@inproceedings{shi-etal-2026-grail,
title = "{GRAIL}: Gradient-Reward Aligned Initialization for {L}o{RA}-based {GRPO}",
author = "Shi, Liwen and
Zeng, Zijie and
Wang, Bowen and
Yunji and
Chen, Zeyu and
Shao, Xinyu and
Sun, Xinzheng and
Xiong, Xuyang and
Wu, Haoyang and
Zhang, Sheng",
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.926/",
pages = "19549--19565",
ISBN = "979-8-89176-462-0",
abstract = "Group Relative Policy Optimization (GRPO) improves LLM reasoning, but LoRA-based GRPO can underperform when the initialized adapter subspace is poorly aligned with reward-induced gradients. We propose GRAIL, Gradient-Reward Aligned Initialization for LoRA-based GRPO, which estimates reward-aligned gradients from grouped rollouts before policy optimization and uses them to construct LoRA factors. The full GRAIL configuration includes a reference-solution fallback only for all-zero-correctness groups during calibration; after initialization, all models use the standard GRPO objective without extra trainable parameters or loss changes. We also study Degenerate-Group Reflection Resampling (DGRR) as an optional extension that resamples all-zero-correctness groups after self-reflection without exposing reference solutions. Across Qwen2.5, Phi-3.5, and Gemma-2 backbones on GSM8K and MATH500, GRAIL consistently improves LoRA-based GRPO, with particularly strong gains on smaller Qwen and Gemma models. Analyses attribute these gains to stronger GRPO-gradient capture, smaller principal angles, and better handling of correctness-degenerate groups."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shi-etal-2026-grail">
<titleInfo>
<title>GRAIL: Gradient-Reward Aligned Initialization for LoRA-based GRPO</title>
</titleInfo>
<name type="personal">
<namePart type="given">Liwen</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zijie</namePart>
<namePart type="family">Zeng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bowen</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name>
<namePart>Yunji</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zeyu</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinyu</namePart>
<namePart type="family">Shao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinzheng</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuyang</namePart>
<namePart type="family">Xiong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haoyang</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sheng</namePart>
<namePart type="family">Zhang</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>Group Relative Policy Optimization (GRPO) improves LLM reasoning, but LoRA-based GRPO can underperform when the initialized adapter subspace is poorly aligned with reward-induced gradients. We propose GRAIL, Gradient-Reward Aligned Initialization for LoRA-based GRPO, which estimates reward-aligned gradients from grouped rollouts before policy optimization and uses them to construct LoRA factors. The full GRAIL configuration includes a reference-solution fallback only for all-zero-correctness groups during calibration; after initialization, all models use the standard GRPO objective without extra trainable parameters or loss changes. We also study Degenerate-Group Reflection Resampling (DGRR) as an optional extension that resamples all-zero-correctness groups after self-reflection without exposing reference solutions. Across Qwen2.5, Phi-3.5, and Gemma-2 backbones on GSM8K and MATH500, GRAIL consistently improves LoRA-based GRPO, with particularly strong gains on smaller Qwen and Gemma models. Analyses attribute these gains to stronger GRPO-gradient capture, smaller principal angles, and better handling of correctness-degenerate groups.</abstract>
<identifier type="citekey">shi-etal-2026-grail</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.926/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>19549</start>
<end>19565</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T GRAIL: Gradient-Reward Aligned Initialization for LoRA-based GRPO
%A Shi, Liwen
%A Zeng, Zijie
%A Wang, Bowen
%A Chen, Zeyu
%A Shao, Xinyu
%A Sun, Xinzheng
%A Xiong, Xuyang
%A Wu, Haoyang
%A Zhang, Sheng
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%A Yunji
%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 shi-etal-2026-grail
%X Group Relative Policy Optimization (GRPO) improves LLM reasoning, but LoRA-based GRPO can underperform when the initialized adapter subspace is poorly aligned with reward-induced gradients. We propose GRAIL, Gradient-Reward Aligned Initialization for LoRA-based GRPO, which estimates reward-aligned gradients from grouped rollouts before policy optimization and uses them to construct LoRA factors. The full GRAIL configuration includes a reference-solution fallback only for all-zero-correctness groups during calibration; after initialization, all models use the standard GRPO objective without extra trainable parameters or loss changes. We also study Degenerate-Group Reflection Resampling (DGRR) as an optional extension that resamples all-zero-correctness groups after self-reflection without exposing reference solutions. Across Qwen2.5, Phi-3.5, and Gemma-2 backbones on GSM8K and MATH500, GRAIL consistently improves LoRA-based GRPO, with particularly strong gains on smaller Qwen and Gemma models. Analyses attribute these gains to stronger GRPO-gradient capture, smaller principal angles, and better handling of correctness-degenerate groups.
%U https://aclanthology.org/2026.findings-emnlp.926/
%P 19549-19565
Markdown (Informal)
[GRAIL: Gradient-Reward Aligned Initialization for LoRA-based GRPO](https://aclanthology.org/2026.findings-emnlp.926/) (Shi et al., Findings 2026)
ACL
- Liwen Shi, Zijie Zeng, Bowen Wang, Yunji, Zeyu Chen, Xinyu Shao, Xinzheng Sun, Xuyang Xiong, Haoyang Wu, and Sheng Zhang. 2026. GRAIL: Gradient-Reward Aligned Initialization for LoRA-based GRPO. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 19549–19565, Budapest, Hungary. Association for Computational Linguistics.