@inproceedings{liu-etal-2026-token,
title = "When Token Supervision Fails: Segment-Level On-Policy Distillation for Long-Horizon Agents",
author = "Liu, Jialong and
Li, Zuchao and
Shi, Yuling and
Zhang, Lefei and
Qi, Baoyuan and
Zhao, Hai",
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.167/",
pages = "3688--3725",
ISBN = "979-8-89176-439-2",
abstract = "On-policy distillation (OPD) for long-horizon language agents fails even when the teacher remains sequence-informative: in a failing Qwen3 configuration, the teacher still separates good from bad rollouts (sequence AUROC $\approx$ 0.75), yet the local token-gradient alignment $\alpha_{\mathrm{diag}}$ decays from 0.71 to 0.09 and standard OPD recovers only 1.3{\%} of the teacher{--}student gap. We restrict our claims to two coupled mechanisms, jointly termed CHOP (\textit{Counterfactual Hierarchical On-Policy distillation}): (i) a per-state reliability gate routing between per-token reverse Kullback{--}Leibler matching and per-segment sliced-Wasserstein matching, and (ii) a counterfactual total-variation gate activating privileged information (PI) only when the PI changes the teacher{'}s distribution. The scope is explicit: CHOP targets \textit{low-reliability, long-horizon traces where token-local support matching plateaus}; in short-horizon or high-reliability regimes Top-$K$ local-support matching (LSM) is already sufficient and CHOP{'}s segment branch is rarely activated. CHOP recovers 31.7{\%} of the anisotropic-Qwen3 gap (vs. 1.3{\%}/14.5{\%}/24.7{\%} for OPD/Top-$K$ LSM/Li et al. 2026); \textit{on SWE-bench Verified under a fixed OpenHands harness, CHOP-deployable prevents the OPD-induced regression} ($6.5\%\!\to\!13.2\%$, matched to off-the-shelf 11.8{\%}, not a significant gain over off-the-shelf) and \textit{oracle PI reveals further headroom} (14.2{\%}, 95{\%} CI $[11.4,17.3]$); on AIME 2024 at a 60K-token \textit{stress cap} (a diagnostic horizon, not a deployment target), CHOP holds 44.9 vs. 42.8 for Top-$K$ LSM ($p=0.032$), while standard OPD collapses to 28.4. Compute-matched, leakage, and PI-removed-teacher controls attribute the gain to supervision-object routing rather than extra compute or PI contamination."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liu-etal-2026-token">
<titleInfo>
<title>When Token Supervision Fails: Segment-Level On-Policy Distillation for Long-Horizon Agents</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jialong</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zuchao</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuling</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lefei</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Baoyuan</namePart>
<namePart type="family">Qi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hai</namePart>
<namePart type="family">Zhao</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>On-policy distillation (OPD) for long-horizon language agents fails even when the teacher remains sequence-informative: in a failing Qwen3 configuration, the teacher still separates good from bad rollouts (sequence AUROC \approx 0.75), yet the local token-gradient alignment α_diag decays from 0.71 to 0.09 and standard OPD recovers only 1.3% of the teacher–student gap. We restrict our claims to two coupled mechanisms, jointly termed CHOP (Counterfactual Hierarchical On-Policy distillation): (i) a per-state reliability gate routing between per-token reverse Kullback–Leibler matching and per-segment sliced-Wasserstein matching, and (ii) a counterfactual total-variation gate activating privileged information (PI) only when the PI changes the teacher’s distribution. The scope is explicit: CHOP targets low-reliability, long-horizon traces where token-local support matching plateaus; in short-horizon or high-reliability regimes Top-K local-support matching (LSM) is already sufficient and CHOP’s segment branch is rarely activated. CHOP recovers 31.7% of the anisotropic-Qwen3 gap (vs. 1.3%/14.5%/24.7% for OPD/Top-K LSM/Li et al. 2026); on SWE-bench Verified under a fixed OpenHands harness, CHOP-deployable prevents the OPD-induced regression (6.5%\!\!13.2%, matched to off-the-shelf 11.8%, not a significant gain over off-the-shelf) and oracle PI reveals further headroom (14.2%, 95% CI [11.4,17.3]); on AIME 2024 at a 60K-token stress cap (a diagnostic horizon, not a deployment target), CHOP holds 44.9 vs. 42.8 for Top-K LSM (p=0.032), while standard OPD collapses to 28.4. Compute-matched, leakage, and PI-removed-teacher controls attribute the gain to supervision-object routing rather than extra compute or PI contamination.</abstract>
<identifier type="citekey">liu-etal-2026-token</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.167/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>3688</start>
<end>3725</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T When Token Supervision Fails: Segment-Level On-Policy Distillation for Long-Horizon Agents
%A Liu, Jialong
%A Li, Zuchao
%A Shi, Yuling
%A Zhang, Lefei
%A Qi, Baoyuan
%A Zhao, Hai
%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 liu-etal-2026-token
%X On-policy distillation (OPD) for long-horizon language agents fails even when the teacher remains sequence-informative: in a failing Qwen3 configuration, the teacher still separates good from bad rollouts (sequence AUROC \approx 0.75), yet the local token-gradient alignment α_diag decays from 0.71 to 0.09 and standard OPD recovers only 1.3% of the teacher–student gap. We restrict our claims to two coupled mechanisms, jointly termed CHOP (Counterfactual Hierarchical On-Policy distillation): (i) a per-state reliability gate routing between per-token reverse Kullback–Leibler matching and per-segment sliced-Wasserstein matching, and (ii) a counterfactual total-variation gate activating privileged information (PI) only when the PI changes the teacher’s distribution. The scope is explicit: CHOP targets low-reliability, long-horizon traces where token-local support matching plateaus; in short-horizon or high-reliability regimes Top-K local-support matching (LSM) is already sufficient and CHOP’s segment branch is rarely activated. CHOP recovers 31.7% of the anisotropic-Qwen3 gap (vs. 1.3%/14.5%/24.7% for OPD/Top-K LSM/Li et al. 2026); on SWE-bench Verified under a fixed OpenHands harness, CHOP-deployable prevents the OPD-induced regression (6.5%\!\!13.2%, matched to off-the-shelf 11.8%, not a significant gain over off-the-shelf) and oracle PI reveals further headroom (14.2%, 95% CI [11.4,17.3]); on AIME 2024 at a 60K-token stress cap (a diagnostic horizon, not a deployment target), CHOP holds 44.9 vs. 42.8 for Top-K LSM (p=0.032), while standard OPD collapses to 28.4. Compute-matched, leakage, and PI-removed-teacher controls attribute the gain to supervision-object routing rather than extra compute or PI contamination.
%U https://aclanthology.org/2026.aacl-long.167/
%P 3688-3725
Markdown (Informal)
[When Token Supervision Fails: Segment-Level On-Policy Distillation for Long-Horizon Agents](https://aclanthology.org/2026.aacl-long.167/) (Liu et al., AACL-IJCNLP 2026)
ACL
- Jialong Liu, Zuchao Li, Yuling Shi, Lefei Zhang, Baoyuan Qi, and Hai Zhao. 2026. When Token Supervision Fails: Segment-Level On-Policy Distillation for Long-Horizon Agents. In 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), pages 3688–3725, Hengqin, China. Association for Computational Linguistics.