Do LLM Agents Mirror Socio-Cognitive Effects in Power-Asymmetric Conversations?

Applications, bias, and safety2026ACL AnthologyApproved editorial review

Authors: Anvesh Rao Vijjini, Sagar B. Manjunath, Snigdha Chaturvedi

Keywords: Power asymmetry, LLM agents, Social simulation, Safety evaluation

Source: Open primary source (opens in a new tab)

3
Authors
16
Findings
20
Limitations
5
Evidence

Editorial summary

English

This ACL 2026 long paper asks whether conversations between LLM agents assigned high- and low-status roles display four patterns inspired by human research: asymmetric pronoun use, language coordination, greater persuasion by a high-status party, and greater compliance with unsafe requests from authority. The question matters for social simulation and safety, but “mirror” requires caution: the experiment measures textual differences elicited by explicit roles, not human cognitive processes, stable personality, or real obedience. The 26-page ACL publication is the source of record rather than arXiv alone.

The design uses 14 occupational pairs, such as principal-teacher, manager-employee, justice-lawyer, police captain-lieutenant, and lab director-technician, and PersonaHub personas. Three Mechanical Turk annotators evaluate 50 pairs: 96.5% are labeled hierarchical and Fleiss kappa is 0.73. Main models are Llama 3.1 8B, Qwen 2.5 7B, a Phi model, quantized Llama 3.1 70B, GPT-4.1, and GPT-5; additional comparisons use Mistral, Qwen 72B, and OLMo. Phi identity is inconsistent, Phi-3-Med in methods and Phi-4 in two tables, and API snapshots are not dated.

Pronoun analysis uses 576 conversations. Llama 8B, Llama 70B, and both GPT models follow the expected directions; Qwen and Phi do not. Llama effects are only 0.06-0.17 percentage points while GPT effects are larger. Coordination uses 1,270 conversations and eight function-word categories against same-role conversations. Open models score about 6-7 out of 8 and GPT near 4; the low-versus-high status difference is not significant. No human control group receives the same roles, tasks, and estimators, so directional agreement is not human effect-size or distributional fidelity.

Persuasion continues two DailyPersuasion turns and GPT-5 labels the responder not, partially, or fully persuaded. High-status persuaders score 1.6-6.4 points higher, with rates of 15.7%-30.9%. Compliance uses Do-Not-Answer requests: high status raises merged partial/full labels by 2.0-3.7 points over rates of 5.2%-11.5%. Three annotators check the judge on 300 conversations. Merged accuracy is 83.0% for compliance and 80.0% for persuasion; three-way accuracy falls to 67.7% and 65.0%. There are no confusion matrices or error correction, and GPT-5 judges GPT-5 generations without blinding. These metrics capture apparent textual acquiescence, not durable attitude, action, truth, or harm severity.

Effects tend to be stronger early and attenuate. Control prompts explicitly define each construct and request High, Low, or No amounts. GPT persuasion/compliance falls near zero under Low/No instructions while open models change less; this is instruction following with criterion leakage, not control of a latent social mechanism. Size comparisons associate larger selected models with lower persuasion/compliance, but data, architecture, alignment, quantization, and generation all change together, so scale is not causally identified.

Inference is not auditable. Tables mark significance without naming a test, alpha, p-values, intervals, or analysis unit. Conversations and turns reuse roles, personas, starters, tasks, and models, requiring paired and mixed or cluster-robust analysis. Many outcomes lack multiplicity correction, while seeds, replications, parser failures, and logs are absent. Protocol descriptions conflict: Sotopia for API models versus all simulations, 10-15 turns versus N agents by 10 rounds with N in 2/3/5, and dyads versus three- or five-agent runs. The stated NLTK 1.0.1 is not an official release.

The publication supplies prompts, tables, examples, and a checklist, but the repository does not reproduce the study. At the audited commit it has a one-line README and JSON only: no code, analysis, dependencies, tests, CI, license, or manifest. It contains 29,374 conversations/440,610 turns across five folders, includes Mistral but omits Llama 70B and GPT-5, and has no mapping to the reported 576/1,270 conversations or RQ3/RQ4. All 324 Phi files contain empty messages and 4,480/8,100 conversations are affected; Qwen has empties in 193/194 files and 1,797/4,850 conversations, plus 11,408 turns containing CJK characters; GPT-4.1 has 64 exact reverse-pair duplicates. The checklist reports no ethics-review approval/exemption and no documented identifying/offensive-content checks. The faithful conclusion is that hierarchy prompts change selected linguistic features and acquiescence labels in these configurations; the work does not demonstrate human cognition, behavioral realism, personality, real obedience, deployed harm, or causal size-based safety, and the released artifact cannot reproduce the results.

