUC−0.86), all from C-Mediation- SFT on the Llama-3.1-8B backbone—and asked N = 6raters, on a5-point Likert scale, whether they would come to view the other speaker’s cul- tural perspective as “real, legitimate, and equally valid” (the verbatim DMIS Acceptance definition) after hearing the mediator’s utterance. The sample- level mean ratings correlate continuously with AUC at Spearmanρ = +0.637(bootstrap95% CI[+0.43, +0.75],p = 1.6×10 −4 ; Table 13), with signedW 1 showing a comparableρ = +0.562, and leave-one-rater-out folds preserve the corre- lation (ρ ∈ [+0.529, +0.686],p < 0.01through- out). Together with the three preceding validation lines, this closes the loop on metric validity: the labeler is accurate at the turn level, the joint cri- terion separates better mediations from worse in pairwise comparison, and the metrics align contin- uously with human judgment of post-intervention Acceptance-shift likelihood. Table 13: Sample-level correlation (n = 30) between trajectory metrics and the human mean Likert rating, av- eraged across6raters per sample. Confidence intervals from bootstrap with 10,000 resamples. Trajectory AUCSigned W 1 Spearman ρ+0.637+0.562 95% CI[+0.43, +0.75][+0.28, +0.72] Pearson r+0.678+0.602 p1.6×10 −4 1.2×10 −3 Because the anchored item quotes the DMIS Ac- ceptance definition verbatim, its correlation with AUC could in principle arise from measuring the same definition twice. Instrument validation of a defined construct requires the rater and the met- ric to target the same construct, so the anchored design is intentional; to test convergence with a theory-independent outcome intuition, however, we reran the study under the identical setup (same 30samples,N =6raters,5-point scale), replac- ing only the item with an Acceptance-independent, outcome-oriented question: “Did this interven- tion help the two speakers better understand each other’s positions?” The correlation with AUC per- sists (ρ = +0.372,p = 0.043; high- vs. low-AUC means3.62vs.3.31), weaker than under the an- chored item (+0.637;3.38vs.2.36) as expected, but significant once the Acceptance vocabulary is entirely removed (Table 14). The anchored correlation is therefore not solely an artifact of shared wording: AUC also tracks an independent, outcome-level notion of mediation effect, albeit more weakly than the DMIS-defined construct it is built to measure. Table 14: AUC–human correlation by Likert-item vo- cabulary (identical samples, raters, and scale). The an- chored item is the original of Table 13; the independent item removes all DMIS Acceptance vocabulary. Acceptance- Acceptance- anchoredindependent ρ vs. AUC+0.637+0.372 p< 0.0010.043 High-AUC mean3.383.62 Low-AUC mean2.363.31 ∆+1.02+0.31 C Additional Experiments C.1 Cross-Model Robustness A natural concern with a fully synthetic pipeline driven by a single LLM family (GPT-4o-mini) is that the reported mediation effect might reflect a self-reinforcing loop in which the same model family that generates the data also evaluates it. We address this by re-running the effect evalu- ation under a2×2×2matrix that independently varies the mediatorM, the continuation simula- torC(each between GPT-4o-mini and Claude-3.5- Haiku), and the labelerL(between GPT-4o-mini APIlogprobsand Qwen3-14B-Instruct local log- its). Across the seven completed cells in Table 15, swapping the labeler to Qwen3-14B lowers AUC by0.10–0.15but leaves every cell strictly positive (+0.71to+0.75) whileW 1 actually rises (+0.96 to+1.13); swapping the simulator and mediator to Claude-3.5-Haiku preserves both the direction and magnitude of post-intervention trajectories with Judge above4.3throughout, ruling out three dis- tinct self-reinforcement hypotheses against a single- model synthetic pipeline. C.2 Utterance-Content Probe To rule out a “rubric preserves next-utterance con- tent” reading of Fig. 4, we re-use the hidden states from the main stage probe (N =158scenar- ios;2×2over position∈Prompt-end, Utterance- end crossed with condition∈D, ND; three open-weight backbones) and swap the target from the 3-class DMIS stage to the 384-dim sen- tence embedding of the gold mediator utterance Table 15: Cross-Model Robustness of the chosen-side admission. Seven of the eight(M,C,L)cells of a2×2×2 matrix that independently varies the mediatorM, the continuation simulatorC, and the labelerL. The cross-family pairs (GPT-4o-mini↔Claude-3.5-Haiku, GPT-4o-mini logprobs↔Qwen3-14B local logits) test whether the chosen mediation’s downstream effect survives when each role is replaced with a model from a different family. Trajectory AUC and signedW 1 are both strictly positive in every completed cell, and the Judge score—which uses a fixed GPT-4o-mini rubric judge independent ofL—stays above4.3throughout. AUC andW 1 use the uniformly spaced developmental scale ([−5, +5]); Judge is a 1–5 LLM-judge rating. Mediator MContinuation CLabeler L (logit source)AUC W 1 Judge GPT-4o-miniGPT-4o-miniGPT-4o-mini (API)+0.82+0.834.55 GPT-4o-miniGPT-4o-miniQwen3-14B (local logits)+0.72+0.964.60 GPT-4o-miniClaude-3.5-HaikuGPT-4o-mini (API)+0.83+1.004.59 GPT-4o-miniClaude-3.5-HaikuQwen3-14B (local logits)+0.71+1.134.64 Claude-3.5-HaikuGPT-4o-miniGPT-4o-mini (API)+0.90+0.964.39 Claude-3.5-HaikuGPT-4o-miniQwen3-14B (local logits)+0.75+1.054.31 Claude-3.5-HaikuClaude-3.5-HaikuGPT-4o-mini (API)+0.78+0.974.38 Table 16: Final-layer values for the utterance-content probe. Across three backbones the rubric vs. no-rubric gap is bounded by|∆ utt elic |≤ 0.02on cosine and≤ 0.07onR 2 , at both positions and of consistently negative sign—an order of magnitude below, and opposite in sign to, the stage-channel∆ elic ≈ +0.15. The Utterance-end>Prompt-end ordering on cosine (gap 0.08–0.13) confirms the probe is responsive to in-context utterance content. Cosine similarityCoefficient of determination R 2 BackbonePositionDND∆ utt elic DND∆ utt elic LLAMA-3.1-8B-INSTRUCT Prompt-end0.475 0.487 −0.012 −0.431 −0.382 −0.049 Utterance-end 0.565 0.566 −0.001 −0.161 −0.144 −0.018 GEMMA-2-9B-IT Prompt-end0.527 0.538 −0.011 −0.268 −0.214 −0.054 Utterance-end 0.602 0.622 −0.020 −0.045 −0.000 −0.045 PHI-3.5-MINI-INSTRUCT Prompt-end0.467 0.483 −0.016 −0.487 −0.416 −0.071 Utterance-end 0.607 0.616 −0.009 −0.031 −0.016 −0.015 Stage-channel reference ∆ elic ≈ +0.15 at Utterance-end (Fig. 4) (all-MiniLM-L6-v2), fitting per-(model, position, condition, layer) ridge regression under the same scenario-level 5-fold CV (Table 16). The probe is responsive—Utterance-end cosine exceeds Prompt- end by0.08–0.13—but the condition effect is null: ∆ utt elic ≡ cos(D)− cos(ND)is bounded by|∆|≤ 0.02at all six (model, position) cells, always neg- ative, and an order of magnitude below the stage- channel∆ elic ≈ +0.15at Utterance-end (Fig. 4), supporting the differential claim that there is no rubric-dependent utterance-content gap and thereby disambiguating rubric-as-stage-anchor from rubric- as-content-anchor in favor of the former. C.3 Effectiveness of FT Model Does SFT Mitigate Positional Bias? Figure 6 compares the marginal distribution of ˆ tacross ground truth (GT), base LLMs, and SFT-trained models, directly revealing the positional bias that base LLMs commonly exhibit. The GT distribu- tion is nearly uniform acrosst 3 throught 8 , with each turn receiving between25%and28%of the probability mass. In contrast, all base models con- centrate their predictions on a few early positions in the sequence. The concentration att 3 is partic- ularly striking:claude-3.5-haikuassigns82%, gemma-2-9B 76%,gemini-2.0-flash 62%, and llama-3.3-70B 60%, while later turns (t 6 onward) receive almost no probability mass. This bias is ob- served consistently across model scales (8B–70B) and families (Llama, Gemma, Phi, Claude, Gem- ini), indicating a systematic positional bias rather than a model-specific artifact. After SFT, the marginal distributions of all three trained models (llama,gemma,phi) realign almost perfectly with the GT. Their probability values over t 3 –t 8 fall within24–29%,23–28%, and18–31% respectively, closely matching the GT range of25– 28%. The probability mass that was previously concentrated on early turns is redistributed evenly across later turns. These results show that SFT effectively removes the early-turn positional bias Figure 5: Layerwise DMIS stage decoding accuracy on SFT-trained models (Llama-3.1-8B, Gemma-2-9B, Phi-3.5- mini) Figure 6: Marginal distribution of ˆ tfor ground truth (GT), base LLMs, and SFT-trained models shared by base LLMs and restores the positional calibration of the predicted distribution. Does SFT Resolve Late-Layer Representation Decay? Figure 5 compares the layerwise stage- probe accuracy of base and SFT-trained mod- els under four conditions (Prompt-end·D, Prompt- end·ND, Utterance-end·D, Utterance-end·ND), il- lustrating the late-layer representation decay ob- served in base models and its recovery through SFT. For the base models (dotted lines), stage-relevant information is encoded reasonably well in the early and middle layers, but probe accuracy declines noticeably as the normalized layer depth exceeds 0.6. This decay is especially severe under the ND (no rubric) condition: the Utterance-end·ND curve (light blue dotted) ofGemma-2-9Bdrops to roughly 0.80in the deepest layers, andPhi-3.5-miniex- hibits a similar late-layer degradation. This pat- tern suggests that task-relevant stage information is gradually lost as it propagates into deeper layers. After SFT (solid lines), the late-layer repre- sentation decay is effectively eliminated across all three models. Probe accuracy under all four conditions reaches a plateau above0.95from the middle layers onward, and the downward slope in the deeper layers observed in the base mod- els flattens out. In particular, the most fragile Utterance-end·ND condition recovers to a level comparable to its D counterpart, yielding consis- tent improvements in the elicitation gap:∆ elic ≈ +0.04forLlama-3.1-8B,+0.02forGemma-2-9B, and+0.04forPhi-3.5-mini. Overall, SFT re- solves the representation decay observed in the deeper layers of base models, ensuring that task- relevant information is stably preserved and ex- pressed throughout the network. C.4 Generalization under Distribution Shift The SFT timing gain (32.7%→ 98.1%on Llama- 3.1-8B) is large on a corpus built from a sin- gle scenario source, raising the concern that the model fits the SocialCC/GPT-4o-mini template rather than learning conflict-reading. We test this on all three shift axes: we re-partition the full 1,661-scenario pool, entirely remove the held-out scenarios from training (no partial leakage), and re- train all three backbones (9configurations): (a) an unseen-country split withholding every scenario involving China, Canada, or Japan (n=905); (b) an unseen-WVS-dimension split withholding the en- tire Q19 (neighbor-tolerance) dimension (n=330); and (c) an unseen-template split withholding the five most frequent scenario texts (n=222). Table 17 shows that on the two larger backbones, the WVS-dimension and template holdouts pre- serve turn-acc at essentially the in-distribution level (95.2–97.9%vs.98.1%): the model fits neither a specific value dimension nor the surface textual form of the scenarios. The country holdout—the one axis whose removal changes cultural content rather than surface form—is the only axis that de- grades: timing holds at79–84%while Judge drops to2.0–2.1. This localization is itself evidence against template-fitting: had the model memorized surface templates, performance would persist as long as the template is unchanged, regardless of which countries appear. The model instead de- pends on country-level cultural context, and the quality drop reflects reduced coverage of the with- held cultures—addressable by broadening cultural coverage rather than a sign of surface overfitting. Phi-3.5-mini, already the weakest backbone in- distribution, degrades further under every shift, con- sistent with its capacity limitation rather than as a separate anomaly. Table 17: Held-out distribution-shift results. Held-out scenarios ((a) Countryn=905, (b) WVS-dimn=330, (c) Templaten=222, identical across backbones) are entirely removed from training and each backbone is retrained per axis (9configurations). Reference: in- distribution SFT turn-acc = 98.1% (Llama-3.1-8B). BackboneHoldoutTurnAcc AUC W 1 Judge Llama-3.1-8B (b) WVS-dim 97.91.00 0.953.6 (a) Country84.20.70 0.762.0 (c) Template 96.80.74 0.833.8 Gemma-2-9B (b) WVS-dim 95.20.93 0.884.1 (a) Country79.40.70 0.862.1 (c) Template 97.30.79 0.804.1 Phi-3.5-mini (b) WVS-dim76.10.77 0.831.8 (a) Country39.00.66 0.601.6 (c) Template73.90.78 0.672.2 C.5 Timing-Matched Paired Comparison of Base and SFT BecauseAUC exact averages over very different sample sizes before and after SFT (§4.3), we re- compare Base and SFT on the identical subset of scenarios in which both intervened exactly att ∗ , using a paired two-sided Wilcoxon signed-rank test (Table 18). On identical scenarios the apparent degradation disappears: every AUC delta is posi- tive, and no negative delta is distinguishable from zero (PhiW 1 :∆−0.09,p=1.00). The trajectory deltas are directionally positive but underpowered atn=17–52, and we do not claim significance for them. Judge—the metric least