Development and application of a generative-AI-driven virtual simulation teaching system for international business negotiation scenarios

May combine two claims.

The moderator's role proved more important than initially expected-without it, persona drift across long sessions became noticeable to learners by the third or fourth turn.

“The moderator's role proved more important than we initially expected-without it, persona drift across long sessions became noticeable to learners by the third or fourth turn”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Average response latency settled at 2.3 seconds per agent turn, which proved acceptable to learners during pilot testing.

“Average response latency settled at 2.3 seconds per agent turn, which proved acceptable to learners during pilot testing.”

May combine two claims.

Cross-cultural scenario construction now scales without the logistical impossibility of recruiting authentic foreign counterparts for every cohort, and individualized feedback no longer waits on instructor bandwidth-the assessment engine returns session-specific commentary minutes after a session closes.

“Cross-cultural scenario construction now scales without the logistical impossibility of recruiting authentic foreign counterparts for every cohort, and individualized feedback no longer waits on instructor bandwidth -the assessment engine returns session-specific commentary minutes after a session closes.”

May combine two claims.

The adjusted effect sizes reported in Table 5 attenuate by roughly 8-14% relative to the unadjusted values but remain large and statistically significant.

“the adjusted effect sizes reported in Table 5 attenuate by roughly 8-14% relative to the unadjusted values but remain large and statistically significant”

May combine two claims.

Feedback usefulness scored highest (4.51), followed by content authenticity (4.39); usability scored lowest (4.08), with most complaints traced to occasional voice-recognition errors when learners spoke with strong regional accents.

“Feedback usefulness scored highest (4.51), followed by content authenticity (4.39); usability scored lowest (4.08), with most complaints traced to occasional voice-recognition errors when learners spoke with strong regional accents.”

May combine two claims.

The cross-border M&A scenario showed the largest experimental advantage, whereas the contract-drafting scenario exhibited a more modest gap.

“The cross-border M&A scenario, which carries the heaviest informational and cultural load, showed the largest experimental advantage, whereas the contract-drafting scenariowhere conventional role-play already exposes students to standard clauses-exhibited a more modest gap [56].”

• cites [56]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

This study designed, built, and tested a virtual simulation teaching system for international business negotiation in which generative AI takes the role of a culturally differentiated counterpart rather than a tutor.

“This study designed, built, and tested a virtual simulation teaching system for international business negotiation in which generative AI takes the role of a culturally differentiated counterpart rather than a tutor.”

May combine two claims.

Subgroup satisfaction differed in shape rather than magnitude: international trade students valued strategic-decision feedback most, English majors gravitated toward cultural and language register dimensions, and law-background students preferred dispute-resolution scenarios where formal argumentation mattered.

“Satisfaction differed in shape rather than magnitude: international trade students valued the strategic-decision feedback most, English majors gravitated toward the cultural and language register dimensions, and law-background students preferred the disputeresolution scenarios where formal argumentation mattered.”

May combine two claims.

Under a load of 40 concurrent sessions, the mean total latency held at 2.34 s (σ = 0.41 s); at 80 concurrent sessions, latency drifted to 3.91 s, still below the 5-second threshold that pilot learners had identified as the upper bound of acceptable wait time.

“Under a load of 40 concurrent sessions, the mean total latency held at 2.34 s (σ = 0.41 s); at 80 concurrent sessions, latency drifted to 3.91 s, still below the 5-second threshold that pilot learners had identified as the upper bound of acceptable wait time.”

May combine two claims.

Text-only turns clear the 5-second study threshold across the full tested range, while voice-augmented turns approach or cross it beyond 60 concurrent sessions.

“Text-only turns clear the study threshold across the full tested range, while voice-augmented turns approach or cross it beyond 60 concurrent sessions -a boundary that constrains the practical scale of voice-first deployments and that we return to in Section V.”

May combine two claims.

Experimental learners returned a composite satisfaction score of 4.34 on a five-point scale, against 3.67 for the control group.

“Experimental learners returned S learner = 4.34 on a five-point scale, against 3.67 for the control group.”

May combine two claims.

Pearson correlation across the twelve indicators reached r = 0.78, ICC(3,1) = 0.87, and a Bland-Altman analysis located the mean bias at only +0.07 rating points on a five-point scale, with 95% limits of agreement of [-0.11, +0.25].