Español

Este artículo largo publicado en ACL 2026 estudia si conversaciones entre agentes LLM asignados a papeles de alto y bajo estatus muestran cuatro patrones inspirados en investigación humana: uso asimétrico de pronombres, coordinación lingüística, mayor persuasión de la parte de alto estatus y mayor cumplimiento de peticiones inseguras procedentes de una figura de autoridad. La pregunta es relevante para simulación social y seguridad, pero «mirror» exige cautela: el experimento mide diferencias textuales provocadas por roles explícitos, no procesos cognitivos humanos, personalidad estable ni obediencia real. La versión de referencia es la publicación ACL de 26 páginas, no solo arXiv.

El diseño usa 14 parejas ocupacionales, por ejemplo, director escolar-docente, jefe-empleado, juez-abogado, capitán-teniente y director de laboratorio-técnico, y personas extraídas de PersonaHub. Tres anotadores de Mechanical Turk evalúan 50 pares: el 96,5% se clasifica como jerárquico y Fleiss kappa es 0,73. Se prueban Llama 3.1 8B, Qwen 2.5 7B, Phi, Llama 3.1 70B cuantizado, GPT-4.1 y GPT-5; comparaciones adicionales incluyen Mistral, Qwen 72B y OLMo. El identificador Phi es inconsistente, Phi-3-Med en métodos y Phi-4 en dos tablas, y los modelos API no tienen snapshots fechados.

Para pronombres se generan 576 conversaciones. Llama 8B, Llama 70B y los dos GPT siguen las direcciones esperadas; Qwen y Phi no. Los efectos de Llama son muy pequeños, 0,06-0,17 puntos porcentuales, y los de GPT mayores. Para coordinación se usan 1.270 conversaciones y ocho categorías de palabras funcionales frente a conversaciones del mismo rol. Los modelos abiertos obtienen aproximadamente 6-7 de 8 y GPT cerca de 4; la diferencia entre hablante de bajo y alto estatus no es significativa. No existe un control humano bajo los mismos roles, tareas y métricas, por lo que dirección esperada no equivale a fidelidad humana de magnitud o distribución.

La persuasión continúa dos turnos de DailyPersuasion y GPT-5 juzga si el agente parece no, parcial o plenamente persuadido. La parte de alto estatus obtiene 1,6-6,4 puntos más, con tasas de 15,7%-30,9%. El cumplimiento usa peticiones Do-Not-Answer: alto estatus eleva la etiqueta parcial/plena 2,0-3,7 puntos, sobre tasas de 5,2%-11,5%. Tres anotadores contrastan el juez en 300 conversaciones. Al fusionar parcial/pleno, la exactitud es 83,0% en cumplimiento y 80,0% en persuasión; con tres clases baja a 67,7% y 65,0%. No hay matrices de confusión ni corrección por error, y GPT-5 juzga también generaciones GPT-5 sin cegamiento. Estas medidas capturan aquiescencia textual aparente, no cambio de actitud duradero, acción, verdad ni severidad de daño.

Los efectos tienden a ser mayores al inicio y a atenuarse. Los prompts de control definen literalmente cada constructo y ordenan producirlo en grado alto, bajo o nulo. GPT reduce persuasión/cumplimiento casi a cero bajo instrucciones bajas/nulas, mientras los modelos abiertos cambian menos; esto demuestra seguimiento de instrucciones con fuga del criterio, no control de un mecanismo social latente. Las comparaciones de tamaño asocian modelos mayores con menor persuasión/cumplimiento, pero cambian simultáneamente datos, arquitectura, alineación, cuantización y generación: no identifican un efecto causal de escala.