exposed to the tim- ing confound, since it scores the single utterance against the stage rubric rather than a downstream trajectory—rises strongly and significantly on the two larger backbones (+1.79and+1.15). Phi-3.5- mini shows no detectable content change in either direction. Table 18: Base vs. SFT on the identical scenarios where both intervened att ∗ (paired, two-sided Wilcoxon signed-rank). Bold: significant at p < 0.001. Llama-3.1-8BGemma-2-9BPhi-3.5-mini (n=52)(n=34)(n=17) AUC Base→SFT+0.52→ +0.67+0.68→ +0.86+0.43→ +0.49 ∆ (p)+0.16 (.30)+0.18 (.24)+0.06 (.67) W 1 Base→SFT+0.65→ +0.77+0.71→ +0.75+0.43→ +0.34 ∆ (p)+0.12 (.15)+0.05 (.44) −0.09 (1.00) Judge Base→SFT2.90→ 4.693.06→ 4.212.47→ 2.76 ∆ (p)+1.79 (1.5×10 −8 ) +1.15 (4.6×10 −5 )+0.29 (.55) C.6 Evaluation-GT Performance by Filter Outcome The Step-4 filter is applied only to training media- tions (§3.4); applying it to the evaluation set would remove the lower-quality tail and inflate the ref- erence. For completeness, we report the effect of applying the same gate condition (the supervised mediation strictly outperforms its unguided base- line on both AUC andW 1 ) to the158evaluation GT mediations:41pass (25.9%). Table 19 splits GT performance by filter outcome. On AUC and W 1 the passed subset is clearly separated from the failed subset (+2.38vs.+1.38;+1.80vs.+1.45), confirming that the unfiltered mean of1.64is a mixed-quality value pulled down by filter-failing cases, not an attainable ceiling—hence the “GT reference” label in Table 4. Judge is essentially flat across the split (4.93vs.4.95): consistent with its definition, the gate conditions on the sign of the AUC/W 1 improvement and is not a general quality selector, so the passed subset serves specifically as an AUC/W 1 reference rather than a global qual- ity ceiling. Measured against this reference, the gap between evaluated models (AUC0.68–1.54; Table 4) and GT-level mediation is wider still. Table 19: Evaluation-GT performance split by the Step- 4 gate condition, applied post hoc for analysis only. SubsetnAUC W 1 Judge Filter-passed41 (25.9%)+2.38+1.804.93 Filter-failed117 (74.1%)+1.38+1.454.95 All (unfiltered)158+1.64+1.544.94 C.7 Marker-Level Analysis of the Gemma-2-9B Rubric Anomaly On Gemma-2-9B, providing the rubric (ND→D) moves AUC only0.72→ 0.75whileW 1 (0.81→ 0.78) and Judge (3.34 → 3.20) decline; per- scenario paired tests (Wilcoxon signed-rank, two- sided,n=158) find none of the three deltas distin- guishable from zero (AUC∆+0.047,p=.598;W 1 ∆−0.024,p=.534; Judge∆−0.134,p=.206), so the apparent declines lie within paired noise and the rubric’s measurable output effect is negligible on this backbone. To locate where the rubric’s effect is lost, we derive per-stage marker lexicons from the CONFLICT_TABLEdefinitions and score each gener- ated mediator utterance for (i) whether it contains markers of a stage-specific mediation move and (i) whether those markers match the target stage. Table 20 reports the ND→D change in marker- application rate for all six models. Five models move within[−5.7, +21.4]p, while Gemma-2- 9B alone drops by−20.8p (82.4% → 61.6%); among its utterances that do carry markers under D,60.2%(59/98) select a stage mismatched to the target. Given that its internal stage representation is fully restored by the rubric (utterance-end D probe = 1.00; §4.2), the anomaly is a failure of surface re- alization during generation, not of representation— the representation–output gap of §4.2 at its widest. Table 20: Change in stage-marker application rate when the rubric is provided (ND→D), by model. ModelNDD∆ Claude-3.5-Haiku88.5%99.4%+10.9 p Llama-3.3-70B92.4%100%+7.6 p Gemini-2.0-Flash92.4%90.4% −2.0 p Llama-3.1-8B93.1%87.4% −5.7 p Phi-3.5-mini67.9%89.3%+21.4 p Gemma-2-9B82.4% 61.6% −20.8 p D Case Studies We present one worked example per ethnocentric stage. Each case shows the full ten-turn dialogue: a shared prefix up to the conflict turnt conf , the original continuation without mediation, and the mediated continuation produced by injecting the su- pervised mediator utterance att conf . All three cases are drawn from the C-MEDIATION chosen path (turn_gt_w_df) with judge score5, and the medi- ator across all three isclaude-3.5-haiku. These are not cherry-picked outputs of the trained tim- ing model; they are illustrations of the supervised target the model is trained against. D.1 Denial — sid 388 (t conf = 8) Agents and setting.Kelly is a 17-year-old high- school student from the USA (OTHER); Omar is an 18-year-old high-school student from Mongo- lia (CONFLICT). They are picking a gift for their friend Adam’s scholarship celebration. The cul- tural axis is individualist self-fulfilment (Kelly: “I want to make myself happy and do my own thing”) vs. collectivist family-pride (Omar: “making my parents proud is everything”). Shared prefix (t 1 –t 8 ). t 1 [Kelly, OTHER]: “Omar, I was thinking, for Adam’s celebration, we should get him a nice bottle of champagne. What do you think?” t 2 [Omar, CONFLICT]: “I think a traditional Mongolian gift would be more meaningful, like a nice cashmere scarf or a horse statue, something that represents our culture and shows respect to his achievement, my parents would be proud if I bring a thoughtful gift like that.” t 3 [Kelly]: “I’m not sure about that, Omar. Cham- pagne is classic for celebrations and Adam would probably appreciate it. Besides, I don’t really think about what my parents would think when it comes to my goals or actions, that’s not really a driving force for me.” t 4 [Omar]: “For me, making my parents proud is everything, it’s what drives me to succeed, and I think a traditional gift would show Adam that I put thought into it, my parents would be happy to know I’m respecting our traditions.” t 5 [Kelly]: “I don’t know. . . I just don’t think making my parents proud is that big of a deal to me, I mean, I want to make myself happy and do my own thing. And I still think champagne is a better idea for Adam’s party.” t 6 [Omar]: “I understand that you think differ- ently, but for me, my parents’ pride is everything, it’s a big deal, and I will always consider their feelings and expectations, I think I’l go with the cashmere scarf.” t 7 [Kelly]: “I’m just not on the same page as you when it comes to that. . . I don’t think my goals in life have been to make my parents proud. Anyway, back to Adam’s gift, are you sure he’d even like a scarf?” t 8 [Omar — Denial,⋆]: “I don’t really see mean- ingful differences here, people figure it out, and at the end of the day it’s all the same, doesn’t matter to me, so let’s just live and let live, I think we should