@inproceedings{fang-etal-2026-llmzero,
title = "{LLMZERO}: Discovering Adaptive Training Strategies for {RL} Post-Training via {LLM} Agents",
author = "Fang, Haoyang and
Zhu, Wei and
Han, Boran and
Pan, Zhenyu and
Yang, Shuo and
Zhang, Shuai and
Gai, Jiading and
Tang, Peng and
Hu, Cuixiong and
Rangwala, Huzefa and
Karypis, George and
Wang, Bernie",
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.1866/",
pages = "39827--39856",
ISBN = "979-8-89176-462-0",
abstract = "RL post-training strategies are dataset-dependent and reveal a recurring empirical pattern: capacity parameters accumulate monotonically across stages, while regularization parameters predominantly oscillate in response to shifting training dynamics. This distinction highlights a potential flaw in fixed training schedules: by forcing all parameters along rigid paths, they fail to capture the dynamic exploration-exploitation tradeoffs that regularization must track. We uncover this through LLMZero, an agentic system that optimizes training trajectories via tree search by diagnosing pathologies at each checkpoint and proposing coordinated multi-parameter transitions. Across four diverse GRPO tasks, LLMZero discovers strategies that improve over the base model by 9{\%} to 140{\%} and over grid search by 6{\%} to 15{\%} (relative), consistently outperforming random search and a skill-based agent under a matched compute budget. The capacity{--}regularization asymmetry is consistent across all four tasks, offering a candidate design heuristic for multi-stage training."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="fang-etal-2026-llmzero">
<titleInfo>
<title>LLMZERO: Discovering Adaptive Training Strategies for RL Post-Training via LLM Agents</title>
</titleInfo>
<name type="personal">
<namePart type="given">Haoyang</namePart>
<namePart type="family">Fang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Boran</namePart>
<namePart type="family">Han</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhenyu</namePart>
<namePart type="family">Pan</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">Shuai</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiading</namePart>
<namePart type="family">Gai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Peng</namePart>
<namePart type="family">Tang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cuixiong</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Huzefa</namePart>
<namePart type="family">Rangwala</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">George</namePart>
<namePart type="family">Karypis</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bernie</namePart>
<namePart type="family">Wang</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>RL post-training strategies are dataset-dependent and reveal a recurring empirical pattern: capacity parameters accumulate monotonically across stages, while regularization parameters predominantly oscillate in response to shifting training dynamics. This distinction highlights a potential flaw in fixed training schedules: by forcing all parameters along rigid paths, they fail to capture the dynamic exploration-exploitation tradeoffs that regularization must track. We uncover this through LLMZero, an agentic system that optimizes training trajectories via tree search by diagnosing pathologies at each checkpoint and proposing coordinated multi-parameter transitions. Across four diverse GRPO tasks, LLMZero discovers strategies that improve over the base model by 9% to 140% and over grid search by 6% to 15% (relative), consistently outperforming random search and a skill-based agent under a matched compute budget. The capacity–regularization asymmetry is consistent across all four tasks, offering a candidate design heuristic for multi-stage training.</abstract>
<identifier type="citekey">fang-etal-2026-llmzero</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1866/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>39827</start>
<end>39856</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T LLMZERO: Discovering Adaptive Training Strategies for RL Post-Training via LLM Agents
%A Fang, Haoyang
%A Zhu, Wei
%A Han, Boran
%A Pan, Zhenyu
%A Yang, Shuo
%A Zhang, Shuai
%A Gai, Jiading
%A Tang, Peng
%A Hu, Cuixiong
%A Rangwala, Huzefa
%A Karypis, George
%A Wang, Bernie
%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 fang-etal-2026-llmzero
%X RL post-training strategies are dataset-dependent and reveal a recurring empirical pattern: capacity parameters accumulate monotonically across stages, while regularization parameters predominantly oscillate in response to shifting training dynamics. This distinction highlights a potential flaw in fixed training schedules: by forcing all parameters along rigid paths, they fail to capture the dynamic exploration-exploitation tradeoffs that regularization must track. We uncover this through LLMZero, an agentic system that optimizes training trajectories via tree search by diagnosing pathologies at each checkpoint and proposing coordinated multi-parameter transitions. Across four diverse GRPO tasks, LLMZero discovers strategies that improve over the base model by 9% to 140% and over grid search by 6% to 15% (relative), consistently outperforming random search and a skill-based agent under a matched compute budget. The capacity–regularization asymmetry is consistent across all four tasks, offering a candidate design heuristic for multi-stage training.
%U https://aclanthology.org/2026.findings-emnlp.1866/
%P 39827-39856
Markdown (Informal)
[LLMZERO: Discovering Adaptive Training Strategies for RL Post-Training via LLM Agents](https://aclanthology.org/2026.findings-emnlp.1866/) (Fang et al., Findings 2026)
ACL
- Haoyang Fang, Wei Zhu, Boran Han, Zhenyu Pan, Shuo Yang, Shuai Zhang, Jiading Gai, Peng Tang, Cuixiong Hu, Huzefa Rangwala, George Karypis, and Bernie Wang. 2026. LLMZERO: Discovering Adaptive Training Strategies for RL Post-Training via LLM Agents. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 39827–39856, Budapest, Hungary. Association for Computational Linguistics.