@inproceedings{zhu-etal-2026-ce,
title = "{CE}-{GAD}: Curriculum-Efficient On-Policy Adversarial Distillation of Large Language Models",
author = "Zhu, Wenhui and
Sang, Hejian and
Chen, Xiwen and
Wang, Zhipeng and
Mohajer Hamidi, Shayan and
Yu, Han and
Wang, Jingjing and
Xu, Yuanda and
Dong, Xuanzhao and
Zhou, Zhengze and
He, Ran and
Markovic-Voronov, Jelena and
Behdin, Kayhan and
Ghosh, Sayan and
Geramifard, Alborz",
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.843/",
pages = "17827--17836",
ISBN = "979-8-89176-462-0",
abstract = "Generative adversarial distillation (GAD) provides on-policy feedback for black-box distillation by training a discriminator that ranks teacher and student responses to the same prompt. While effective, GAD is also costly: every step pays for a discriminator update, an actor update over the full response, and the indirect cost of recovering whenever the discriminator and the actor progress at incompatible rates. We argue that the right way to make GAD efficient is not to make individual updates cheaper but to decide **when each update is worth its cost**. We introduce **curriculum-efficient GAD (CE-GAD)**, an adaptive recipe that monitors a small set of training signals and uses them to:\par - (i) skip discriminator updates that would not change the reward, - (ii) constrain actor updates and prune per-token gradients to the subset that actually carries signal, and - (iii) rescale rewards and normalize advantages to preserve within-batch contrast under saturation.\par Evaluations across four models from two open-source families (Qwen2.5 at 3B/7B/14B and Llama-3.2-3B) demonstrate that CE-GAD matches or surpasses vanilla GAD on quality while winning on 12 out of 16 individual benchmark configurations. Specifically, on Qwen2.5-7B, CE-GAD improves the average score by up to **+3{\%}** relative compared to the base model, while reducing per-step wall-clock time by up to **8x** relative to vanilla GAD."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhu-etal-2026-ce">
<titleInfo>
<title>CE-GAD: Curriculum-Efficient On-Policy Adversarial Distillation of Large Language Models</title>
</titleInfo>
<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">Hejian</namePart>
<namePart type="family">Sang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiwen</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhipeng</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shayan</namePart>
<namePart type="family">Mohajer Hamidi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Han</namePart>
<namePart type="family">Yu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jingjing</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuanda</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuanzhao</namePart>
<namePart type="family">Dong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhengze</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ran</namePart>
<namePart type="family">He</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jelena</namePart>
<namePart type="family">Markovic-Voronov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kayhan</namePart>
<namePart type="family">Behdin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sayan</namePart>
<namePart type="family">Ghosh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alborz</namePart>
<namePart type="family">Geramifard</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>Generative adversarial distillation (GAD) provides on-policy feedback for black-box distillation by training a discriminator that ranks teacher and student responses to the same prompt. While effective, GAD is also costly: every step pays for a discriminator update, an actor update over the full response, and the indirect cost of recovering whenever the discriminator and the actor progress at incompatible rates. We argue that the right way to make GAD efficient is not to make individual updates cheaper but to decide **when each update is worth its cost**. We introduce **curriculum-efficient GAD (CE-GAD)**, an adaptive recipe that monitors a small set of training signals and uses them to:\par - (i) skip discriminator updates that would not change the reward, - (ii) constrain actor updates and prune per-token gradients to the subset that actually carries signal, and - (iii) rescale rewards and normalize advantages to preserve within-batch contrast under saturation.\par Evaluations across four models from two open-source families (Qwen2.5 at 3B/7B/14B and Llama-3.2-3B) demonstrate that CE-GAD matches or surpasses vanilla GAD on quality while winning on 12 out of 16 individual benchmark configurations. Specifically, on Qwen2.5-7B, CE-GAD improves the average score by up to **+3%** relative compared to the base model, while reducing per-step wall-clock time by up to **8x** relative to vanilla GAD.</abstract>
<identifier type="citekey">zhu-etal-2026-ce</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.843/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>17827</start>
<end>17836</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T CE-GAD: Curriculum-Efficient On-Policy Adversarial Distillation of Large Language Models
%A Zhu, Wenhui
%A Sang, Hejian
%A Chen, Xiwen
%A Wang, Zhipeng
%A Mohajer Hamidi, Shayan
%A Yu, Han
%A Wang, Jingjing
%A Xu, Yuanda
%A Dong, Xuanzhao
%A Zhou, Zhengze
%A He, Ran
%A Markovic-Voronov, Jelena
%A Behdin, Kayhan
%A Ghosh, Sayan
%A Geramifard, Alborz
%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 zhu-etal-2026-ce
%X Generative adversarial distillation (GAD) provides on-policy feedback for black-box distillation by training a discriminator that ranks teacher and student responses to the same prompt. While effective, GAD is also costly: every step pays for a discriminator update, an actor update over the full response, and the indirect cost of recovering whenever the discriminator and the actor progress at incompatible rates. We argue that the right way to make GAD efficient is not to make individual updates cheaper but to decide **when each update is worth its cost**. We introduce **curriculum-efficient GAD (CE-GAD)**, an adaptive recipe that monitors a small set of training signals and uses them to:\par - (i) skip discriminator updates that would not change the reward, - (ii) constrain actor updates and prune per-token gradients to the subset that actually carries signal, and - (iii) rescale rewards and normalize advantages to preserve within-batch contrast under saturation.\par Evaluations across four models from two open-source families (Qwen2.5 at 3B/7B/14B and Llama-3.2-3B) demonstrate that CE-GAD matches or surpasses vanilla GAD on quality while winning on 12 out of 16 individual benchmark configurations. Specifically, on Qwen2.5-7B, CE-GAD improves the average score by up to **+3%** relative compared to the base model, while reducing per-step wall-clock time by up to **8x** relative to vanilla GAD.
%U https://aclanthology.org/2026.findings-emnlp.843/
%P 17827-17836
Markdown (Informal)
[CE-GAD: Curriculum-Efficient On-Policy Adversarial Distillation of Large Language Models](https://aclanthology.org/2026.findings-emnlp.843/) (Zhu et al., Findings 2026)
ACL
- Wenhui Zhu, Hejian Sang, Xiwen Chen, Zhipeng Wang, Shayan Mohajer Hamidi, Han Yu, Jingjing Wang, Yuanda Xu, Xuanzhao Dong, Zhengze Zhou, Ran He, Jelena Markovic-Voronov, Kayhan Behdin, Sayan Ghosh, and Alborz Geramifard. 2026. CE-GAD: Curriculum-Efficient On-Policy Adversarial Distillation of Large Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 17827–17836, Budapest, Hungary. Association for Computational Linguistics.