@inproceedings{cheng-etal-2026-evocurr,
title = "{E}vo{C}urr: Self-evolving Curriculum with Behavior Code Generation for Complex Decision-making",
author = "Cheng, Yang and
Ma, Weiyu and
Wang, Zilai and
Zhu, Wenhui and
Lv, Tangjie and
Hu, Yujing and
Elhoseiny, Mohamed and
Deng, Yue and
Zhao, Jian",
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.1649/",
pages = "35089--35124",
ISBN = "979-8-89176-462-0",
abstract = "Complex decision-making often requires agents to progress through intermediate tasks rather than solve the final target directly. Existing LLM self-refinement methods typically iterate on a fixed target, while curriculum-learning methods often rely on hand-designed schedules, training-time optimization, or domain-specific difficulty metrics. To smooth the learning curve through adaptive curriculum design, we introduce EvoCurr, a general inference-time framework that co-evolves curricula and executable policies. A Designer proposes verifiable intermediate tasks from the latest accepted progress and recent failures, while a Solver generates or trains policies for these tasks. A task-policy pair is accepted only when it passes hard feasibility checks and reaches a specified performance threshold; meanwhile, an accepted-floor constraint preserves the latest verified checkpoint and prevents failed harder attempts from overwriting mastered behavior. We evaluate EvoCurr across code-as-policy and closed-loop MARL settings, using StarCraft II super-late-game micromanagement as our hardest stress test. EvoCurr demonstrates effective micromanagement across three LLM backbones without LLM parameter training. Improvements in additional experiments on Flatland, POGEMA, Overcooked, and LiveCodeBench Hard demonstrate the generality of EvoCurr."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="cheng-etal-2026-evocurr">
<titleInfo>
<title>EvoCurr: Self-evolving Curriculum with Behavior Code Generation for Complex Decision-making</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yang</namePart>
<namePart type="family">Cheng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Weiyu</namePart>
<namePart type="family">Ma</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zilai</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenhui</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tangjie</namePart>
<namePart type="family">Lv</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yujing</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohamed</namePart>
<namePart type="family">Elhoseiny</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Deng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jian</namePart>
<namePart type="family">Zhao</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>Complex decision-making often requires agents to progress through intermediate tasks rather than solve the final target directly. Existing LLM self-refinement methods typically iterate on a fixed target, while curriculum-learning methods often rely on hand-designed schedules, training-time optimization, or domain-specific difficulty metrics. To smooth the learning curve through adaptive curriculum design, we introduce EvoCurr, a general inference-time framework that co-evolves curricula and executable policies. A Designer proposes verifiable intermediate tasks from the latest accepted progress and recent failures, while a Solver generates or trains policies for these tasks. A task-policy pair is accepted only when it passes hard feasibility checks and reaches a specified performance threshold; meanwhile, an accepted-floor constraint preserves the latest verified checkpoint and prevents failed harder attempts from overwriting mastered behavior. We evaluate EvoCurr across code-as-policy and closed-loop MARL settings, using StarCraft II super-late-game micromanagement as our hardest stress test. EvoCurr demonstrates effective micromanagement across three LLM backbones without LLM parameter training. Improvements in additional experiments on Flatland, POGEMA, Overcooked, and LiveCodeBench Hard demonstrate the generality of EvoCurr.</abstract>
<identifier type="citekey">cheng-etal-2026-evocurr</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1649/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>35089</start>
<end>35124</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T EvoCurr: Self-evolving Curriculum with Behavior Code Generation for Complex Decision-making
%A Cheng, Yang
%A Ma, Weiyu
%A Wang, Zilai
%A Zhu, Wenhui
%A Lv, Tangjie
%A Hu, Yujing
%A Elhoseiny, Mohamed
%A Deng, Yue
%A Zhao, Jian
%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 cheng-etal-2026-evocurr
%X Complex decision-making often requires agents to progress through intermediate tasks rather than solve the final target directly. Existing LLM self-refinement methods typically iterate on a fixed target, while curriculum-learning methods often rely on hand-designed schedules, training-time optimization, or domain-specific difficulty metrics. To smooth the learning curve through adaptive curriculum design, we introduce EvoCurr, a general inference-time framework that co-evolves curricula and executable policies. A Designer proposes verifiable intermediate tasks from the latest accepted progress and recent failures, while a Solver generates or trains policies for these tasks. A task-policy pair is accepted only when it passes hard feasibility checks and reaches a specified performance threshold; meanwhile, an accepted-floor constraint preserves the latest verified checkpoint and prevents failed harder attempts from overwriting mastered behavior. We evaluate EvoCurr across code-as-policy and closed-loop MARL settings, using StarCraft II super-late-game micromanagement as our hardest stress test. EvoCurr demonstrates effective micromanagement across three LLM backbones without LLM parameter training. Improvements in additional experiments on Flatland, POGEMA, Overcooked, and LiveCodeBench Hard demonstrate the generality of EvoCurr.
%U https://aclanthology.org/2026.findings-emnlp.1649/
%P 35089-35124
Markdown (Informal)
[EvoCurr: Self-evolving Curriculum with Behavior Code Generation for Complex Decision-making](https://aclanthology.org/2026.findings-emnlp.1649/) (Cheng et al., Findings 2026)
ACL
- Yang Cheng, Weiyu Ma, Zilai Wang, Wenhui Zhu, Tangjie Lv, Yujing Hu, Mohamed Elhoseiny, Yue Deng, and Jian Zhao. 2026. EvoCurr: Self-evolving Curriculum with Behavior Code Generation for Complex Decision-making. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 35089–35124, Budapest, Hungary. Association for Computational Linguistics.