arXiv cs.CL· Ziyi Zhu, Olivier Tieleman, Caitlin A. Stamatis, Luka Smyth, Thomas D. Hull, Daniel R. Cahn, Jinghong Chen, Matteo Malgaroli·· 11 小时前AI 评分29
DIAL：用直接迭代对抗学习实现逼真的多轮对话模拟
DIAL: Direct Iterative Adversarial Learning for Realistic Multi-Turn Dialogue Simulation
AI 导读
研究者提出 Direct Iterative Adversarial Learning（DIAL），通过生成器（用户模拟器）与判别器的对抗动态迭代提升用户模拟器逼真度。在心理健康支持场景中，DIAL 恢复了监督微调削弱的词汇多样性，并大幅降低判别器准确率，使模拟失败发生率与真实失败发生率高度相关，同时保持较低的失败模式分布散度。
正文 · 原文
Abstract:Realistic user simulation is crucial for training and evaluating multi-turn dialogue systems, yet creating simulators that accurately replicate human behavior remains a significant challenge. An effective simulator must expose the failure modes of the systems under evaluation. This work introduces Direct Iterative Adversarial Learning (DIAL), an adversarial framework that iteratively enhances user simulator realism through a competitive dynamic between a generator (user simulator) and a discriminator. When applied to mental health support, a domain characterized by diverse failure types and a critical dependence on realistic user behavior for failure detection, DIAL restores lexical diversity diminished by supervised fine-tuning and drastically reduces discriminator accuracy. The resulting simulator exhibits a strong correlation between simulated and real failure occurrence rates while maintaining low distributional divergence of failure modes. These findings indicate that DIAL is a promising method for developing realistic user simulators in multi-turn dialogue, facilitating reliable and cost-effective system evaluation prior to deployment.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2512.20773 [cs.CL] |
| (or arXiv:2512.20773v5 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2512.20773 arXiv-issued DOI via DataCite |
|
| Journal reference: | Findings of the Association for Computational Linguistics: EMNLP 2026, pages 925-942, Budapest, Hungary, October 2026 |
| Related DOI: | https://doi.org/10.18653/2026.findings-emnlp.46
DOI(s) linking to related resources |
Submission history
From: Ziyi Zhu [view email]
[v1]
Tue, 23 Dec 2025 21:21:08 UTC (241 KB)
[v2]
Mon, 5 Jan 2026 11:37:37 UTC (247 KB)
[v3]
Wed, 18 Feb 2026 18:41:36 UTC (244 KB)
[v4]
Wed, 6 May 2026 14:26:15 UTC (196 KB)
[v5]
Tue, 3 Nov 2026 21:09:41 UTC (204 KB)
来源：arXiv cs.CL · arxiv.org