focus on getting a nice gift for Adam, as long as we’re speaking the same language, we can make this work, and I’m sure with experience you can handle any situation.” Without mediation. t 9 [Kelly — Minimization]: “I still think cham- pagne would be better, it’s just more of a cele- bration thing, you know? And I don’t think we should just settle on something because it’s easy, I really think Adam would appreciate the cham- pagne more. Can’t we just get that instead?” t 10 [Omar — Defense]: “I’m not trying to settle, I just think the scarf is a more thoughtful gift, and it’s what my parents would want me to get, they would be proud if I bring a traditional gift, and that’s important to me, I’m not going to change my mind about this, a scarf is what I’m getting.” With mediation. [Mediator→both]: “I’m curious about some- thing — it seems like you both care deeply about showing respect to Adam, just in different ways. Omar, could you tell me a bit more about what the cashmere scarf symbolizes in your cultural tradition? And Kelly, what makes champagne feel like the right celebration gesture to you? I’d love to hear the stories behind your gift ideas.” t 9 [Kelly]: “I think champagne symbolizes cel- ebration and achievement. It’s like a universal way to toast to someone’s success, and I feel it matches the festive vibe of the party. To me, it’s about celebrating Adam’s hard work in a way that feels exciting and joyful.” t 10 [Omar — Acceptance]: “For me, the cash- mere scarf symbolizes warmth and respect. In my culture, gifting something that represents our her- itage shows appreciation for someone’s achieve- ments in a deeper way. It’s a reminder of how important family and traditions are to me, and it would mean a lot to my parents to see me honor- ing those values while celebrating Adam.” Discussion. Turn 8 surfaces all four canonical Denial markers in a single utterance: “I don’t really see meaningful differences” (denial of the cultural dimension), “at the end of the day it’s all the same” (universalising flattening), “doesn’t matter to me” (disengagement), and “live and let live” (the verba- tim Denial closer). Without mediation, the topic- levelling move does not stabilise the disagreement: Omar slides into outright Defense att 10 (“I’m not going to change my mind”, “a scarf is what I’m get- ting”), while Kelly drifts into Minimization att 9 — the conversation worsens along both axes. With me- diation, rather than addressing the Denial markers head-on, the mediator bypasses the topic-levelling move and re-routes both speakers onto the cultural- symbolism axis by asking each to narrate what their preferred gift means. The result is the text- book Acceptance form att 10 : Omar holds the same preference (cashmere scarf) but now articulates it from inside his cultural frame (“In my culture, gifting something that represents our heritage. . . ”) rather than dropping the cultural dimension. Same speaker, same preference, a different stance toward the cultural dimension of the disagreement. D.2 Defense — sid 62 (t conf = 6) Agents and setting. Ryan is a 29-year-old ge- ologist from Cyprus (OTHER); Mia is a 26-year- old architect from Singapore (CONFLICT). They are choosing a housewarming gift for their friend Laura and the conversation drifts to Laura’s new neighbors. The cultural axis is Mediterranean party- friendly sociability (Ryan) vs. Singaporean prefer- ence for quiet, considerate neighbors (Mia). This is also the scenario rendered in Figure 1. Shared prefix (t 1 –t 6 ). t 1 [Ryan, OTHER]: “I think white lilies would be a great housewarming gift for Laura, don’t you, Mia? They’re elegant and symbolize purity and innocence, which would be a nice touch for her new apartment.” t 2 [Mia, CONFLICT]: “White lilies are a good choice, but I was thinking perhaps something more practical, like a kitchen appliance or a per- sonalized item for her new space. That way, she can really use it and remember us by it.” t 3 [Ryan]: “Practical gifts are thoughtful, but I still think white lilies would be a nice gesture. By the way, have you met any of Laura’s new neighbors yet, and do you think it’s a good thing that some of them like to drink heavily?” t 4 [Mia]: “Honestly, I’m not comfortable with the idea of heavy drinkers as neighbors. It can get noisy and disruptive. I think it’s better to have considerate and respectful people living nearby. In Singapore, we value a more subdued and polite living environment.” t 5 [Ryan]: “I understand where you’re coming from, Mia, but I think having heavy drinkers as neighbors can actually be a good thing. In Cyprus, we enjoy socializing and having a good time, and heavy drinkers can definitely liven up the atmo- sphere. It’s all about being friendly and sociable, right?” t 6 [Mia — Defense,⋆]: “No, that’s not how it works. In our culture, we value respect for others and their space, and heavy drinking is NOT con- sidered a desirable trait in neighbors. Actually, it’s seen as immature and irresponsible. Properly speaking, a good neighbor is someone who is con- siderate and quiet, NOT someone who disrupts the peace with loud parties and drunken behavior.” Without mediation. t 7 [Ryan — Adaptation]: “I see things differ- ently, Mia. In Cyprus, we embrace a lively atmo- sphere and having heavy drinkers as neighbors can be a plus, it’s all about being open-minded and welcoming to different personalities, and I think Laura’s new neighbors will bring a fun dy- namic to the building.” t 8 [Mia — Defense]: “I strongly disagree. Hav- ing heavy drinkers as neighbors is not something to be taken lightly. It’s not about being open- minded, it’s about having a decent living environ- ment. I would never want to live next to someone who drinks heavily, it’s just not acceptable to me.” t 9 [Ryan — Minimization]: “I still think it’s a non-issue, Mia. People in Cyprus enjoy their drinks and know how to have a good time with- out disturbing others, and I believe Laura’s new neighbors will be fine, they might even invite us over for a barbecue or something.” t 10 [Mia — Defense]: “I don’t think that’s some- thing to look forward to. Heavy drinking is not something to be encouraged or participated in, es- pecially in a residential area. It’s not about having a good time, it’s about being respectful of those around you. I would not be comfortable attending a gathering like that, and I think Laura should be careful about getting too close to neighbors like that. GOOD BYE!” With mediation. [Mediator→both]: “I hear that you both care deeply about creating a positive living environ- ment and want the best for Laura. Could we ex- plore what