@inproceedings{li-etal-2026-expanding,
title = "Expanding the Solvable Frontier: Policy-Adaptive Curriculum Learning for Reasoning {RL}",
author = "Li, Guocong and
Zhang, Jinjian and
Wang, Ping and
Guo, Siyan and
Sun, Yukai and
Huang, Hao and
Zhou, Wei and
Mo, Linjian and
Xu, Hongxia and
Wu, Jian",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.535/",
pages = "12329--12345",
ISBN = "979-8-89176-459-0",
abstract = "Existing curriculum learning methods for reasoning reinforcement learning often treat problem difficulty as a static property of the problem. We argue that difficulty is instead policy-dependent: problems beyond the current policy provide sparse learning signals, while problems already solved by greedy decoding may still yield informative stochastic trajectories during RL. We propose PAD-Curriculum, a policy-adaptive curriculum framework for GRPO that estimates the student{'}s evolving solvable frontier from student-generated trajectories. For currently solved problems, PAD-Curriculum ranks examples by a normalized Scorer-to-Student KL discrepancy computed against a frozen stronger scorer; unsolved problems are deferred and periodically re-evaluated as the policy improves. Experiments on GSM8K, MATH, and NuminaMath-CoT with Llama-3.1 and Qwen3 models show consistent gains over standard GRPO and curriculum/distillation baselines, with particularly strong improvements on challenging mathematical reasoning benchmarks. Analyses further show that adaptive discrepancy estimation is more effective than static scorer-based difficulty."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-etal-2026-expanding">
<titleInfo>
<title>Expanding the Solvable Frontier: Policy-Adaptive Curriculum Learning for Reasoning RL</title>
</titleInfo>
<name type="personal">
<namePart type="given">Guocong</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jinjian</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ping</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Siyan</namePart>
<namePart type="family">Guo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yukai</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hao</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Linjian</namePart>
<namePart type="family">Mo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hongxia</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jian</namePart>
<namePart type="family">Wu</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Existing curriculum learning methods for reasoning reinforcement learning often treat problem difficulty as a static property of the problem. We argue that difficulty is instead policy-dependent: problems beyond the current policy provide sparse learning signals, while problems already solved by greedy decoding may still yield informative stochastic trajectories during RL. We propose PAD-Curriculum, a policy-adaptive curriculum framework for GRPO that estimates the student’s evolving solvable frontier from student-generated trajectories. For currently solved problems, PAD-Curriculum ranks examples by a normalized Scorer-to-Student KL discrepancy computed against a frozen stronger scorer; unsolved problems are deferred and periodically re-evaluated as the policy improves. Experiments on GSM8K, MATH, and NuminaMath-CoT with Llama-3.1 and Qwen3 models show consistent gains over standard GRPO and curriculum/distillation baselines, with particularly strong improvements on challenging mathematical reasoning benchmarks. Analyses further show that adaptive discrepancy estimation is more effective than static scorer-based difficulty.</abstract>
<identifier type="citekey">li-etal-2026-expanding</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.535/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>12329</start>
<end>12345</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Expanding the Solvable Frontier: Policy-Adaptive Curriculum Learning for Reasoning RL
%A Li, Guocong
%A Zhang, Jinjian
%A Wang, Ping
%A Guo, Siyan
%A Sun, Yukai
%A Huang, Hao
%A Zhou, Wei
%A Mo, Linjian
%A Xu, Hongxia
%A Wu, Jian
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F li-etal-2026-expanding
%X Existing curriculum learning methods for reasoning reinforcement learning often treat problem difficulty as a static property of the problem. We argue that difficulty is instead policy-dependent: problems beyond the current policy provide sparse learning signals, while problems already solved by greedy decoding may still yield informative stochastic trajectories during RL. We propose PAD-Curriculum, a policy-adaptive curriculum framework for GRPO that estimates the student’s evolving solvable frontier from student-generated trajectories. For currently solved problems, PAD-Curriculum ranks examples by a normalized Scorer-to-Student KL discrepancy computed against a frozen stronger scorer; unsolved problems are deferred and periodically re-evaluated as the policy improves. Experiments on GSM8K, MATH, and NuminaMath-CoT with Llama-3.1 and Qwen3 models show consistent gains over standard GRPO and curriculum/distillation baselines, with particularly strong improvements on challenging mathematical reasoning benchmarks. Analyses further show that adaptive discrepancy estimation is more effective than static scorer-based difficulty.
%U https://aclanthology.org/2026.emnlp-main.535/
%P 12329-12345
Markdown (Informal)
[Expanding the Solvable Frontier: Policy-Adaptive Curriculum Learning for Reasoning RL](https://aclanthology.org/2026.emnlp-main.535/) (Li et al., EMNLP 2026)
ACL
- Guocong Li, Jinjian Zhang, Ping Wang, Siyan Guo, Yukai Sun, Hao Huang, Wei Zhou, Linjian Mo, Hongxia Xu, and Jian Wu. 2026. Expanding the Solvable Frontier: Policy-Adaptive Curriculum Learning for Reasoning RL. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 12329–12345, Budapest, Hungary. Association for Computational Linguistics.