@inproceedings{li-etal-2026-januspo,
title = "{J}anus{PO}: Asymmetric Clip and {KL} for Entropy-Stable Group Policy Optimization in Tool-Augmented {LLM} Reasoning",
author = "Li, Hao and
Zhuo, Man Fung and
Wu, Guangxin and
Wu, Yusen and
Wu, Jiayu",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.200/",
pages = "4433--4450",
ISBN = "979-8-89176-439-2",
abstract = "Group-relative policy optimization (GRPO, DAPO, ARPO) has become the dominant recipe for RL fine-tuning of reasoning LLMs, yet in the resource-constrained regime most practitioners train in (a single GPU, small batches) it consistently suffers from rapid entropy collapse that kills exploration, together with sparse credit assignment in long rollouts, which is most extreme under multi-turn tool use but already present in long single-turn chain-of-thought. We address both failure modes with JANUSPO, a drop-in GRPO modification that introduces asymmetric PPO clipping $\epsilon_{+}<\epsilon_{-}$ to tolerate larger downward than upward updates, pairs it with an asymmetric KL schedule $\beta_{pos}<\beta_{neg}$ that doubles as a free entropy regulator without any explicit upper bound, and redistributes the terminal reward through a turn-weighted decomposition $w_{t}(n)$ that generalises naturally from single-turn CoT to multi-turn rollouts. On Qwen2.5-3B-Instruct / GSM8K, JANUSPO improves from 74.8{\%} to 84.3$\pm$0.2{\%} (mean$\pm$std over 2 seeds) across 14 controlled ablations, and the same recipe lifts 7B from 85.3{\%} to 93.6{\%} with no hyper-parameter retuning. On multi-turn DeepSearch (103 GAIA + 100 HLE), SFT-warmup-then-JANUSPO is the only RL recipe whose paired comparison against SFT survives best@8 outlier verification, achieving strict dominance on all four reward and F1 columns while a matched-init GRPO baseline{'}s apparent F1 wins collapse to 0{\%}. Beyond this main result, we document a systematic negative finding: SFT warm-up is a necessary precondition for GRPO-style multi-turn RL on format-unaware base models."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-etal-2026-januspo">
<titleInfo>
<title>JanusPO: Asymmetric Clip and KL for Entropy-Stable Group Policy Optimization in Tool-Augmented LLM Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hao</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Man</namePart>
<namePart type="given">Fung</namePart>
<namePart type="family">Zhuo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Guangxin</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yusen</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiayu</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-439-2</identifier>
</relatedItem>
<abstract>Group-relative policy optimization (GRPO, DAPO, ARPO) has become the dominant recipe for RL fine-tuning of reasoning LLMs, yet in the resource-constrained regime most practitioners train in (a single GPU, small batches) it consistently suffers from rapid entropy collapse that kills exploration, together with sparse credit assignment in long rollouts, which is most extreme under multi-turn tool use but already present in long single-turn chain-of-thought. We address both failure modes with JANUSPO, a drop-in GRPO modification that introduces asymmetric PPO clipping ε_+<ε_- to tolerate larger downward than upward updates, pairs it with an asymmetric KL schedule β_pos<β_neg that doubles as a free entropy regulator without any explicit upper bound, and redistributes the terminal reward through a turn-weighted decomposition w_t(n) that generalises naturally from single-turn CoT to multi-turn rollouts. On Qwen2.5-3B-Instruct / GSM8K, JANUSPO improves from 74.8% to 84.3\pm0.2% (mean\pmstd over 2 seeds) across 14 controlled ablations, and the same recipe lifts 7B from 85.3% to 93.6% with no hyper-parameter retuning. On multi-turn DeepSearch (103 GAIA + 100 HLE), SFT-warmup-then-JANUSPO is the only RL recipe whose paired comparison against SFT survives best@8 outlier verification, achieving strict dominance on all four reward and F1 columns while a matched-init GRPO baseline’s apparent F1 wins collapse to 0%. Beyond this main result, we document a systematic negative finding: SFT warm-up is a necessary precondition for GRPO-style multi-turn RL on format-unaware base models.</abstract>
<identifier type="citekey">li-etal-2026-januspo</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.200/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>4433</start>
<end>4450</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T JanusPO: Asymmetric Clip and KL for Entropy-Stable Group Policy Optimization in Tool-Augmented LLM Reasoning
%A Li, Hao
%A Zhuo, Man Fung
%A Wu, Guangxin
%A Wu, Yusen
%A Wu, Jiayu
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F li-etal-2026-januspo
%X Group-relative policy optimization (GRPO, DAPO, ARPO) has become the dominant recipe for RL fine-tuning of reasoning LLMs, yet in the resource-constrained regime most practitioners train in (a single GPU, small batches) it consistently suffers from rapid entropy collapse that kills exploration, together with sparse credit assignment in long rollouts, which is most extreme under multi-turn tool use but already present in long single-turn chain-of-thought. We address both failure modes with JANUSPO, a drop-in GRPO modification that introduces asymmetric PPO clipping ε_+<ε_- to tolerate larger downward than upward updates, pairs it with an asymmetric KL schedule β_pos<β_neg that doubles as a free entropy regulator without any explicit upper bound, and redistributes the terminal reward through a turn-weighted decomposition w_t(n) that generalises naturally from single-turn CoT to multi-turn rollouts. On Qwen2.5-3B-Instruct / GSM8K, JANUSPO improves from 74.8% to 84.3\pm0.2% (mean\pmstd over 2 seeds) across 14 controlled ablations, and the same recipe lifts 7B from 85.3% to 93.6% with no hyper-parameter retuning. On multi-turn DeepSearch (103 GAIA + 100 HLE), SFT-warmup-then-JANUSPO is the only RL recipe whose paired comparison against SFT survives best@8 outlier verification, achieving strict dominance on all four reward and F1 columns while a matched-init GRPO baseline’s apparent F1 wins collapse to 0%. Beyond this main result, we document a systematic negative finding: SFT warm-up is a necessary precondition for GRPO-style multi-turn RL on format-unaware base models.
%U https://aclanthology.org/2026.aacl-long.200/
%P 4433-4450
Markdown (Informal)
[JanusPO: Asymmetric Clip and KL for Entropy-Stable Group Policy Optimization in Tool-Augmented LLM Reasoning](https://aclanthology.org/2026.aacl-long.200/) (Li et al., AACL-IJCNLP 2026)
ACL