“Pearson correlation across the twelve indicators reached r = 0.78, ICC(3,1) = 0.87, and a Bland-Altman analysis (Figure 9) located the mean bias at only +0.07 rating points on a five-point scale, with 95% limits of agreement of [-0.11, +0.25].”

May combine two claims.

Eight instructors used the system across the semester; their evaluation averaged 4.21 on pedagogical value, with the highest marks given to the assessment dashboard's ability to surface individual weaknesses that would otherwise stay invisible in cohort-sized classes.

“Eight instructors used the system across the semester.Their evaluation, captured through a structured rubric, averaged 4.21 on pedagogical value, with the highest marks given to the assessment dashboard's ability to surface individual weaknesses that would otherwise stay invisible in cohort-sized classes.”

Classified as probably not a substantial claim. May combine two claims.

Instructors move from staging role-plays to curating scenarios, calibrating rubrics, and intervening at the moments the AI cannot read.

“Instructors move from staging role-plays to curating scenarios, calibrating rubrics, and intervening at the moments the AI cannot read.”

May combine two claims.

Semi-structured interviews attributed the extra practice to three factors: the freshness of rotating counterpart profiles, the low social cost of failing in front of a non-judgmental AI, and the immediate rubric feedback at session close.

“the semi-structured interviews reported in Section 4.3 attribute the extra practice to three factors -the freshness of rotating counterpart profiles, the low social cost of failing in front of a non-judgmental AI, and the immediate rubric feedback at session close.”

The quote only partly states this claim.

At six-month follow-up (n = 54 from the experimental group), self-reported confidence in handling cross-cultural negotiation remained 0.71 standard deviations above baseline.

“Self-reported confidence in handling cross-cultural negotiation remained 0.71 standard deviations above baseline, computed as:”

May combine two claims.

Strategic decision-making: experimental mean 83.8 (SD 6.8) vs control 74.5 (SD 7.6), t=7.18, p<0.001, d=1.29.

“83.8 (6.8) 74.5 (7.6) 7.18 < 0.001 1.29”

May combine two claims.

The experimental group moved from a baseline mean of 3.41 to a posttest mean of 4.17 on the CCSS, while the control group rose only from 3.38 to 3.62.

“The experimental group moved from a baseline mean of 3.41 (on a five-point scale) to a posttest mean of 4.17, while the control group rose only from 3.38 to 3.62 [57].”

• cites [57]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Ablation studies showed that removing RAG dropped quality from 4.41 to 3.74; removing the cultural reasoning module dropped it to 3.91; removing the moderator agent caused a milder fall to 4.18 but visibly increased persona drift across long sessions.

“Removing RAG dropped quality from 4.41 to 3.74; removing the cultural reasoning module dropped it to 3.91; removing the moderator agent caused a milder fall to 4.18 but visibly increased persona drift across long sessions.”

May combine two claims.

Information & analysis: experimental mean 82.3 (SD 7.1) vs control 76.8 (SD 7.9), t=4.08, p<0.001, d=0.73.

“82.3 (7.1) 76.8 (7.9) 4.08 < 0.001 0.73”

May combine two claims.

In the experimental cohort, voice input was invoked in roughly 32% of sessions, most heavily in the first two weeks and again during the M&A scenarios; the remaining 68% of sessions ran text-only.

“Telemetry from the experimental cohort indicates that voice input was invoked in roughly 32% of sessions, most heavily in the first two weeks and again during the M&A scenarios; the remaining 68% of sessions ran text-only.”

May combine two claims.

Overall negotiation score: experimental mean 84.6 (SD 6.2) vs control 75.1 (SD 7.4), t=7.74, p<0.001, d=1.39.

“84.6 (6.2) | 75.1 (7.4) | 7.74 | < 0.001 | 1.39”

May combine two claims.

Under the quality index, GPT-4o reached Q = 4.41, Qwen-2.5-72B reached Q = 4.18, and a baseline GPT-3.5-turbo configuration trailed at Q = 3.62.

“Under this index, GPT-4o reached Q = 4.41, Qwen-2.5-72Breached Q = 4.18, and a baseline GPT-3.5-turboconfiguration trailed at Q = 3.62 [59].”

• cites [59]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionThe authors state this as an assumption: Cultural appropriateness matters most in this domain.

May combine two claims.

Cross-cultural sensitivity: experimental mean 86.9 (SD 5.4) vs control 74.2 (SD 8.1), t=10.27, p<0.001, d=1.84.