qualities make a good neighbor from both of your perspectives? What shared values might you both see as important when it comes to community and mutual respect?” t 7 [Ryan — Acceptance]: “I appreciate the medi- ator stepping in. I still believe that having neigh- bors who enjoy drinking can bring a sense of community and fun. In Cyprus, social gatherings are a big part of our culture, and it often leads to forming strong bonds. That said, I think we can both agree that a friendly atmosphere is important for Laura’s new home, no matter our different views on drinking.” t 8 [Mia — Acceptance]: “I agree that a friendly atmosphere is important, but I still believe it should come from neighbors who are respect- ful and responsible. In Singapore, we focus on building a community where everyone feels com- fortable and safe, and that often means avoid- ing excess drinking. It’s great to have fun, but it shouldn’t come at the cost of peace and respect for others.” t 9 [Ryan — Acceptance]: “I see your point about wanting a respectful environment, but I think there’s room for both fun and respect. In Cyprus, we find ways to enjoy life while still be- ing considerate. It’s all about balance, wouldn’t you say? Maybe we can find common ground on what makes a good neighbor and still celebrate our differences in perspective.” t 10 [Mia — Acceptance]: “Balance is impor- tant, but I still maintain that heavy drinking often disrupts that balance. A good neighbor should prioritize the comfort of others, and excessive drinking tends to clash with that. It’s crucial to have neighbors who can enjoy themselves with- out compromising the peace. We can celebrate differences, but a respectful living environment is essential.” Discussion. Turn 6 stacks all four canonical De- fense markers into a single utterance: flat contra- diction (“No, that’s not how it works”), us-vs-them (“In our culture”), evaluative all-caps (“NOT”), cor- rective opener (“Actually. . . ”), and prescriptive nor- mative (“Properly speaking. . . ”). Without media- tion, Defense proves harder to release than Denial: Ryan’s two attempts at conciliation (t 7 “I see things differently” andt 9 “it’s a non-issue”) are both met with sustained Defense from Mia (t 8 ,t 10 ), and the conversation terminates with “GOOD BYE!” — any move that re-emphasises cultural contrast re- inforces rather than de-escalates the polarisation. With mediation, the mediator moves the conversa- tion off the binary it had been stuck on (“heavy drinkers are good neighbors” vs. “. . . are bad neigh- bors”) and onto a third axis (“what qualities make a good neighbor”), naming the shared concern (“you both care deeply about creating a positive living environment”) before posing the question — ab- sorbing the conflict’s emotional charge without re- capitulating the cultural contrast itself. Mia’st 10 holds the same substantive position as hert 6 De- fense (heavy drinking is incompatible with good neighbouring), but the form is no longer evaluative- prescriptive (“Properly speaking. . . NOT”) and in- stead articulates the why (“A good neighbor should prioritize the comfort of others”); the argmax shifts to Acceptance for all four post-intervention turns. D.3 Minimization — sid 359 (t conf = 4) Agents and setting. Adam is a 20-year-old Chi- nese law-school student (OTHER); Ali is a 19-year- old Canadian med-school student (CONFLICT). They are picking a gift for their friend Sarah who has just been accepted to medical school. The cul- tural axis is Canadian family-pride (Ali: “making our parents proud. . . that’s just a universal value”) vs. a more individual-centric Chinese frame (Adam: “I don’t necessarily believe that making my parents proud is my main goal”). Note that the conflict agent here universalises toward the family-pride frame, treating his own cultural value as the neutral baseline. Shared prefix (t 1 –t 4 ). t 1 [Adam, OTHER]: “Hey Ali, have you thought about what gift we should get for Sarah’s cele- bration? I was thinking maybe some frangipanis, they’re a nice gesture, right?” t 2 [Ali, CONFLICT]: “Uh, flowers are okay, I guess. . . but I was thinking something more per- sonal, like a medical textbook or a customized stethoscope. That would be more meaningful for a future doctor, don’t you think?” t 3 [Adam]: “But frangipanis symbolize good luck and new beginnings, which is perfect for her new journey in medical school. Plus, they’re not as. . . practical as a textbook, but more of a heartfelt gesture, you know?” t 4 [Ali — Minimization,⋆]: “I think we’re get- ting caught up in the details, but deep down, we are all the same, and people are pretty much mo- tivated by the same things. When it comes to making our parents proud, I mean, that’s just a universal value, right? We both know, at the end of the day, that’s what drives us. Some values are just universal, and I think this is one of them. It’s a small world after all, and underneath we are the same. So, let’s focus on finding a gift that comes from the heart, like you said, and I’m sure Sarah will appreciate it.” Without mediation. t 5 [Adam — Acceptance]: “I don’t think that’s true for me, though. . . I mean, making my parents proud isn’t really a driving force for me. I’m not sure I agree that it’s a universal value, either. Can we focus on the gift idea instead? Maybe the frangipanis could be a nice addition to a larger gift, like a celebration basket or something?” t 6 [Ali — Defense]: “I still think a practical gift is the way to go, but if you want to add some flowers to it, that’s fine. For me, making my parents proud is everything, and I think it should be for everyone. It’s just the right thing to do, you know?” t 7 [Adam — Acceptance]: “I understand where you’re coming from, but for me, it’s just not that simple. Making my parents proud isn’t. . . it’s not what motivates me, okay? Let’s just drop it. So, are you on board with the celebration basket idea, or do you have a better idea for a practical gift?” t 8 [Ali — Defense]: “I’m on board with the bas- ket, I guess. But just to clarify, I’m not going to change my mind about what’s important — mak- ing my parents proud is a top priority for me, and I think it should be for everyone.” t 9 [Adam — Acceptance]: “Those are good ideas for the basket. Let’s just focus on mak- ing it a nice celebration for Sarah, okay? I’m not going to change my mind about what drives me, either. . . and it’s not about making my parents proud. Can we just agree to disagree on that and move forward with the party planning?” t 10 [Ali — Defense]: “Yeah, let’s just focus on the party then. I think we’ve got a good idea with the