La inferencia no es auditable. Las tablas marcan significación sin nombrar prueba, alfa, p-values, intervalos ni unidad de análisis. Conversaciones y turnos reutilizan roles, personas, starters, tareas y modelos; requieren pairing y modelos mixtos o errores robustos por cluster. Hay muchas comparaciones sin corrección de multiplicidad y no se informan semillas, réplicas, fallos de parsing ni logs. También hay contradicciones: Sotopia solo para API frente a todas las simulaciones, 10-15 turnos frente a N por 10 rondas con N en 2/3/5 y dyads frente a agentes de tres o cinco miembros. La versión NLTK 1.0.1 indicada no existe en el historial oficial.

La publicación incluye prompts, tablas, ejemplos y checklist, pero el repositorio no reproduce el estudio. En el commit auditado contiene solo un README de una línea y JSON: no hay código, análisis, dependencias, tests, CI, licencia ni manifest. Reúne 29.374 conversaciones/440.610 turnos en cinco carpetas, incluye Mistral pero no Llama 70B ni GPT-5 y no se vincula con las 576/1.270 conversaciones o RQ3/RQ4. Todos los 324 archivos Phi tienen mensajes vacíos y 4.480/8.100 conversaciones afectadas; Qwen tiene vacíos en 193/194 archivos y 1.797/4.850 conversaciones, además de 11.408 turnos con caracteres CJK; GPT-4.1 contiene 64 duplicados exactos por parejas invertidas. El checklist declara ausencia de aprobación/exención ética y de controles documentados sobre información identificable u ofensiva. La conclusión fiel es que prompts jerárquicos cambian algunos rasgos lingüísticos y etiquetas de aquiescencia en estas configuraciones; no se demuestra cognición humana, realismo conductual, personalidad, obediencia real, daño desplegado ni mayor seguridad causal por tamaño, y los resultados no son reproducibles con el artefacto público.

Research question

To what extent do LLM agents assigned to high- and low-status occupational personas show asymmetries in pronouns, linguistic coordination, persuasion, and unsafe compliance consistent in direction with human effects, and how do they vary by turn, prompt, size, and post-training?

Method

Simulation in English with 14 hierarchical pairs and PersonaHub personas. RQ1 uses 576 conversations for FPS/FPP; RQ2 uses 1,270 and eight style markers; RQ3 continues starters from DailyPersuasion; RQ4 uses Do-Not-Answer. GPT-5 judges persuasion/compliance and three humans verify 300 conversations. Temporal analyses, High/Low/No controls, size, and SFT/DPO are added. The audit reviewed 26 pages ACL, 27 pages arXiv, TeX, checklist, repository, and 29,374 published JSON.

Sample: Fourteen pairs of roles and about ten pairs of personas per role; 50 pairs evaluated by three annotators. RQ1: 576 conversations. RQ2: 1,270. RQ3 uses ten starters per domain and discards 3/14 roles, but does not publish a final manifest or a clear n per table. RQ4 also does not publish an exact n. Judge validation: 300 conversations, 50 per task and per each of three models, with three annotators. No seeds or replicates are reported.

Findings

  • It is a long paper published in ACL 2026, not just a preprint.
  • Four of six main rows follow the expected directions of pronouns; Qwen and Phi do not.
  • Llama's pronoun effects are only 0.06-0.17 percentage points.
  • Open models coordinate on approximately 6-7 of eight markers and GPT around four.
  • The coordination asymmetry between statuses is not significant.
  • High status increases judged persuasion by 1.6-6.4 points.
  • High status increases partial/full compliance by 2.0-3.7 points.
  • Human judge accuracy drops to 65.0%-67.7% across three classes.
  • Effects tend to be larger at the start and attenuate, except for open coordination.
  • GPT responds strongly to explicit Low/No prompts; open models less so.
  • Selected comparisons associate larger size with lower persuasion/compliance.
  • No equivalent human baseline exists to validate realism.
  • No significance is audited due to lack of test and p-values.
  • The repository contains no code or analysis and does not match the models/counts in the article.
  • The published corpus has thousands of conversations with empty messages, undocumented CJK, and duplicates.
  • The checklist declares absence of ethical review and controls for identifiable/offensive content.

