@inproceedings{zhu-etal-2026-learnability,
title = "On the Learnability of Antidistillation Reasoning Traces",
author = "Zhu, Hongyu and
Liang, Sichu and
Wang, Wenwen and
Chen, Bofan and
Wang, Hanyi and
Wang, Shi-Lin",
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.867/",
doi = "10.18653/2026.findings-emnlp.867",
pages = "18312--18335",
ISBN = "979-8-89176-462-0",
abstract = "Unauthorized distillation turns reasoning traces into a conduit for transferring proprietary models' reasoning behavior to smaller students. Antidistillation Sampling (ADS) perturbs decoding to preserve task utility for legitimate users while obstructing student imitation. Our token-level Metropolis{--}Hastings diagnosis reveals that current ADS-style defenses leave the trace{'}s intrinsic learnability largely intact: most tokens still follow the original teacher, while a few teacher-uncertain positions induce extreme loss and dominate optimization. ADS hides learnable reasoning behind a sparse high-loss tail instead of eliminating it. We therefore reformulate reasoning distillation as offline reinforcement learning from imperfect demonstrations, with credit assignment that reinforces high-quality reasoning behavior while suppressing disruptive segments. Across two teachers, six students, and six datasets, the resulting learner extracts reasoning from current anti-distillation traces, outperforming vanilla maximum-likelihood training and eight supervised fine-tuning baselines by up to 21{\%}. We then turn the same principle into an adaptive defense that directly targets high-value reasoning behaviors, exploiting the perceptual gap between humans and models to preserve readability while disrupting model-visible Unicode sequences. This strategy sharply degrades both classical distillation and the adaptive learner, and scales to protect GPT-5.5 traces. These results suggest that robust anti-distillation should optimize for low \textit{intrinsic learnability}, not merely high token-level surprise."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhu-etal-2026-learnability">
<titleInfo>
<title>On the Learnability of Antidistillation Reasoning Traces</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hongyu</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sichu</namePart>
<namePart type="family">Liang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenwen</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bofan</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hanyi</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shi-Lin</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>Unauthorized distillation turns reasoning traces into a conduit for transferring proprietary models’ reasoning behavior to smaller students. Antidistillation Sampling (ADS) perturbs decoding to preserve task utility for legitimate users while obstructing student imitation. Our token-level Metropolis–Hastings diagnosis reveals that current ADS-style defenses leave the trace’s intrinsic learnability largely intact: most tokens still follow the original teacher, while a few teacher-uncertain positions induce extreme loss and dominate optimization. ADS hides learnable reasoning behind a sparse high-loss tail instead of eliminating it. We therefore reformulate reasoning distillation as offline reinforcement learning from imperfect demonstrations, with credit assignment that reinforces high-quality reasoning behavior while suppressing disruptive segments. Across two teachers, six students, and six datasets, the resulting learner extracts reasoning from current anti-distillation traces, outperforming vanilla maximum-likelihood training and eight supervised fine-tuning baselines by up to 21%. We then turn the same principle into an adaptive defense that directly targets high-value reasoning behaviors, exploiting the perceptual gap between humans and models to preserve readability while disrupting model-visible Unicode sequences. This strategy sharply degrades both classical distillation and the adaptive learner, and scales to protect GPT-5.5 traces. These results suggest that robust anti-distillation should optimize for low intrinsic learnability, not merely high token-level surprise.</abstract>
<identifier type="citekey">zhu-etal-2026-learnability</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.867</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.867/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>18312</start>
<end>18335</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T On the Learnability of Antidistillation Reasoning Traces
%A Zhu, Hongyu
%A Liang, Sichu
%A Wang, Wenwen
%A Chen, Bofan
%A Wang, Hanyi
%A Wang, Shi-Lin
%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 zhu-etal-2026-learnability
%X Unauthorized distillation turns reasoning traces into a conduit for transferring proprietary models’ reasoning behavior to smaller students. Antidistillation Sampling (ADS) perturbs decoding to preserve task utility for legitimate users while obstructing student imitation. Our token-level Metropolis–Hastings diagnosis reveals that current ADS-style defenses leave the trace’s intrinsic learnability largely intact: most tokens still follow the original teacher, while a few teacher-uncertain positions induce extreme loss and dominate optimization. ADS hides learnable reasoning behind a sparse high-loss tail instead of eliminating it. We therefore reformulate reasoning distillation as offline reinforcement learning from imperfect demonstrations, with credit assignment that reinforces high-quality reasoning behavior while suppressing disruptive segments. Across two teachers, six students, and six datasets, the resulting learner extracts reasoning from current anti-distillation traces, outperforming vanilla maximum-likelihood training and eight supervised fine-tuning baselines by up to 21%. We then turn the same principle into an adaptive defense that directly targets high-value reasoning behaviors, exploiting the perceptual gap between humans and models to preserve readability while disrupting model-visible Unicode sequences. This strategy sharply degrades both classical distillation and the adaptive learner, and scales to protect GPT-5.5 traces. These results suggest that robust anti-distillation should optimize for low intrinsic learnability, not merely high token-level surprise.
%R 10.18653/2026.findings-emnlp.867
%U https://aclanthology.org/2026.findings-emnlp.867/
%U https://doi.org/10.18653/2026.findings-emnlp.867
%P 18312-18335
Markdown (Informal)
[On the Learnability of Antidistillation Reasoning Traces](https://aclanthology.org/2026.findings-emnlp.867/) (Zhu et al., Findings 2026)
ACL
- Hongyu Zhu, Sichu Liang, Wenwen Wang, Bofan Chen, Hanyi Wang, and Shi-Lin Wang. 2026. On the Learnability of Antidistillation Reasoning Traces. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 18312–18335, Budapest, Hungary. Association for Computational Linguistics.