@inproceedings{cao-etal-2026-disentangled,
title = "Disentangled Privileged States for Simulator-Assisted Multi-Turn Dialogue Policy Training",
author = "Cao, Zhiyong and
Liu, Dunqiang and
Zhou, Sicheng and
Cong, Lin",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1852/",
pages = "39543--39564",
ISBN = "979-8-89176-462-0",
abstract = "Simulator-assisted reinforcement learning is a practical way to train multi-turn dialogue agents, but policy optimization can be misled when simulated users provide inconsistent feedback across long interaction trajectories. In this paper, we focus on a simulator-side question: can structured latent state tracking improve measured simulator consistency and provide useful process supervision for dialogue policy training? Motivated by a standard model-based RL simulation-bound perspective, we introduce **DPS**, a framework that decouples task-conditioned privileged state tracking from surface utterance realization. These privileged states are hidden from the agent but used during training to derive dense, state-aware process rewards that complement sparse outcome signals. We also construct RecruitmentBench, a recruitment persuasion dialogue benchmark grounded in anonymized real-world logs. Experiments on RecruitmentBench and two public datasets use offline real-log-grounded and cross-simulator protocols to evaluate intrinsic simulator consistency and downstream simulator-assisted policy performance. Results indicate that DPS improves measured simulator fidelity and yields stronger fine-tuned policy performance in these controlled evaluation settings, while remaining competitive with strong prompt-based LLM agents. We further calibrate our LLM-based evaluator against human judgments on MATH-CHAT. Code is available at: https://github.com/TraceLite-AI/DPS."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="cao-etal-2026-disentangled">
<titleInfo>
<title>Disentangled Privileged States for Simulator-Assisted Multi-Turn Dialogue Policy Training</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zhiyong</namePart>
<namePart type="family">Cao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dunqiang</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sicheng</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lin</namePart>
<namePart type="family">Cong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Simulator-assisted reinforcement learning is a practical way to train multi-turn dialogue agents, but policy optimization can be misled when simulated users provide inconsistent feedback across long interaction trajectories. In this paper, we focus on a simulator-side question: can structured latent state tracking improve measured simulator consistency and provide useful process supervision for dialogue policy training? Motivated by a standard model-based RL simulation-bound perspective, we introduce **DPS**, a framework that decouples task-conditioned privileged state tracking from surface utterance realization. These privileged states are hidden from the agent but used during training to derive dense, state-aware process rewards that complement sparse outcome signals. We also construct RecruitmentBench, a recruitment persuasion dialogue benchmark grounded in anonymized real-world logs. Experiments on RecruitmentBench and two public datasets use offline real-log-grounded and cross-simulator protocols to evaluate intrinsic simulator consistency and downstream simulator-assisted policy performance. Results indicate that DPS improves measured simulator fidelity and yields stronger fine-tuned policy performance in these controlled evaluation settings, while remaining competitive with strong prompt-based LLM agents. We further calibrate our LLM-based evaluator against human judgments on MATH-CHAT. Code is available at: https://github.com/TraceLite-AI/DPS.</abstract>
<identifier type="citekey">cao-etal-2026-disentangled</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1852/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>39543</start>
<end>39564</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Disentangled Privileged States for Simulator-Assisted Multi-Turn Dialogue Policy Training
%A Cao, Zhiyong
%A Liu, Dunqiang
%A Zhou, Sicheng
%A Cong, Lin
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F cao-etal-2026-disentangled
%X Simulator-assisted reinforcement learning is a practical way to train multi-turn dialogue agents, but policy optimization can be misled when simulated users provide inconsistent feedback across long interaction trajectories. In this paper, we focus on a simulator-side question: can structured latent state tracking improve measured simulator consistency and provide useful process supervision for dialogue policy training? Motivated by a standard model-based RL simulation-bound perspective, we introduce **DPS**, a framework that decouples task-conditioned privileged state tracking from surface utterance realization. These privileged states are hidden from the agent but used during training to derive dense, state-aware process rewards that complement sparse outcome signals. We also construct RecruitmentBench, a recruitment persuasion dialogue benchmark grounded in anonymized real-world logs. Experiments on RecruitmentBench and two public datasets use offline real-log-grounded and cross-simulator protocols to evaluate intrinsic simulator consistency and downstream simulator-assisted policy performance. Results indicate that DPS improves measured simulator fidelity and yields stronger fine-tuned policy performance in these controlled evaluation settings, while remaining competitive with strong prompt-based LLM agents. We further calibrate our LLM-based evaluator against human judgments on MATH-CHAT. Code is available at: https://github.com/TraceLite-AI/DPS.
%U https://aclanthology.org/2026.findings-emnlp.1852/
%P 39543-39564
Markdown (Informal)
[Disentangled Privileged States for Simulator-Assisted Multi-Turn Dialogue Policy Training](https://aclanthology.org/2026.findings-emnlp.1852/) (Cao et al., Findings 2026)
ACL