@inproceedings{hu-etal-2026-dupo,
title = "{DUPO}: Dual-Mode Unified Policy Optimization for Tool-Using Agents",
author = "Hu, Zexin and
Hu, Mu and
Yang, Lijuan and
Wang, Yisen and
Zhu, Xiaogang and
Jiao, KangJian and
Fu, Yu and
Gu, Lihong and
Wang, Zhiyong and
Hu, Kun",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.836/",
pages = "19405--19421",
ISBN = "979-8-89176-459-0",
abstract = "Tool-using large language models (LLMs) have improved at producing executable API calls, but reliable tool use requires coordinating semantic intent with executable invocation, especially in multi-step settings. Existing tool-learning objectives and traces collapse these two roles into a single homogeneous generation stream, so the learning signal cannot distinguish a flawed plan from a flawed call. We propose Dual-Mode Unified Policy Optimization (DUPO), a typed trajectory optimization framework that represents each step with a PLAN span for the current subgoal and a TOOL span for the API invocation. DUPO follows a PLAN-to-TOOL rollout factorization, optimizes the two modes with mode-separated policy ratios, and introduces Chain-Aware Upstream Advantage, so that early planning errors, which propagate through the rest of the trajectory, receive stronger credit than later steps. To support this optimization, we introduce PlanArena, a human-annotated dataset of 20,646 trajectories across 7,941 scenarios with explicit PLAN{--}TOOL supervision. Experiments on BFCL, API-Bank, and Tau2 show that DUPO improves tool-use performance, with gains concentrated in multi-step settings. Our code and dataset will be released upon publication."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hu-etal-2026-dupo">
<titleInfo>
<title>DUPO: Dual-Mode Unified Policy Optimization for Tool-Using Agents</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zexin</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mu</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lijuan</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yisen</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaogang</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">KangJian</namePart>
<namePart type="family">Jiao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yu</namePart>
<namePart type="family">Fu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lihong</namePart>
<namePart type="family">Gu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhiyong</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kun</namePart>
<namePart type="family">Hu</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Tool-using large language models (LLMs) have improved at producing executable API calls, but reliable tool use requires coordinating semantic intent with executable invocation, especially in multi-step settings. Existing tool-learning objectives and traces collapse these two roles into a single homogeneous generation stream, so the learning signal cannot distinguish a flawed plan from a flawed call. We propose Dual-Mode Unified Policy Optimization (DUPO), a typed trajectory optimization framework that represents each step with a PLAN span for the current subgoal and a TOOL span for the API invocation. DUPO follows a PLAN-to-TOOL rollout factorization, optimizes the two modes with mode-separated policy ratios, and introduces Chain-Aware Upstream Advantage, so that early planning errors, which propagate through the rest of the trajectory, receive stronger credit than later steps. To support this optimization, we introduce PlanArena, a human-annotated dataset of 20,646 trajectories across 7,941 scenarios with explicit PLAN–TOOL supervision. Experiments on BFCL, API-Bank, and Tau2 show that DUPO improves tool-use performance, with gains concentrated in multi-step settings. Our code and dataset will be released upon publication.</abstract>
<identifier type="citekey">hu-etal-2026-dupo</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.836/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>19405</start>
<end>19421</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T DUPO: Dual-Mode Unified Policy Optimization for Tool-Using Agents
%A Hu, Zexin
%A Hu, Mu
%A Yang, Lijuan
%A Wang, Yisen
%A Zhu, Xiaogang
%A Jiao, KangJian
%A Fu, Yu
%A Gu, Lihong
%A Wang, Zhiyong
%A Hu, Kun
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F hu-etal-2026-dupo
%X Tool-using large language models (LLMs) have improved at producing executable API calls, but reliable tool use requires coordinating semantic intent with executable invocation, especially in multi-step settings. Existing tool-learning objectives and traces collapse these two roles into a single homogeneous generation stream, so the learning signal cannot distinguish a flawed plan from a flawed call. We propose Dual-Mode Unified Policy Optimization (DUPO), a typed trajectory optimization framework that represents each step with a PLAN span for the current subgoal and a TOOL span for the API invocation. DUPO follows a PLAN-to-TOOL rollout factorization, optimizes the two modes with mode-separated policy ratios, and introduces Chain-Aware Upstream Advantage, so that early planning errors, which propagate through the rest of the trajectory, receive stronger credit than later steps. To support this optimization, we introduce PlanArena, a human-annotated dataset of 20,646 trajectories across 7,941 scenarios with explicit PLAN–TOOL supervision. Experiments on BFCL, API-Bank, and Tau2 show that DUPO improves tool-use performance, with gains concentrated in multi-step settings. Our code and dataset will be released upon publication.
%U https://aclanthology.org/2026.emnlp-main.836/
%P 19405-19421
Markdown (Informal)
[DUPO: Dual-Mode Unified Policy Optimization for Tool-Using Agents](https://aclanthology.org/2026.emnlp-main.836/) (Hu et al., EMNLP 2026)
ACL
- Zexin Hu, Mu Hu, Lijuan Yang, Yisen Wang, Xiaogang Zhu, KangJian Jiao, Yu Fu, Lihong Gu, Zhiyong Wang, and Kun Hu. 2026. DUPO: Dual-Mode Unified Policy Optimization for Tool-Using Agents. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 19405–19421, Budapest, Hungary. Association for Computational Linguistics.