“86.9 (5.4) 74.2 (8.1) 10.27 < 0.001 1.84”

Classified as probably not a substantial claim.

The system has a retrieval-augmented scenario sampler that kept training material fresh across sessions.

“a retrieval-augmented scenario sampler that kept training material fresh across sessions”

May combine two claims.

Strategy diversity index: experimental mean 0.71 (SD 0.09) vs control 0.54 (SD 0.11), t=9.42, p<0.001, d=1.69.

“0.71 (0.09) 0.54 (0.11) 9.42 < 0.001 1.69”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Two instructors raised concerns about overreliance: learners might accept the AI's feedback as authoritative when it is, in fact, probabilistic.

“Two instructors raised concerns about overreliance: learners might accept the AI's feedback as authoritative when it is, in fact, probabilistic.”

May combine two claims.

The experimental group outperformed the control group on every competence dimension, with the gap widest on cross-cultural sensitivity and strategic decision-making.

“The experimental group outperformed the control group on every dimension, with the gap widest on cross-cultural sensitivity and strategic decisionmaking-a pattern that matches the design intent of the AI module rather than reflecting any generalized improvement.”

May combine two claims.

Empirical concurrent throughput reached 14.6 successful turns per second on the production cluster, with ρsuccess = 0.984.

“Empirical Θ reached 14.6 successful turns per second on the production cluster, with ρsuccess = 0.984.”

May combine two claims.

Experimental learners initiated 41% more voluntary sessions than their control-group peers, and their voluntary session length grew rather than shrank across weeks.

“Experimental learners initiated 41% more voluntary sessions than their control-group peers, and their voluntary session length grew rather than shrank across weeks.”

May combine two claims.

CCSS gain: experimental mean 0.76 (SD 0.21) vs control 0.24 (SD 0.19), t=14.43, p<0.001, d=2.59.

“0.76 (0.21) 0.24 (0.19) 14.43 < 0.001 2.59”

May combine two claims.

Sessions completed: experimental mean 31.4 (SD 4.7) vs control 22.3 (SD 3.9), t=11.74, p<0.001, d=2.10.

“31.4 (4.7) 22.3 (3.9) 11.74 < 0.001 2.10”

May combine two claims.

The largest experimental advantage appears in the cross-border M&A scenario, the smallest in contract drafting.

“The largest experimental advantage appears in the cross-border M&A scenario, the smallest in contract drafting.”

Probably not a substantial claim: may not pass S3 (load-bearing).

End-to-end response latency per agent turn was measured across five concurrent-session loads, stratified by input modality.

“End-to-end response latency per agent turn across five concurrent-session loads, stratified by input modality (text-only vs voice-augmented).”

May combine two claims.

Knowledge posttest: experimental mean 87.2 (SD 5.5) vs control 81.3 (SD 6.4), t=5.51, p<0.001, d=0.99.

“87.2 (5.5) 81.3 (6.4) 5.51 < 0.001 0.99”

May combine two claims.

The two profiles (learner satisfaction and instructor evaluation) align closely on content authenticity, feedback usefulness, and pedagogical value, and diverge most on the usability axis where voice-recognition errors dominated open-text complaints.

“the two profiles align closely on content authenticity, feedback usefulness, and pedagogical value, and diverge most on the usability axis where voice-recognition errors dominated open-text complaints.”

May combine two claims.

The pedagogical appeal of RAG is twofold: it curbs hallucination and it lets instructors update the underlying knowledge base without retraining the model.

“The pedagogical appeal of RAG is twofold.It curbs hallucination-a serious concern when a model is meant to enact a Japanese trading partner or a Brazilian buyer with culturally specific norms-and it lets instructors update the underlying knowledge base without retraining the model [23].”

• cites [23]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Evaluating educational effectiveness draws on Kirkpatrick's four-level model: reaction, learning, behavior, and results.

“Evaluating educational effectiveness draws on Kirkpatrick's four-level model: reaction, learning, behavior, and results [30].”

• cites [30]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

Kolb's experiential learning cycle provides a rhythm for structuring each negotiation session and its debrief.

“Kolb's experiential learning cycleconcrete experience, reflective observation, abstract conceptualization, active experimentation-provides a rhythm for structuring each negotiation session and its debrief [27].”

• cites [27]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Voice input was optional, adopted in roughly a third of sessions, and its confounding with modality-dependent latency has been reported openly rather than absorbed silently.

