@inproceedings{wu-etal-2026-unified,
title = "A Unified Per-Token Gating Family for On-Policy Distillation: {FKL}/{RKL} Mixing with Multi-Channel and Bias Coefficients",
author = "Wu, Suwan and
Lin, Yumeng and
Yuan, Pengcheng and
Jiang, Xiaolong",
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.1082/",
pages = "22921--22933",
ISBN = "979-8-89176-462-0",
abstract = "Per-token gating of forward/reverse KL losses has emerged as a key technique for on-policy distillation (OPD), but existing methods such as EOPD (Agarwal et al., 2024) and ToDi (Jung et al., 2025) each commit to a single gating signal and a single direction. It is unclear which signal works best for a given task, or whether gating helps consistently. We introduce a unified analysis framework: a four-coefficient parameterization $\sigma(a, h_t + b, u_x + c + d, \text{gap}_t)$ in which EOPD and ToDi appear as 1D restrictions, exposing multi-channel composition and an explicit bias term as additional structural degrees of freedom. On TweetEval emotion and hate, our framework improves over the matched-magnitude 1D restrictions in 33 of 36 comparable configurations (91.7{\%}). A 26-configuration mean-match isolation experiment shows that dynamic OPD outperforms effective-KL-matched static baselines in 19 of 26 cases, with no configuration falling below the $-1\sigma$ noise band relative to its matched static counterpart. The optimal $(a,b,c,d)$ is task-conditional: emotion favors entropy, hate favors prompt-level uncertainty."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wu-etal-2026-unified">
<titleInfo>
<title>A Unified Per-Token Gating Family for On-Policy Distillation: FKL/RKL Mixing with Multi-Channel and Bias Coefficients</title>
</titleInfo>
<name type="personal">
<namePart type="given">Suwan</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yumeng</namePart>
<namePart type="family">Lin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pengcheng</namePart>
<namePart type="family">Yuan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaolong</namePart>
<namePart type="family">Jiang</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>Per-token gating of forward/reverse KL losses has emerged as a key technique for on-policy distillation (OPD), but existing methods such as EOPD (Agarwal et al., 2024) and ToDi (Jung et al., 2025) each commit to a single gating signal and a single direction. It is unclear which signal works best for a given task, or whether gating helps consistently. We introduce a unified analysis framework: a four-coefficient parameterization σ(a, h_t + b, u_x + c + d, \textgap_t) in which EOPD and ToDi appear as 1D restrictions, exposing multi-channel composition and an explicit bias term as additional structural degrees of freedom. On TweetEval emotion and hate, our framework improves over the matched-magnitude 1D restrictions in 33 of 36 comparable configurations (91.7%). A 26-configuration mean-match isolation experiment shows that dynamic OPD outperforms effective-KL-matched static baselines in 19 of 26 cases, with no configuration falling below the -1σ noise band relative to its matched static counterpart. The optimal (a,b,c,d) is task-conditional: emotion favors entropy, hate favors prompt-level uncertainty.</abstract>
<identifier type="citekey">wu-etal-2026-unified</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1082/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>22921</start>
<end>22933</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T A Unified Per-Token Gating Family for On-Policy Distillation: FKL/RKL Mixing with Multi-Channel and Bias Coefficients
%A Wu, Suwan
%A Lin, Yumeng
%A Yuan, Pengcheng
%A Jiang, Xiaolong
%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 wu-etal-2026-unified
%X Per-token gating of forward/reverse KL losses has emerged as a key technique for on-policy distillation (OPD), but existing methods such as EOPD (Agarwal et al., 2024) and ToDi (Jung et al., 2025) each commit to a single gating signal and a single direction. It is unclear which signal works best for a given task, or whether gating helps consistently. We introduce a unified analysis framework: a four-coefficient parameterization σ(a, h_t + b, u_x + c + d, \textgap_t) in which EOPD and ToDi appear as 1D restrictions, exposing multi-channel composition and an explicit bias term as additional structural degrees of freedom. On TweetEval emotion and hate, our framework improves over the matched-magnitude 1D restrictions in 33 of 36 comparable configurations (91.7%). A 26-configuration mean-match isolation experiment shows that dynamic OPD outperforms effective-KL-matched static baselines in 19 of 26 cases, with no configuration falling below the -1σ noise band relative to its matched static counterpart. The optimal (a,b,c,d) is task-conditional: emotion favors entropy, hate favors prompt-level uncertainty.
%U https://aclanthology.org/2026.findings-emnlp.1082/
%P 22921-22933
Markdown (Informal)
[A Unified Per-Token Gating Family for On-Policy Distillation: FKL/RKL Mixing with Multi-Channel and Bias Coefficients](https://aclanthology.org/2026.findings-emnlp.1082/) (Wu et al., Findings 2026)
ACL