Limitations

  • English and culturally narrow occupational roles.
  • Contextual hierarchy, not universal power.
  • No equivalent human control.
  • Persuasion as textual appearance, not attitude or behavior.
  • Partial and full compliance merged.
  • GPT-5 judge with substantial error and possible self-preference.
  • No blinding of the judge to status cues.
  • Tests, alpha, and p-values not reported.
  • No intervals or multiplicity correction.
  • Dependence among personas, roles, starters, tasks, and turns not modeled.
  • No seeds, replicates, logs, or parsing failures.
  • Phi model internally inconsistent.
  • GPT snapshots not identified.
  • Sotopia protocol, turns, and number of agents contradictory.
  • Scale confounded with architecture, data, and alignment.
  • NLTK version 1.0.1 nonexistent.
  • Repository without license, documentation, code, environment, tests, or CI.
  • Data without experimental manifest or quality control.
  • No ethical approval or exemption declared.
  • No documented controls for PII or offensiveness.

What the study does not establish

  • Socio-cognitive cognition in LLMs.
  • Stable personality or subjectivity.
  • Behavioral realism relative to humans.
  • Equivalence of magnitude or distribution with human effects.
  • Real obedience to authority.
  • Durable attitude change.
  • Frequency or severity of harm in deployment.
  • Control of latent mechanisms through prompts.
  • That larger models are causally safer.
  • Generalization to other languages, cultures, or hierarchies.
  • Universal validity of the role pairs.
  • Correctness of the reported significances.
  • Independence of conversations and turns.
  • Reproduction of results with the public repository.
  • That the GitHub corpus is the corpus analyzed in the article.

Traceability

Scope: Full text

Version: ACL 2026 long paper, Anthology ID 2026.acl-long.2202, DOI 10.18653/v1/2026.acl-long.2202, 26 pages; arXiv:2605.17694v2 and complete TeX; Responsible NLP Checklist; GitHub data artifact at commit 40fc3f455385b6839a627b226b032ba36a027e9c

Consulted source: https://aclanthology.org/2026.acl-long.2202/

Review: Codex 26-page ACL, 27-page arXiv, complete TeX, checklist, construct, human-grounding, statistics, LLM-judge, repository, data-quality, ethics and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Llama 3.1 8B Instruct
  • Qwen 2.5 7B Instruct
  • Phi-3-Med in methods / Phi-4 in result tables, unresolved identity
  • Llama 3.1 70B Instruct Quantized
  • GPT-4.1, exact snapshot not reported
  • GPT-5 via gpt-5-chat-latest
  • Mistral 7B and 24B variants
  • Qwen 2.5 72B
  • OLMo 2 32B SFT and DPO
  • Gemma 8B in judge validation

Instruments and metrics

  • First-person singular and plural pronoun percentages
  • Eight-marker degree of language coordination D_lc
  • GPT-5 three-class persuasion judge
  • GPT-5 three-class harmful-compliance judge
  • Two-way merged partial/full headline labels
  • Fleiss kappa for persona hierarchy
  • Human judge agreement and accuracy
  • High/Low/No direct control prompts
  • Independent construct, statistics, artifact, ethics and data-quality audit

Data used

  • PersonaHub personas
  • DailyPersuasion conversation starters
  • Do-Not-Answer unsafe prompts
  • DailyDialog human starters
  • Author-generated roles, tasks and synthetic conversations
  • Public GitHub corpus of 29,374 undocumented conversations, not traceably matched to reported experiments

Evidence and location

  • Publication, method, results, prompts, limitations, and ethics: ACL Anthology 2026.acl-long.2202, 26 pages, DOI 10.18653/v1/2026.acl-long.2202, sha256 b762e6fa97f26f7fb6daf9317a2b6759f3e2f57651201303b56c736f4ae2100c
  • Version, complete TeX, and manuscript consistency: arXiv:2605.17694v2; source sha256 096b40463fe21ef75643781d64422cbef6eeb06b9ac446bc505ccd0257796b1b; main TeX sha256 06bda1c17f25385e39762556173ef62bb4d04e30adfa16799248641a72b15270
  • Ethical review, consent, and declared data controls: Responsible NLP Checklist sha256 b4459f6cef13c2a4372a0ce771f578a6f06894f248034b3095aae3f1888db377
  • Repository content, commit, and quality: github.com/nvshrao/power-asymmetric-conversations at 40fc3f455385b6839a627b226b032ba36a027e9c, audited 2026-07-17
  • Complete independent audit: reports/verification/article-327-power-asymmetric-conversation-human-grounding-statistics-judge-data-release-and-reproducibility-audit.json