“Voice input was optional, adopted in roughly a third of sessions, and its confounding with modality-dependent latency has been reported openly rather than absorbed silently.”

May combine two claims.

Baseline equivalence was checked across grade point average, prior English proficiency (CET-4 scores), and a pretest of negotiation knowledge; no statistically meaningful differences emerged between groups.

“Baseline equivalence was checked across grade point average, prior English proficiency (CET-4 scores), and a pretest of negotiation knowledge; no statistically meaningful differences emerged between groups.”

Classified as probably not a substantial claim. May combine two claims.

The assigned protocol was symmetric: six mandatory sessions per scenario for every learner in either arm, giving 24 scheduled sessions per student.

“The assigned protocol was symmetric: six mandatory sessions per scenario for every learner in either arm, giving 24 scheduled sessions per student.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

A recent synthesis by Yan and colleagues formalized the promises and challenges of GenAI across the learning cycle and identified authentic interactional practice as one of the least explored frontiers.

“a recent synthesis by Yan and colleagues in Nature Human Behaviour formalized the promises and challenges of GenAI across the learning cycle and identified authentic interactional practice as one of the least explored frontiers [66].”

• cites [66]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The full backend, prompt templates, evaluation pipeline, and anonymized session-trace samples have been packaged and are available with the article's Related files, and a Code Availability section records the archive.

“the full backend, prompt templates, evaluation pipeline, and anonymized session-trace samples have been packaged and are available with the article's Related files, and a Code Availability section in the Declarations records the archive”

May combine two claims.

Cronbach's α measured on a pilot cohort (n = 34) reached 0.86 across the full rubric, with dimension-level α values between 0.78 and 0.89.

“Cronbach's α measured on a pilot cohort (n = 34) reached 0.86 across the full rubric, with dimension-level α values between 0.78 and 0.89.”

May combine two claims.

Large language models such as GPT-4 and its successors can sustain coherent multi-turn dialogue, adopt distinct personas, and respond to nuanced prompts in ways that approximate human conversational partners.

“Large language models such as GPT-4 and its successors can sustain coherent multi-turn dialogue, adopt distinct personas, and respond to nuanced prompts in ways that approximate human conversational partners [5].”

• cites [5]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Randomization used a computer-generated sequence (block size four, seed pre-registered with the departmental research office); allocation concealment was preserved through sequentially numbered opaque sealed envelopes opened only in the presence of a witness after each student had completed baseline testing.

“Randomization used a computer-generated sequence (block size four, seed pre-registered with the departmental research office); allocation concealment was preserved through sequentially numbered opaque sealed envelopes opened only in the presence of a witness after each student had completed baseline testing.”

May combine two claims.

Interaction quality in traditional international business negotiation courses depends on the seniority and improvisational skill of fellow students, leaving weaker groups without a competent counterpart.

“Interaction quality depends on the seniority and improvisational skill of fellow students, leaving weaker groups without a competent counterpart.”

May combine two claims.

Both groups received identical lecture content and identical scenario themes across eight weeks of practical training.

“Both groups received identical lecture content and identical scenario themesinternational trade contract negotiation, cross-border merger and acquisition negotiation, technology licensing negotiation, and dispute resolution negotiation -across eight weeks of practical training.”

May combine two claims. The quote only partly states this claim.

These initial findings suggest that generative AI may extend the constructivist promise of virtual simulation into interactional domains previously bottlenecked by partner availability, with gains concentrated where conventional...

“These initial findings -obtained in one cohort of Chinese undergraduates -suggest that generative AI may extend the constructivist promise of virtual simulation into interactional domains previously bottlenecked by partner availability”

• AssumptionRests on an assumption: The twelve-indicator rubric used for assessment is a valid measure of negotiation competence.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Each phase of international business negotiation calls upon a distinct skill cluster: information gathering and BATNA estimation in preparation, anchoring and framing at the opening, concession management and packaging through the middle, and closure rituals at the end.

“Each phase calls upon a distinct skill cluster: information gathering and BATNA estimation in preparation, anchoring and framing at the opening, concession management and packaging through the middle, and closure rituals at the end.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

AI counterparts inherit biases from their training corpora that ongoing curation can mitigate but not eliminate.

“AI counterparts inherit biases from their training corpora that ongoing curation can mitigate but not eliminate”

Probably not a substantial claim: may not pass S3 (load-bearing).