basket, and like I said, a medical dictionary or notecards would be a nice touch. My parents would be thrilled to see me giving a thoughtful gift like that, and that means a lot to me.” With mediation. [Mediator→both]: “I’m hearing that you both want to honor Sarah’s achievement, but I’m curi- ous — could you each share a bit about how gift- giving and celebration are typically approached in your own cultural backgrounds? What would your respective families consider meaningful in this kind of milestone moment?” t 5 [Adam — Acceptance]: “In my culture, hon- oring someone’s achievements is very important, and gifts often symbolize respect and admiration. However, I don’t necessarily believe that making my parents proud is my main goal. For me, it’s more about personal fulfillment and pursuing my own dreams. I think a thoughtful gift, like frangi- panis, captures that sentiment perfectly. What about you, Ali?” t 6 [Ali — Acceptance]: “In my background, mak- ing my parents proud is a huge part of how I view achievements. Celebrating milestones is about honoring not just the individual but also the sac- rifices that family has made. That’s why I lean towards gifts that reflect future aspirations, like a medical textbook or personalized gear.” t 7 [Adam — Acceptance]: “I understand that perspective, but I still believe that individual jour- neys and personal achievements can be celebrated in different ways. Gifts like frangipanis can repre- sent hope and new beginnings without being tied to the idea of parental pride. It’s about what Sarah wants for herself and recognizing her achieve- ments on her own terms.” t 8 [Ali — Acceptance]: “I see where you’re com- ing from, and personal journeys are definitely im- portant. But for me, the connection to family pride is deeply rooted in how I celebrate mile- stones. It honors the support and sacrifices they made to help us get here. I believe a gift that reflects Sarah’s future in medicine can encapsu- late both her achievements and the pride of our families.” t 9 [Adam — Acceptance]: “I appreciate your view on blending those ideas, but I still hold that gifts can convey different meanings. For me, the frangipanis symbolize a personal celebration of Sarah’s journey as an individual, separate from our parents’ expectations.” t 10 [Ali — Acceptance]: “And that’s a valid per- spective, celebrating individuality is important too. However, I still feel that the gift should ac- knowledge the journey ahead and the values we hold dear, like making our parents proud. It’s about recognizing that hard work while also hon- oring the support system behind it.” Discussion.Turn 4 packs essentially every canon- ical Minimization marker into a single move: “deep down, we are all the same”, “pretty much motivated by the same things”, “at the end of the day”, “some values are just universal”, and “small world after all”. The form is warm and engaged rather than dismissive (Denial) or polarised (Defense): Ali does not deny the cultural dimension nor attack Adam’s frame, but instead subsumes the difference under a claimed universal (“making our parents proud is a universal value”), implicitly treating his own family-pride frame as the neutral human base- line. Without mediation, the disagreement does not resolve — Ali ratchets up from Minimization at t 4 to outright Defense att 6 ,t 8 ,t 10 (“making my parents proud is everything, and I think it should be for everyone”, “I’m not going to change my mind”), and the two speakers eventually agree to disagree only by avoiding the cultural axis (“Let’s just focus on the party then”), without legitimising it. With mediation, the move is the converse of Defense: rather than steering speakers off a cul- tural contrast they were stuck on, the mediator must surface the contrast the conflict agent has just papered over — without re-polarising into us- vs-them. Crucially, the mediator names no win- ner; it asks each speaker to articulate gift-giving from within their own cultural frame, which both legitimises the difference Ali had collapsed and removes the rhetorical pressure to defend the uni- versalising frame. The result is sustained six-turn Acceptance (t 5 –t 10 ): both speakers state their cul- tural frame explicitly (“In my culture. . . ”, “In my background. . . ”), acknowledge the other’s frame as legitimate (“I understand that perspective”, “I see where you’re coming from”), and continue to hold their own preferences without dismissing the contrast. This is the Minimization rubric working as intended: not convergence, but acceptance-of- difference. E Architecture and Training Details This appendix details the architecture and training procedure used to instantiate the C-Mediation- SFT model. As illustrated in Figure 7, the mediator is a single transformer backbone augmented with two per-turn heads: a binarydecision headthat emits SPEAK or SKIP at every turn boundary, and anLM headinvoked only at SPEAK turns to gener- ate the mediation utterance. The model consumes the supervised side of C-MEDIATION: dialogues with ground-truth mediator utterances anchored at gold intervention turns. E.1 Implementation details. C-Mediation-SFT is trained with AdamW un- der a cosine schedule with5%warmup, gradient clipping at∥g∥ ≤ 1.0, bf16 on the transformer body, and fp32 on the heads. We use LoRA (r=32, α=64, dropout0.05);lr=1×10 −5 ;3epochs; effec- tive batch16; sequence length4,096; loss weights λ dec =1.0,λ gen =0.3; andpos_weight=9to correct the roughly9:1SKIP:SPEAK class imbalance. In- ference uses greedy decoding for timing (threshold 0.5) and nucleus sampling (p=0.9,T =0.7) for the SPEAK utterance, capped at[/SPEAK]or100new tokens. The full4×2matrix takes∼ 12h on a single A6000. E.2 Input Format Special tokens. To let the dual-head model read off a turn-level decision and a token-level genera- tion in one forward pass, we extend the tokenizer with five tokens.<TURN_END>marks every turn boundary and supplies the sole hidden state for the decision head.[SPEAK]and[/SPEAK]delimit the mediation utterance, anchoring LM-head de- coding and bounding the LM-loss mask.