@inproceedings{wu-etal-2026-stark,
title = "{STARK}: Strategic Advantage Reinforcement with Confidence-Driven Exploration for Long-Horizon Agentic Learning",
author = "Wu, Jinyang and
Chang, Xinan and
Shen, Yuhao and
Lu, Zhengxi and
Yang, Shuo and
Feng, Mingkuan and
Yi, Jiangyan and
Zhang, Shuai and
Lian, Zheng and
Wen, Zhengqi and
Tao, Jianhua",
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.126/",
pages = "2902--2928",
ISBN = "979-8-89176-459-0",
abstract = "Reinforcement learning (RL) for long-horizon agents faces two key challenges. First, parallel linear rollouts tend to converge toward homogeneous reasoning patterns, leading to cognitive-mode collapse. Second, existing methods refine credit assignment along only a single dimension, making it difficult to simultaneously achieve state-aligned comparison and meta-reasoning discrimination. We propose STARK (**S**trategic and **T**actical **A**dvantage **R**einforcement via Confidence Guided **K**eypoint Exploration) to resolve both issues. STARK performs corrective exploration through confidence-driven branching at uncertain decision points and injects reflection into new branches. To empower these keypoint-expanded trajectories, STARK introduces a dual-track strategic and tactical advantage estimator: The strategic track compares reasoning modes under shared environmental observations, while the tactical track evaluates execution quality within each meta-reasoning tag group. Extensive experiments on ALFWorld, ScienceWorld, and WebShop show that STARK consistently outperforms strong baselines across model scales and task domains, and enables a 1.5B model to surpass GPT-5 on the out-of-domain settings."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wu-etal-2026-stark">
<titleInfo>
<title>STARK: Strategic Advantage Reinforcement with Confidence-Driven Exploration for Long-Horizon Agentic Learning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jinyang</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinan</namePart>
<namePart type="family">Chang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuhao</namePart>
<namePart type="family">Shen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhengxi</namePart>
<namePart type="family">Lu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shuo</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mingkuan</namePart>
<namePart type="family">Feng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiangyan</namePart>
<namePart type="family">Yi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shuai</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zheng</namePart>
<namePart type="family">Lian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhengqi</namePart>
<namePart type="family">Wen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jianhua</namePart>
<namePart type="family">Tao</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>Reinforcement learning (RL) for long-horizon agents faces two key challenges. First, parallel linear rollouts tend to converge toward homogeneous reasoning patterns, leading to cognitive-mode collapse. Second, existing methods refine credit assignment along only a single dimension, making it difficult to simultaneously achieve state-aligned comparison and meta-reasoning discrimination. We propose STARK (**S**trategic and **T**actical **A**dvantage **R**einforcement via Confidence Guided **K**eypoint Exploration) to resolve both issues. STARK performs corrective exploration through confidence-driven branching at uncertain decision points and injects reflection into new branches. To empower these keypoint-expanded trajectories, STARK introduces a dual-track strategic and tactical advantage estimator: The strategic track compares reasoning modes under shared environmental observations, while the tactical track evaluates execution quality within each meta-reasoning tag group. Extensive experiments on ALFWorld, ScienceWorld, and WebShop show that STARK consistently outperforms strong baselines across model scales and task domains, and enables a 1.5B model to surpass GPT-5 on the out-of-domain settings.</abstract>
<identifier type="citekey">wu-etal-2026-stark</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.126/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>2902</start>
<end>2928</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T STARK: Strategic Advantage Reinforcement with Confidence-Driven Exploration for Long-Horizon Agentic Learning
%A Wu, Jinyang
%A Chang, Xinan
%A Shen, Yuhao
%A Lu, Zhengxi
%A Yang, Shuo
%A Feng, Mingkuan
%A Yi, Jiangyan
%A Zhang, Shuai
%A Lian, Zheng
%A Wen, Zhengqi
%A Tao, Jianhua
%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 wu-etal-2026-stark
%X Reinforcement learning (RL) for long-horizon agents faces two key challenges. First, parallel linear rollouts tend to converge toward homogeneous reasoning patterns, leading to cognitive-mode collapse. Second, existing methods refine credit assignment along only a single dimension, making it difficult to simultaneously achieve state-aligned comparison and meta-reasoning discrimination. We propose STARK (**S**trategic and **T**actical **A**dvantage **R**einforcement via Confidence Guided **K**eypoint Exploration) to resolve both issues. STARK performs corrective exploration through confidence-driven branching at uncertain decision points and injects reflection into new branches. To empower these keypoint-expanded trajectories, STARK introduces a dual-track strategic and tactical advantage estimator: The strategic track compares reasoning modes under shared environmental observations, while the tactical track evaluates execution quality within each meta-reasoning tag group. Extensive experiments on ALFWorld, ScienceWorld, and WebShop show that STARK consistently outperforms strong baselines across model scales and task domains, and enables a 1.5B model to surpass GPT-5 on the out-of-domain settings.
%U https://aclanthology.org/2026.emnlp-main.126/
%P 2902-2928
Markdown (Informal)
[STARK: Strategic Advantage Reinforcement with Confidence-Driven Exploration for Long-Horizon Agentic Learning](https://aclanthology.org/2026.emnlp-main.126/) (Wu et al., EMNLP 2026)
ACL
- Jinyang Wu, Xinan Chang, Yuhao Shen, Zhengxi Lu, Shuo Yang, Mingkuan Feng, Jiangyan Yi, Shuai Zhang, Zheng Lian, Zhengqi Wen, and Jianhua Tao. 2026. STARK: Strategic Advantage Reinforcement with Confidence-Driven Exploration for Long-Horizon Agentic Learning. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 2902–2928, Budapest, Hungary. Association for Computational Linguistics.