Two independent coders processed the transcripts under an open-then-axial coding scheme (Cohen's κ = 0.83 on the initial pass).

“Two independent coders processed the transcripts under an open-then-axial coding scheme (Cohen's κ = 0.83 on the initial pass).”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

In an education deployment of our size, LoRA turns culturally specialized adapter training from a prohibitive line item into a routine one.

“In an education deployment of our size, LoRA turns culturally specialized adapter training from a prohibitive line item into a routine one.”

Classified as probably not a substantial claim. May combine two claims.

The sample is one cohort of third-year Chinese undergraduates at one university within one program, running across a single fifteen-week semester.

“The sample is one cohort of third-year Chinese undergraduates at one university within one program, running across a single fifteen-week semester.”

Classified as probably not a substantial claim.

Outcome measurement leans on blind human raters and validated external instruments (CCSS, knowledge exam) alongside the twelve-indicator rubric.

“Outcome measurement leans on blind human raters and validated external instruments (CCSS, knowledge exam) alongside the twelve-indicator rubric”

Classified as probably not a substantial claim.

Negotiation competence can be decomposed into four trainable strands: information and analytical skill, communication and persuasion skill, cross-cultural sensitivity, and strategic decision-making under uncertainty.

“From these considerations we decompose negotiation competence into four trainable strands: information and analytical skill, communication and persuasion skill, crosscultural sensitivity, and strategic decision-making under uncertainty [34].”

• cites [34]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

All primary outcomes are analyzed both as raw group comparisons and, in a pre-registered secondary analysis, as ANCOVA-adjusted estimates that condition on pretest score, GPA, CET-4 score, and total voluntary-practice minutes.

“All primary outcomes reported below are analyzed both as raw group comparisons and, in a pre-registered secondary analysis, as ANCOVA-adjusted estimates that condition on pretest score, GPA, CET-4 score, and total voluntary-practice minutes.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). The quote only partly states this claim.

The decay in confidence was shallower than the typical post-course drop reported in conventional curricula.

“a decay shallower than the typical post-course drop reported in conventional curricula [62].”

• cites [62]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The twelve-indicator rubric was developed in-house and its cross-context transportability remains an open question.

“the rubric itself was developed in-house and its cross-context transportability remains an open question.”

Probably not a substantial claim: may not pass S3 (load-bearing).

International work by Kasneci and colleagues mapped the opportunities and risks of large language models in classrooms.

“International work by Kasneci and colleagues mapped the opportunities and risks of large language models in classrooms [12]”

• cites [12]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

Brown, Collins, and Duguid extended this view through situated cognition, arguing that concepts inhabit the activities and contexts within which they are produced.

“Brown, Collins, and Duguid extended this view through situated cognition, arguing that concepts inhabit the activities and contexts within which they are produced [26].”

• cites [26]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

The retrieval-augmented design curbs hallucination but does not eliminate it.

“The retrievalaugmented design curbs hallucination but does not eliminate it.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The system-only indicators inform learner-facing feedback but do not enter the primary outcome analysis in Section 4.1.

“The system-only indicators inform learner-facing feedback but do not enter the primary outcome analysis in Section 4.1.”

Classified as probably not a substantial claim.

Reliability of AI-generated content remains the most pressing limitation.

“Reliability of AI-generated content remains the most pressing limitation.”

May combine two claims.

Casting a generative model as counterpart rather than tutor supplies the missing mechanism to instantiate context, and psychological fidelity is the axis on which a counterpart-shaped AI can deliver.

“Casting a generative model as counterpart rather than tutor supplies exactly that missing constituent, and it extends the constructivist and fidelity vocabularies in a specific way: the fidelity axis that had received the least attention in business-education simulation, psychological fidelity in the Kuipers sense [63], is precisely the axis on which a counterpart-shaped AI can deliver.”

• cites [63]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Sampling for interviews was purposive and stratified: six students from each of three posttest score terciles, balanced by gender, with recruitment closed once thematic saturation was reached across the three strata.

“Sampling was purposive and stratified: we drew six students from each of three posttest score terciles, balanced by gender, and closed recruitment once thematic saturation was reached across the three strata.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

The ambivalence is not a defect but a marker that the simulation crossed into the territory where genuine practice begins.

“We see this ambivalence not as a defect but as a marker that the simulation crossed into the territory where genuine practice begins.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The six-month retention analysis covers only the experimental arm and therefore reads as a within-group trajectory, not a between-arm contrast.