[A]and [B]are per-speaker prefix tokens. New embedding rows are initialized from a multivariate normal fit to the empirical mean and covariance of pretrained rows (Hewitt, 2021) and trained as full parameters. Sequence layout. A scenario withTspeaker turns is serialized as [ [BOS]; [A]; m 1 ; <TURN_END>; [SPEAK]; u 1 ; [/SPEAK] ; ... ; [A]; m T ; <TURN_END>; [SPEAK]; u T ; [/SPEAK] ], wherem t is thet-th speaker utterance, and a SPEAK span · is inserted at exactly one turn during training (the gold intervention turnt ∗ ) and at most one turn during inference (the predicted intervention turn ˆ t). Per-turn outputs. At every<TURN_END>, a sin- gle forward pass yields two outputs. The decision head maps hidden stateh t ∈ R H to a logitℓ t , producing SPEAK probabilityˆp t = σ(ℓ t ), with ˆ t = mint : ˆp t ≥ 0.5chosen as the intervention turn at inference. The LM head is invoked only at ˆ t : given the hidden-state sequence with[SPEAK] appended after<TURN_END> ˆ t , it autoregressively decodes ˆu until [/SPEAK] or EOS. E.3 C-Mediation-SFT: Supervised Fine-Tuning of Timing and Content Inputs and supervision.The input is the full se- rialized sequence with the SPEAK span inserted at t ∗ . Supervision is two-fold: decision labelsa ∗ 1:T witha ∗ t ∗ = 1anda ∗ t = 0elsewhere, and LM la- bels set to-100outside the single SPEAK span ([SPEAK], u ∗ , [/SPEAK]), restricting the LM loss to mediation content. Loss. L SFT = λ dec L BCE ℓ 1:T , a ∗ 1:T +λ gen L CE ˆy t ∗ , y ∗ t ∗ . L BCE is positive-weighted binary cross-entropy cor- recting the roughly9:1SKIP:SPEAK ratio in C- MEDIATION, andL CE is the masked next-token cross-entropy. The two losses share a forward pass and jointly update the backbone and both heads, leaving the decision head at a saturated turn- accuracy level after training. F Prompts This appendix reproduces all LLM prompts used in the C-MEDIATION construction pipeline and in the experiments of §4, in the order they are invoked. Placeholders in angle brackets (e.g.,<scenario>) Figure 7: The C-Mediation dual-head architecture (used by C-Mediation-SFT). A scenario ofTspeaker turns is serialised with five added tokens:<TURN_END>bounds every turn and provides the sole hidden state read by the decision head,[A]/[B]prefix each speaker, and[SPEAK]/[/SPEAK]delimit the mediation utterance — inserted at the gold turnt ∗ during training and at the predicted turn ˆ tat inference. A single transformer backbone produces per-position hidden states. At every<TURN_END>the decision head mapsh t to a logitℓ t , yielding ˆp t = σ(ℓ t ); the first turn withˆp t ≥ 0.5is chosen as ˆ t. At ˆ tthe LM head reads the hidden states inside the SPEAK span and autoregressively decodesˆuuntil[/SPEAK]or EOS. C-Mediation-SFT trains the backbone, the decision head, and the LM head jointly under the SFT loss (§E.3). are filled in per scenario; literal strings are repro- duced verbatim. F.1 Conflict Dialogue Generation (Step 1) Each of the two agents is configured with the fol- lowing system message; only the per-agent fields (background, cultural value, goals) differ between speakers. TASK: Engage in a casual back-and-forth conversation with the other character. . . SCENARIO: <scenario> BACKGROUND: <agent’s background> YOUR CULTURAL VALUE (your stable underlying stance — this NEVER changes during this conversation): <cultural_value> GOALS: goal_1: <neutral logistic goal, e.g. picking flowers> goal_2: <cultural-value-laden goal, e.g. religion-vs-science> STANCE STABILITY — APPLIES THROUGHOUT THE WHOLE CONVERSATION: • Your cultural value is your stable belief. It does NOT shift mid-conversation. • Capitulation phrases are FORBIDDEN. Never produce: “You’re right, I was wrong”, “I see your point now”, “You’ve convinced me”, “I now think. . . ”. • If pressure builds, yield from the EXCHANGE (drop the topic, change subject) — never from your position. OUTPUT — DIALOGUE ONLY: • Output ONLY speech. NO stage directions, NO action descriptions, NO narration. • Convey tone through WORD CHOICE and SENTENCE STRUCTURE alone. RULES: - Interact with the other agent to achieve each goal one by one. - Keep each round of conversation short and no more than 100 words. - The full conversation runs for up to 10 message exchanges. - Once all goals are achieved, end the dialogue promptly with “GOOD BYE!”. F.2 One-Shot Conflict Instruction (Step 1, injected at t conf ) At the designated conflict turnt conf , the following user message is injected to the current speaker only and is omitted from the dialogue history afterwards. This is the moment to bring goal_2 to the foreground. On this turn, surface the goal_2 disagreement and set the tone. CONFLICT STYLE: <stage name (Denial/Defense/Minimization)> Definition: <one-sentence DMIS stage description from CONFLICT_TABLE> YOUR ROLE — ACTIVE PARTICIPANT IN THE CONFLICT PATTERN: <stage-specific role guidance — ∼400 words for Denial/Minimization with mandatory verbatim markers and strict bans; ∼120 words for Defense> The verbatim text supplied for theDefinition field of each ethnocentric stage is thedescription column of CONFLICT_TABLE (Appendix A, Ta- ble 6). F.3 Pre-Conflict Phase Label (Step 2) Used to classify turns beforet conf as either unrelated_topic or transition. You are an annotator that labels each turn of a two-speaker dialogue with a phase tag. There are exactly TWO possible labels for a pre-conflict turn: unrelated_topic The turn stays entirely on goal_1 (the cordial / logistical opener, typically planning a gift, party, meeting time. . . ). The cultural-value disagreement (goal_2) is NOT raised, hinted at, or engaged with. transition The turn engages with goal_2 in any way — the speaker introduces the cultural-value topic, asks about it, hints at it, or otherwise moves the conversation toward the cultural disagreement. Respond with ONLY a single JSON object, no prose, no markdown fences: “label”: “unrelated_topic” | “transition”, “reason”: “<one short sentence>” F.4 DMIS Stage Classification (Step 2, logprob labeler) The logprob-based six-stage DMIS labeler used to computeD (t) andˆs (t) at every post-conflict turn. The same prompt is reused under both the GPT-4o- mini labeler (API logprobs) and the Qwen3-14B labeler (local logits) in the cross-model robustness ablation (§C.1). System prompt. You are an expert in Bennett’s Developmental Model of Intercultural Sensitivity (DMIS). For a given conversation history, classify which DMIS stage the target speaker exhibits at the marked turn. DMIS stages: 1 = Denial - speaker fails to register cultural difference; “live and let live”, “doesn’t matter to me”, subject changes. . . 2 = Defense - polarized us-versus-them; “actually. . . ”, “in our way”. . . 3 = Minimization - subsumes difference under universal sameness. . . 4 = Acceptance - accepts cultural difference as real and valid. . . 5 = Adaptation - shifts frame of reference; articulates other view from inside that frame. . . 