@inproceedings{yang-etal-2026-token,
title = "Token-Level On-Policy Contrastive Learning for Black-box Distillation Without Full Rollouts",
author = "Yang, Jian and
Lim, Chern Hong and
Vu, Thuy-Trang",
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.34/",
pages = "705--718",
ISBN = "979-8-89176-439-2",
abstract = "Black-box distillation trains student models using only teacher-generated outputs, making it practical when teacher models are accessible only through APIs. However, existing methods either rely mainly on positive imitation or require costly full-response rollouts to obtain corrective signals. We show that such corrective signals can be obtained locally from teacher trajectories alone. At each position, the teacher token provides a positive anchor, while a high-probability alternative sampled from the student provides an informative negative. Based on this insight, we propose Token-level On-policy contrastive Learning for blackbox Distillation (TOLD), which samples a single negative token from the student{'}s distribution and optimizes a Bradley-Terry objective anchored by negative log-likelihood. This enables the student to reinforce teacher tokens while suppressing its own likely mistakes, without teacher logits, reward models, or full-response rollouts. Experiments on reasoning and instruction-following tasks show that TOLD outperforms or matches strong black-box distillation baselines and achieves up to 1.88 times training speedup over rollout-based on-policy baselines."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yang-etal-2026-token">
<titleInfo>
<title>Token-Level On-Policy Contrastive Learning for Black-box Distillation Without Full Rollouts</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jian</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chern</namePart>
<namePart type="given">Hong</namePart>
<namePart type="family">Lim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Thuy-Trang</namePart>
<namePart type="family">Vu</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>Black-box distillation trains student models using only teacher-generated outputs, making it practical when teacher models are accessible only through APIs. However, existing methods either rely mainly on positive imitation or require costly full-response rollouts to obtain corrective signals. We show that such corrective signals can be obtained locally from teacher trajectories alone. At each position, the teacher token provides a positive anchor, while a high-probability alternative sampled from the student provides an informative negative. Based on this insight, we propose Token-level On-policy contrastive Learning for blackbox Distillation (TOLD), which samples a single negative token from the student’s distribution and optimizes a Bradley-Terry objective anchored by negative log-likelihood. This enables the student to reinforce teacher tokens while suppressing its own likely mistakes, without teacher logits, reward models, or full-response rollouts. Experiments on reasoning and instruction-following tasks show that TOLD outperforms or matches strong black-box distillation baselines and achieves up to 1.88 times training speedup over rollout-based on-policy baselines.</abstract>
<identifier type="citekey">yang-etal-2026-token</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.34/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>705</start>
<end>718</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Token-Level On-Policy Contrastive Learning for Black-box Distillation Without Full Rollouts
%A Yang, Jian
%A Lim, Chern Hong
%A Vu, Thuy-Trang
%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 yang-etal-2026-token
%X Black-box distillation trains student models using only teacher-generated outputs, making it practical when teacher models are accessible only through APIs. However, existing methods either rely mainly on positive imitation or require costly full-response rollouts to obtain corrective signals. We show that such corrective signals can be obtained locally from teacher trajectories alone. At each position, the teacher token provides a positive anchor, while a high-probability alternative sampled from the student provides an informative negative. Based on this insight, we propose Token-level On-policy contrastive Learning for blackbox Distillation (TOLD), which samples a single negative token from the student’s distribution and optimizes a Bradley-Terry objective anchored by negative log-likelihood. This enables the student to reinforce teacher tokens while suppressing its own likely mistakes, without teacher logits, reward models, or full-response rollouts. Experiments on reasoning and instruction-following tasks show that TOLD outperforms or matches strong black-box distillation baselines and achieves up to 1.88 times training speedup over rollout-based on-policy baselines.
%U https://aclanthology.org/2026.aacl-long.34/
%P 705-718
Markdown (Informal)
[Token-Level On-Policy Contrastive Learning for Black-box Distillation Without Full Rollouts](https://aclanthology.org/2026.aacl-long.34/) (Yang et al., AACL-IJCNLP 2026)
ACL