@inproceedings{yang-etal-2026-sample,
title = "Sample Learning State Guided Dynamic {SFT}{--}{RL} Integration Post-training",
author = "Yang, Zheng and
Liu, Yilin and
Li, Zhuoran and
Gao, Wenhao and
Li, Bingwei and
Sun, Haozhe and
Yuan, Caixia and
Wan, Fan and
Chen, Wu and
Wang, Xiaojie",
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.2011/",
doi = "10.18653/2026.findings-emnlp.2011",
pages = "42798--42818",
ISBN = "979-8-89176-462-0",
abstract = "Integrating supervised fine-tuning (SFT) and reinforcement learning (RL) has become a common post-training paradigm for improving the reasoning capabilities of large language models. Existing methods typically rely on sequential SFT{--}RL pipelines or static loss weights, overlooking sample-level heterogeneity and evolving learning states. We propose Sample Learning State (SLS), a rollout-based characterization of each training sample using two complementary signals: sample mastery and exploration-trajectory dispersion. These signals estimate whether reward-driven updates are reliable for the core task and whether current rollouts provide informative contrast for RL optimization, respectively. Based on SLS, we design a state-aware coefficient that dynamically balances SFT and RL losses, increasing the RL weight only when the model demonstrates sufficient core-task competence and the sampled trajectories provide useful reward contrast. Experiments on Qwen2.5-Math models show that SLS achieves state-of-the-art results on four in-distribution and two out-of-distribution mathematical reasoning benchmarks. In multi-reward settings, SLS preserves core reasoning performance while mitigating reward hacking induced by auxiliary rewards."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yang-etal-2026-sample">
<titleInfo>
<title>Sample Learning State Guided Dynamic SFT–RL Integration Post-training</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zheng</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yilin</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhuoran</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenhao</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bingwei</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haozhe</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Caixia</namePart>
<namePart type="family">Yuan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fan</namePart>
<namePart type="family">Wan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wu</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaojie</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>Integrating supervised fine-tuning (SFT) and reinforcement learning (RL) has become a common post-training paradigm for improving the reasoning capabilities of large language models. Existing methods typically rely on sequential SFT–RL pipelines or static loss weights, overlooking sample-level heterogeneity and evolving learning states. We propose Sample Learning State (SLS), a rollout-based characterization of each training sample using two complementary signals: sample mastery and exploration-trajectory dispersion. These signals estimate whether reward-driven updates are reliable for the core task and whether current rollouts provide informative contrast for RL optimization, respectively. Based on SLS, we design a state-aware coefficient that dynamically balances SFT and RL losses, increasing the RL weight only when the model demonstrates sufficient core-task competence and the sampled trajectories provide useful reward contrast. Experiments on Qwen2.5-Math models show that SLS achieves state-of-the-art results on four in-distribution and two out-of-distribution mathematical reasoning benchmarks. In multi-reward settings, SLS preserves core reasoning performance while mitigating reward hacking induced by auxiliary rewards.</abstract>
<identifier type="citekey">yang-etal-2026-sample</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.2011</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2011/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>42798</start>
<end>42818</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Sample Learning State Guided Dynamic SFT–RL Integration Post-training
%A Yang, Zheng
%A Liu, Yilin
%A Li, Zhuoran
%A Gao, Wenhao
%A Li, Bingwei
%A Sun, Haozhe
%A Yuan, Caixia
%A Wan, Fan
%A Chen, Wu
%A Wang, Xiaojie
%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 yang-etal-2026-sample
%X Integrating supervised fine-tuning (SFT) and reinforcement learning (RL) has become a common post-training paradigm for improving the reasoning capabilities of large language models. Existing methods typically rely on sequential SFT–RL pipelines or static loss weights, overlooking sample-level heterogeneity and evolving learning states. We propose Sample Learning State (SLS), a rollout-based characterization of each training sample using two complementary signals: sample mastery and exploration-trajectory dispersion. These signals estimate whether reward-driven updates are reliable for the core task and whether current rollouts provide informative contrast for RL optimization, respectively. Based on SLS, we design a state-aware coefficient that dynamically balances SFT and RL losses, increasing the RL weight only when the model demonstrates sufficient core-task competence and the sampled trajectories provide useful reward contrast. Experiments on Qwen2.5-Math models show that SLS achieves state-of-the-art results on four in-distribution and two out-of-distribution mathematical reasoning benchmarks. In multi-reward settings, SLS preserves core reasoning performance while mitigating reward hacking induced by auxiliary rewards.
%R 10.18653/2026.findings-emnlp.2011
%U https://aclanthology.org/2026.findings-emnlp.2011/
%U https://doi.org/10.18653/2026.findings-emnlp.2011
%P 42798-42818
Markdown (Informal)
[Sample Learning State Guided Dynamic SFT–RL Integration Post-training](https://aclanthology.org/2026.findings-emnlp.2011/) (Yang et al., Findings 2026)
ACL
- Zheng Yang, Yilin Liu, Zhuoran Li, Wenhao Gao, Bingwei Li, Haozhe Sun, Caixia Yuan, Fan Wan, Wu Chen, and Xiaojie Wang. 2026. Sample Learning State Guided Dynamic SFT–RL Integration Post-training. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 42798–42818, Budapest, Hungary. Association for Computational Linguistics.