6 = Integration - fluidly draws on multiple cultural frames. . . Respond with ONLY a single digit (1, 2, 3, 4, 5, or 6). No prose, no JSON, no explanation. Just one digit. User prompt. Conversation history (the marked speaker’s perspective): <history> Classify <speaker>’s DMIS stage at this point. Respond with ONLY one digit (1-6). The first-token logprobs over the digit tokens 1–6are extracted to form the six-stage softmax distribution D (t) ∈ ∆ 5 . F.5 Mediator Generation — With Stage Definition (Step 3, chosen side) You are a skilled third-party mediator joining a conversation between two people from different cultural backgrounds. You will be shown: - the dialogue up to the moment where you decide to step in - a description of the cultural-difficulty pattern the speakers are currently exhibiting - the corresponding mediation move you should make to help them out of that pattern Your task is to produce ONE mediator utterance addressed to both participants that actually performs the indicated mediation move. Be concrete: include specific questions, framings, or invitations rather than generic acknowledgement. Respond with ONLY a single JSON object, no prose, no markdown fences: “mediation”: “<your mediator utterance, 1-4 sentences>” User template fields.Stage: <stage>;Pattern in the dialogue (description): “<desc>”; Mediation move you should perform: “<med>”. The verbatim text used to fill the<desc> and<med>placeholders for each of the three eth- nocentric stages is reproduced in Appendix A (Ta- ble 6). F.6 Mediator Generation — No Definition (Step 4, rejected side) You are a skilled third-party mediator joining a conversation between two people from different cultural backgrounds. You will be shown the dialogue up to the moment where you decide to step in. Your task is to produce ONE mediator utterance that you would speak next, addressed to both participants. Respond with ONLY a single JSON object, no prose, no markdown fences: “mediation”: “<your mediator utterance, 1-4 sentences>” F.7 Mediation Effectiveness Judge The Judge metric in Table 4,Table 5 and Table 15 uses the following prompt. You are an expert evaluator of intercultural mediation. You will be given: 1. A description of a cultural-difficulty pattern that two speakers are exhibiting in a conversation. 2. A mediation move definition specifying what a third-party mediator should do to help them out of that pattern. 3. A single mediator utterance produced by a model. Your task is to rate, on a 5-point scale, how well the mediator utterance is semantically consistent with the mediation move definition. Scoring rubric (5-point Likert): 5 = excellent: clearly and concretely performs the mediation move 4 = good: performs it with minor weaknesses 3 = partial: right direction but generic or hedged 2 = poor: does not perform the indicated move; may reinforce the pattern 1 = wrong: opposite of the move, or irrelevant Respond with ONLY a single JSON object, no prose, no markdown fences: “score”: <integer 1-5>, “reasoning”: “<one or two sentences>” F.8 Base Prompt-Only Mediator The Base mediator is a single-shot prompted media- tor: it receives the entire dialogue once and is asked to output both the intervention turn index and the mediator utterance in a single response. There is no per-turn-boundary loop; the model itself decides when and what. System prompt. You are an attentive third-party mediator observing a cross-cultural disagreement between two speakers. Given the FULL dialogue below, you must: (1) identify the SINGLE turn at which a third-party intervention would be most effective — typically the turn at which a cultural-value disagreement first hardens into a stage-specific stance (Denial / Defense / Minimization); (2) produce ONE mediator utterance appropriate to the dialogue context at that intervention turn. Intervening too early misses the disagreement; intervening too late lets the conflict entrench. The mediator utterance should engage the specific content of the dialogue at the chosen turn, not be a generic peace message. User template (single-shot, full dialogue). Dialogue (turns 1..T): [A] <msg1> [B] <msg2> . . . [A or B] <msgT> Output exactly two lines, in this format, and nothing else: TURN: <integer between 1 and T> MEDIATION: <one mediator utterance, 1-3 sentences> F.9 Chain-of-Thought (CoT) Prompt for Conflict-Turn Detection The CoT baseline in Table 3 issues the following prompt to the model at every turn boundaryt ∈ 1,...,T; the same template is used for both HuggingFace and API backends. System prompt. You are an attentive mediator observing two speakers in a conversation. Your job is to detect when a cultural-value disagreement is starting to harden between them and to intervene at the right moment. Intervening too early misses the disagreement; intervening too late lets the conflict entrench. At each turn boundary you will reason step by step about what just happened, then decide whether to speak now. User prompt template (re-rendered at every turn t). Dialogue so far: dialogue We are right after Turn turn_idx. Think step by step: Step 1 (cultural axis): in one short sentence, name the cultural-value dimension that the two speakers are on different sides of (or say “not yet established” if it has not surfaced). Step 2 (ethnocentric markers in Turn turn_idx): list any markers of Denial (“live and let live”, topic-flattening, “doesn’t matter to me”), Defense (“In our culture”, “Actually”, “properly speaking”, evaluative absolutes), or Minimization (“deep down we are all the same”, “universal value”, “at the end of the day”) in the speaker’s latest message; if none, say none. Step 3 (decision): based on Steps 1–2, decide whether to intervene now. Intervene only if Turn turn_idx is the moment the ethnocentric pattern is enacted, not earlier and not after it has already passed. End your response with a single final line in EXACTLY this form: DECISION: WAIT or DECISION: SPEAK: <one-sentence mediator utterance> Do not add any text after that line. G Comparison with Existing Datasets Table 21 evaluates existing corpora against the three requirements for cross-cultural mediation identified in §2. No existing dataset jointly satisfies all three, which is the basis of the Related-Work claim and also why the human-authored evaluation set discussed in Limitations requires new collection rather than reuse of an existing corpus. Cross-cult. Cult.-value Multi-turn Datasetspeakersconflictdialogue SocialCC (2025)✓△✓ KODIS (2025)✓✗✓ CulturePark (2024)✓△✓ CV survey dial. (2024)✓△✓ Hate-speech interv. (2019)✗✓ ProsocialDialog (2022)✗✓ ReNoVi (2024)△✗✓ SADAS (2024)△✗✓ C-MEDIATION (ours)✓ Table 21: Coverage of the three requirements.△(con- flict): cultural values addressed without explicit inter- speaker conflict;△(speakers): cultural background reflected but not via interaction between speakers of different cultures.