@inproceedings{zhou-etal-2026-capo,
title = "{CAPO}-{SLT}: Confidence-Aware Policy Optimization for Stable Sign Language Translation Generation",
author = "Zhou, Ling and
Chen, Yuhao and
Cheng, Lin and
Yao, Chengwen and
Sun, Donghui and
Chen, Xiaoxin",
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.850/",
doi = "10.18653/2026.findings-emnlp.850",
pages = "17999--18014",
ISBN = "979-8-89176-462-0",
abstract = "Recent sign language translation methods have advanced visual-language alignment, yet autoregressive generation remains vulnerable to unstable policy updates under ambiguous visual evidence. Such instability may cause locally plausible continuations to accumulate into sentence-level semantic deviations, especially when reinforcement learning is used with noisy reward and advantage estimates. Existing PPO-style policy optimization applies confidence-agnostic clipping bounds to all generated tokens, which limits its ability to distinguish between over-amplified confident predictions and low-probability tokens that may still carry useful corrective signals. To address this issue, we propose CAPO-SLT, a reinforcement learning based framework for stable sign language translation generation. CAPO introduces a confidence-aware policy optimization strategy that adapts the upper clipping bound according to old-policy token probability. It applies more conservative updates to high-confidence predictions while preserving update capacity for low-probability positive-advantage tokens. In addition, CAPO caps negative-advantage updates to improve robustness under noisy supervision. This design provides a simple drop-in replacement for standard clipping and stabilizes generation-stage optimization without changing the underlying visual encoder or reward function. Experiments on Chinese and American sign language benchmarks show consistent improvements over pose-based baselines under the same input setting, suggesting that CAPO-SLT provides an effective optimization strategy for stable sign language translation generation."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhou-etal-2026-capo">
<titleInfo>
<title>CAPO-SLT: Confidence-Aware Policy Optimization for Stable Sign Language Translation Generation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ling</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuhao</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lin</namePart>
<namePart type="family">Cheng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chengwen</namePart>
<namePart type="family">Yao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Donghui</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaoxin</namePart>
<namePart type="family">Chen</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>Recent sign language translation methods have advanced visual-language alignment, yet autoregressive generation remains vulnerable to unstable policy updates under ambiguous visual evidence. Such instability may cause locally plausible continuations to accumulate into sentence-level semantic deviations, especially when reinforcement learning is used with noisy reward and advantage estimates. Existing PPO-style policy optimization applies confidence-agnostic clipping bounds to all generated tokens, which limits its ability to distinguish between over-amplified confident predictions and low-probability tokens that may still carry useful corrective signals. To address this issue, we propose CAPO-SLT, a reinforcement learning based framework for stable sign language translation generation. CAPO introduces a confidence-aware policy optimization strategy that adapts the upper clipping bound according to old-policy token probability. It applies more conservative updates to high-confidence predictions while preserving update capacity for low-probability positive-advantage tokens. In addition, CAPO caps negative-advantage updates to improve robustness under noisy supervision. This design provides a simple drop-in replacement for standard clipping and stabilizes generation-stage optimization without changing the underlying visual encoder or reward function. Experiments on Chinese and American sign language benchmarks show consistent improvements over pose-based baselines under the same input setting, suggesting that CAPO-SLT provides an effective optimization strategy for stable sign language translation generation.</abstract>
<identifier type="citekey">zhou-etal-2026-capo</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.850</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.850/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>17999</start>
<end>18014</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T CAPO-SLT: Confidence-Aware Policy Optimization for Stable Sign Language Translation Generation
%A Zhou, Ling
%A Chen, Yuhao
%A Cheng, Lin
%A Yao, Chengwen
%A Sun, Donghui
%A Chen, Xiaoxin
%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 zhou-etal-2026-capo
%X Recent sign language translation methods have advanced visual-language alignment, yet autoregressive generation remains vulnerable to unstable policy updates under ambiguous visual evidence. Such instability may cause locally plausible continuations to accumulate into sentence-level semantic deviations, especially when reinforcement learning is used with noisy reward and advantage estimates. Existing PPO-style policy optimization applies confidence-agnostic clipping bounds to all generated tokens, which limits its ability to distinguish between over-amplified confident predictions and low-probability tokens that may still carry useful corrective signals. To address this issue, we propose CAPO-SLT, a reinforcement learning based framework for stable sign language translation generation. CAPO introduces a confidence-aware policy optimization strategy that adapts the upper clipping bound according to old-policy token probability. It applies more conservative updates to high-confidence predictions while preserving update capacity for low-probability positive-advantage tokens. In addition, CAPO caps negative-advantage updates to improve robustness under noisy supervision. This design provides a simple drop-in replacement for standard clipping and stabilizes generation-stage optimization without changing the underlying visual encoder or reward function. Experiments on Chinese and American sign language benchmarks show consistent improvements over pose-based baselines under the same input setting, suggesting that CAPO-SLT provides an effective optimization strategy for stable sign language translation generation.
%R 10.18653/2026.findings-emnlp.850
%U https://aclanthology.org/2026.findings-emnlp.850/
%U https://doi.org/10.18653/2026.findings-emnlp.850
%P 17999-18014
Markdown (Informal)
[CAPO-SLT: Confidence-Aware Policy Optimization for Stable Sign Language Translation Generation](https://aclanthology.org/2026.findings-emnlp.850/) (Zhou et al., Findings 2026)
ACL