“The six-month retention analysis covers only the experimental arm and therefore reads as a within-group trajectory, not a between-arm contrast.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The eight-week intervention is short relative to the career-length consolidation of negotiation skill.

“the eight-week intervention is short relative to the career-length consolidation of negotiation skill”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The second research stream explores how generative AI can support writing instruction, programming education, and conversational language practice, with growing evidence that adaptive feedback raises learner engagement.

“The second stream, situated within educational technology, explores how generative AI can support writing instruction, programming education, and conversational language practice, with growing evidence that adaptive feedback raises learner engagement [11].”

• cites [11]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Conventional international business negotiation courses struggle to deliver authentic cross-cultural interaction, immediate individualized feedback, and scenario variety at scale.

“Conventional international business negotiation courses struggle to deliver authentic cross-cultural interaction, immediate individualized feedback, and scenario variety at scale”

• Also stated in: § A R T I C L E I N P R E S S · boosted

  “Traditional international business negotiation courses, in our observation and confirmed by recent surveys, exhibit four persistent weaknesses.”

• cites [14]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [15]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The first contribution is a modular architecture for pedagogical multi-agent dialogue that other interaction-heavy domains, from counseling to diplomatic practice, could plausibly adapt.

“The first is methodologicala modular architecture for pedagogical multi-agent dialogue that other interaction-heavy domains, from counseling to diplomatic practice, could plausibly adapt.”

May combine two claims.

The fixed branching logic of earlier implementations limited replay value and made it difficult to expose learners to the improvisational quality that real negotiations demand.

“The fixed branching logic limited replay value and made it difficult to expose learners to the improvisational quality that real negotiations demand.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Hall's distinction between high-context and low-context communication adds a complementary lens, capturing how much meaning travels through explicit verbal content versus surrounding cues such as silence, hierarchy, and prior relationship.

“Hall's distinction between highcontext and low-context communication adds a complementary lens, capturing how much meaning travels through explicit verbal content versus surrounding cues such as silence, hierarchy, and prior relationship [33].”

• cites [33]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Chinese universities face mounting pressure to cultivate practitioners capable of handling high-stakes negotiations under uncertain geopolitical conditions, a demand that conventional classroom instruction struggles to meet.

“Chinese universities, in particular, face mounting pressure to cultivate practitioners capable of handling high-stakes negotiations under uncertain geopolitical conditions, a demand that conventional classroom instruction struggles to meet [3].”

• cites [3]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Return rates were 59/60 in the experimental arm and 58/61 in the control arm, with non-response driven by illness during the final week.

“Return rates were high: 59/60 in the experimental arm and 58/61 in the control arm, with non-response driven by illness during the final week.”

Probably not a substantial claim: may not pass S2 (checkable).

Knowledge, according to constructivist accounts, is not transmitted intact from instructor to learner but assembled through active engagement with authentic problems.

“Knowledge, according to constructivist accounts, is not transmitted intact from instructor to learner but assembled through active engagement with authentic problems [25].”

• cites [25]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

What remains underexplored is the intersection of these strands: a purpose-built negotiation environment in which AI agents act as culturally differentiated counterparties rather than as generic conversational tools, and in which the human learner-not the model-is the object of evaluation.

“What remains underexplored is the intersection of these strands: a purpose-built negotiation environment in which AI agents act as culturally differentiated counterparties rather than as generic conversational tools, and in which the human learner -not the model -is the object of evaluation [13].”

• cites [13]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Virtual simulation systems are conventionally decomposed into a presentation layer, an interaction layer, a logic and content layer, and a data layer, with the design principles of fidelity, interactivity, scalability, and assessability guiding each.

“Architecturally, virtual simulation systems are conventionally decomposed into a presentation layer, an interaction layer, a logic and content layer, and a data layer, with the design principles of fidelity, interactivity, scalability, and assessability guiding each [29].”

• cites [29]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

The primary outcome analysis continues to rely on blind human ratings, and automated scores enter only in learner-facing feedback and internal engineering telemetry.

“the primary outcome analysis therefore continues to rely on blind human ratings, and the automated scores enter only in learner-facing feedback and in the internal engineering telemetry.”

May combine two claims.

Scenarios in traditional international business negotiation courses tend to be static, drawn from textbook cases that students often anticipate before the role-play begins.

“Scenarios tend to be static, drawn from textbook cases that students often anticipate before the role-play begins [14].”

