@inproceedings{wang-etal-2026-apex,
title = "{APEX} : Adaptively Balancing Off-Policy Guidance and On-Policy Exploration for {LLM} Reasoning",
author = "Wang, Yuchen and
Zhang, Zikang and
Xu, Duo and
Xu, Kunpeng and
Zhu, Qihui and
Huang, Wei and
Wu, Siying and
Cheng, Caihong and
He, Xin and
Yang, Yang and
Jiang, Guanjun and
Sun, Xiaoyan",
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.329/",
pages = "6964--6975",
ISBN = "979-8-89176-462-0",
abstract = "In post-training large language models (LLMs), supervised fine-tuning (SFT) provides expert guidance but can overfit, while reinforcement learning (RL) enables exploration but often suffers from cold-start and sparse-reward issues. Although combining SFT with RL is intuitive, naively mixing off-policy expert data into on-policy RL can induce distribution shift and destabilize optimization. We propose \textbf{APEX}, a framework that harmonizes off-policy guidance and on-policy exploration via a Cognitive State Awareness mechanism. Specifically, APEX leverages a perplexity-based Student{'}s -test to compare the sequence-level perplexity distributions of on-policy generations against candidate offline data. This comparison diagnoses whether failures reflect sampling stochasticity or missing knowledge. Guided by this diagnosis, APEX adaptively gates GRPO optimization and modulates SFT supervision. The framework preferentially selects offline samples that are statistically consistent with the current on-policy distribution to guide RL updates, while filtering out high-shift data and verifiably incorrect samples that could destabilize training. APEX also uses a dynamically weighted SFT loss when needed. Extensive experiments show that APEX improves over SFT-then-RL by +7.10 points in average pass@1 on six math benchmarks and by +15.50 points on three general reasoning benchmarks."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-apex">
<titleInfo>
<title>APEX : Adaptively Balancing Off-Policy Guidance and On-Policy Exploration for LLM Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yuchen</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zikang</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Duo</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kunpeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qihui</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Siying</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Caihong</namePart>
<namePart type="family">Cheng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xin</namePart>
<namePart type="family">He</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yang</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Guanjun</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaoyan</namePart>
<namePart type="family">Sun</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>In post-training large language models (LLMs), supervised fine-tuning (SFT) provides expert guidance but can overfit, while reinforcement learning (RL) enables exploration but often suffers from cold-start and sparse-reward issues. Although combining SFT with RL is intuitive, naively mixing off-policy expert data into on-policy RL can induce distribution shift and destabilize optimization. We propose APEX, a framework that harmonizes off-policy guidance and on-policy exploration via a Cognitive State Awareness mechanism. Specifically, APEX leverages a perplexity-based Student’s -test to compare the sequence-level perplexity distributions of on-policy generations against candidate offline data. This comparison diagnoses whether failures reflect sampling stochasticity or missing knowledge. Guided by this diagnosis, APEX adaptively gates GRPO optimization and modulates SFT supervision. The framework preferentially selects offline samples that are statistically consistent with the current on-policy distribution to guide RL updates, while filtering out high-shift data and verifiably incorrect samples that could destabilize training. APEX also uses a dynamically weighted SFT loss when needed. Extensive experiments show that APEX improves over SFT-then-RL by +7.10 points in average pass@1 on six math benchmarks and by +15.50 points on three general reasoning benchmarks.</abstract>
<identifier type="citekey">wang-etal-2026-apex</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.329/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>6964</start>
<end>6975</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T APEX : Adaptively Balancing Off-Policy Guidance and On-Policy Exploration for LLM Reasoning
%A Wang, Yuchen
%A Zhang, Zikang
%A Xu, Duo
%A Xu, Kunpeng
%A Zhu, Qihui
%A Huang, Wei
%A Wu, Siying
%A Cheng, Caihong
%A He, Xin
%A Yang, Yang
%A Jiang, Guanjun
%A Sun, Xiaoyan
%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 wang-etal-2026-apex
%X In post-training large language models (LLMs), supervised fine-tuning (SFT) provides expert guidance but can overfit, while reinforcement learning (RL) enables exploration but often suffers from cold-start and sparse-reward issues. Although combining SFT with RL is intuitive, naively mixing off-policy expert data into on-policy RL can induce distribution shift and destabilize optimization. We propose APEX, a framework that harmonizes off-policy guidance and on-policy exploration via a Cognitive State Awareness mechanism. Specifically, APEX leverages a perplexity-based Student’s -test to compare the sequence-level perplexity distributions of on-policy generations against candidate offline data. This comparison diagnoses whether failures reflect sampling stochasticity or missing knowledge. Guided by this diagnosis, APEX adaptively gates GRPO optimization and modulates SFT supervision. The framework preferentially selects offline samples that are statistically consistent with the current on-policy distribution to guide RL updates, while filtering out high-shift data and verifiably incorrect samples that could destabilize training. APEX also uses a dynamically weighted SFT loss when needed. Extensive experiments show that APEX improves over SFT-then-RL by +7.10 points in average pass@1 on six math benchmarks and by +15.50 points on three general reasoning benchmarks.
%U https://aclanthology.org/2026.findings-emnlp.329/
%P 6964-6975
Markdown (Informal)
[APEX : Adaptively Balancing Off-Policy Guidance and On-Policy Exploration for LLM Reasoning](https://aclanthology.org/2026.findings-emnlp.329/) (Wang et al., Findings 2026)
ACL
- Yuchen Wang, Zikang Zhang, Duo Xu, Kunpeng Xu, Qihui Zhu, Wei Huang, Siying Wu, Caihong Cheng, Xin He, Yang Yang, Guanjun Jiang, and Xiaoyan Sun. 2026. APEX : Adaptively Balancing Off-Policy Guidance and On-Policy Exploration for LLM Reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 6964–6975, Budapest, Hungary. Association for Computational Linguistics.