Enjoy Your Talk: A Human-Centered Benchmark for Multi-Turn Dialogue with Decoupled User Simulation, Target Modeling, and Judging

Personas, identity, and agents2026arXivApproved editorial review

Authors: Jinglan Gong, Jiefan Lu, Hewei Guo, Kehan Li, Zhiyuan Han, Jihang Jiang, Wenwen Tong, Lewei Lu

Keywords: Persona fidelity, User simulation, Multi-turn dialogue

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

8
Authors
4
Findings
5
Limitations
3
Evidence

Editorial summary

English

Seventeen models complete 200 dialogues each, 100 with Nemotron Personas USA and 100 with PersonaMem-v2. The target separates label prediction and generation; Gemini 3.1 Pro Thinking judges subjective metrics and FICR, embedding drift is computed, and a subset is repeated with DeepSeek-V4-Pro.

3,400 target dialogues and more than 30,000 turns. Cross-judge checking uses 46 dialogues and 11 cells; the human pilot covers 59 turn-cells with five annotators. Latent-intent accuracy on PersonaMem ranged from .09 to .80. Enabling reasoning in Gemma-4 increased that accuracy by .466 and .497 by condition. FICR saturated on Nemotron and ranged from .53 to .88 on PersonaMem. The warm-up effect appeared in 16 of 17 models, with GPT-5.5 reversing it.

A simulator and the main judge share the Gemini 3.1 family. The reasoning claim depends on one paired family. Full human validation is still pending. The cross-judge comparison is small. Only English and mainly United States profiles are evaluated. Human-centered does not mean human-validated. It does not demonstrate fidelity to specific real people. It does not establish that small subjective differences are significant.

Español

Diecisiete modelos realizan 200 diálogos cada uno, 100 con Nemotron Personas USA y 100 con PersonaMem-v2. El objetivo separa predicción de etiquetas y generación; Gemini 3.1 Pro Thinking juzga métricas subjetivas y FICR, se calcula drift con embeddings y una submuestra se repite con DeepSeek-V4-Pro.

3.400 diálogos objetivo, más de 30.000 turnos. La comprobación entre jueces usa 46 diálogos y 11 celdas; el piloto humano cubre 59 celdas-turno con cinco anotadores. La precisión de intención latente en PersonaMem varió de .09 a .80. Activar razonamiento en Gemma-4 elevó esa precisión en .466 y .497 según condición. FICR quedó saturada en Nemotron y varió de .53 a .88 en PersonaMem. El warm-up apareció en 16 de 17 modelos, con GPT-5.5 en sentido inverso.

Un simulador y el juez principal comparten familia Gemini 3.1. La afirmación sobre razonamiento depende de una sola familia emparejada. La validación humana completa está pendiente. La comparación entre jueces es pequeña. Solo se evalúan inglés y perfiles principalmente estadounidenses. Human-centered no significa human-validated. No demuestra fidelidad a personas reales concretas. No establece que diferencias subjetivas pequeñas sean significativas.

Research question

Which models maintain persona, intent, and emotional dynamics in long dialogues when simulator, target, and judge are decoupled?

Method

Seventeen models complete 200 dialogues each, 100 with Nemotron Personas USA and 100 with PersonaMem-v2. The target separates label prediction and generation; Gemini 3.1 Pro Thinking judges subjective metrics and FICR, embedding drift is computed, and a subset is repeated with DeepSeek-V4-Pro.

Sample: 3,400 target dialogues and more than 30,000 turns. Cross-judge checking uses 46 dialogues and 11 cells; the human pilot covers 59 turn-cells with five annotators.

Findings

  • Latent-intent accuracy on PersonaMem ranged from .09 to .80.
  • Enabling reasoning in Gemma-4 increased that accuracy by .466 and .497 by condition.
  • FICR saturated on Nemotron and ranged from .53 to .88 on PersonaMem.
  • The warm-up effect appeared in 16 of 17 models, with GPT-5.5 reversing it.

Limitations

  • A simulator and the main judge share the Gemini 3.1 family.
  • The reasoning claim depends on one paired family.
  • Full human validation is still pending.
  • The cross-judge comparison is small.
  • Only English and mainly United States profiles are evaluated.

What the study does not establish

  • Human-centered does not mean human-validated.
  • It does not demonstrate fidelity to specific real people.
  • It does not establish that small subjective differences are significant.

Traceability

Scope: Full text

Version: arxiv; 16-page full text reviewed 2026-07-18

Consulted source: https://arxiv.org/abs/2607.10428

Review: Codex full-text and visual 16-page methodological, statistical and claim-boundary review, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • 17 target models
  • Gemini 3.1 Pro Thinking judge
  • DeepSeek-V4-Pro judge

Instruments and metrics

  • EYT-Bench
  • FICR
  • Embedding intent drift
  • Persona and empathy ratings

Data used

  • Nemotron-Personas-USA
  • PersonaMem-v2

Evidence and location

  • Research question, method, results, and discussion: Full text, pp. 1-16, visually reviewed on 18/07/2026
  • Figures, tables, results, and limitations: Primary PDF sha256 7b102b766bcf84338a990338055a9dd45698085df064b2d57ab424a146592b0c; methods, results, limitations, and appendices
  • Editorial decision and claim boundary: Critical record article-409, complete cross-check of 16 pages