• cites [14]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim. May combine two claims.

Blind rubric scoring proceeded on this dataset: for every session in both arms, the full transcript was extracted, stripped of role identifiers, model-specific stylistic cues, and any timestamp or telemetry markers, and re-tagged with a random session code before being routed to two trained graders blind to group assignment (inter-rater agreement κ = 0.81).

“Blind rubric scoring proceeded on this dataset: for every session in both arms, the full transcript was extracted, stripped of role identifiers, model-specific stylistic cues, and any timestamp or telemetry markers, and re-tagged with a random session code before being routed to two trained graders blind to group assignment (inter-rater agreement κ = 0.81 );”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Parallel work in the NLP community, most visibly the NegotiationArena platform of Bianchi and colleagues, has begun to probe LLM negotiation behavior as an object of study, though this line of work targets agent-to-agent evaluation rather than human learning.

“Parallel work in the NLP community -most visibly the NegotiationArena platform of Bianchi and colleagues -has begun to probe LLM negotiation behavior as an object of study, though this line of work targets agent-to-agent evaluation rather than human learning [67].”

• cites [67]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Attrition was low but not zero: two students in the experimental group withdrew for medical reasons and one control student transferred programs mid-semester, leaving 61/62/61 for primary analysis under an intention-to-treat frame.

“Attrition was low but not zero: two students in the experimental group withdrew for medical reasons and one control student transferred programs mid-semester, leaving 61/62/61 for primary analysis under an intention-to-treat frame.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The first research stream consistently identifies cross-cultural competence as the most fragile dimension of the trained skillset.

“The first, rooted in business and language education, examines pedagogical models for negotiation training-case-based learning, role-play, and simulation games-and consistently identifies cross-cultural competence as the most fragile dimension of the trained skillset [10].”

• cites [10]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Culture shapes every phase of international business negotiation, sometimes invisibly.

“Culture shapes every one of these phases, sometimes invisibly.”

Classified as probably not a substantial claim. May combine two claims.

The questionnaire was distributed to both arms with parallel item wording so that the satisfaction gap could be interpreted as a cross-arm comparison rather than as an absolute score.

“The questionnaire was distributed to both arms -experimental learners rated the AI simulation system and control learners rated their conventional role-play curriculum against parallel item wording -so that the satisfaction gap could be interpreted as a cross-arm comparison rather than as an absolute score.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Effect sizes in this range are unusually large for educational interventions of this length.

“effect sizes in this range are unusually large for educational interventions of this length”

May combine two claims.

The retention figure is a within-group change from baseline rather than a between-arm contrast, and it should not be read as evidence that AI-driven training decays more slowly than conventional training.

“As a consequence, the retention figure is a within-group change from baseline rather than a between-arm contrast, and it should not be read as evidence that AI-driven training decays more slowly than conventional training.”

May combine two claims.

We mitigated cultural stereotyping through expert review of the case base and through deliberate within-nationality variation in personality parameters.

“We mitigated this through expert review of the case base and through deliberate within-nationality variation in personality parameters”

May combine two claims.

The selective pattern suggests the intervention worked through the mechanisms the design anticipated, but it does not by itself rule out competing explanations.

“This selective pattern is itself informative; it suggests the intervention worked through the mechanisms the design anticipated, but it does not by itself rule out competing explanations that we take up in turn.”

May combine two claims.

Content quality was assessed across three dimensions-fluency, contextual relevance, and cultural appropriateness-each rated by three expert annotators on a five-point scale across 600 sampled agent turns, with inter-rater reliability α = 0.79 on Krippendorff's α.

“Content quality was assessed across three dimensions-fluency, contextual relevance, and cultural appropriateness-each rated by three expert annotators on a five-point scale across 600 sampled agent turns.Inter-rater reliability reached α = 0.79 on Krippendorff's α.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The study received Institutional Review Board approval and written informed consent from every participant.

“After receiving Institutional Review Board approval and obtaining written informed consent from every participant, we enrolled 124 third-year students from the International Business Negotiation course.”

Classified as probably not a substantial claim.

Psychological fidelity is at the heart of authentic simulation.

“the axis Kuipers and colleagues placed at the heart of authentic simulation”

• cites Kuipers et al., 2019Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

Post-intervention negotiation skill scores across the four competence dimensions were scored by two blind human graders on redacted transcripts.

“both scored by two blind human graders on redacted transcripts”