@inproceedings{sun-etal-2026-binned,
title = "Binned Relative Policy Optimization towards Efficient Thinking of {VLM}s",
author = "Sun, ShuaiHang and
Li, Zongzhao and
Shi, Kai and
Yang, Jun and
Yang, Zhenyu and
Lu, Haonan and
Huang, Wenbing",
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.1965/",
pages = "41903--41918",
ISBN = "979-8-89176-462-0",
abstract = "Recent VLMs improve multimodal reasoning by generating explicit chains of thought, but often overthink with redundant reasoning. Although conventional length penalties mitigate this redundancy, they rely on coarse global rewards that fail to discern intermediate process quality or adapt to task complexity. To address this dilemma, we propose Binned Relative Policy Optimization (BRPO), a training-efficient framework that enables non-trivial process-level evaluation without introducing costly process reward models. Specifically, BRPO leverages high entropy lexical markers to segment reasoning trajectories into discrete intermediate prefixes, evaluating their problem-solving potential through forced answer completions. Building on this, we introduce Bin Normalization along with a length-aware adjustment to map prefixes into process bins, dynamically calibrating rewards and regulating length constraints based on reasoning stages and visual complexity. Experiments on six multimodal math benchmarks show that our method improves both accuracy and efficiency: on Qwen2.5-VL-7B, it raises average accuracy from 47.8{\%} to 53.6{\%} while reducing the average response length by 39.0{\%}."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="sun-etal-2026-binned">
<titleInfo>
<title>Binned Relative Policy Optimization towards Efficient Thinking of VLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">ShuaiHang</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zongzhao</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kai</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhenyu</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haonan</namePart>
<namePart type="family">Lu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenbing</namePart>
<namePart type="family">Huang</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 VLMs improve multimodal reasoning by generating explicit chains of thought, but often overthink with redundant reasoning. Although conventional length penalties mitigate this redundancy, they rely on coarse global rewards that fail to discern intermediate process quality or adapt to task complexity. To address this dilemma, we propose Binned Relative Policy Optimization (BRPO), a training-efficient framework that enables non-trivial process-level evaluation without introducing costly process reward models. Specifically, BRPO leverages high entropy lexical markers to segment reasoning trajectories into discrete intermediate prefixes, evaluating their problem-solving potential through forced answer completions. Building on this, we introduce Bin Normalization along with a length-aware adjustment to map prefixes into process bins, dynamically calibrating rewards and regulating length constraints based on reasoning stages and visual complexity. Experiments on six multimodal math benchmarks show that our method improves both accuracy and efficiency: on Qwen2.5-VL-7B, it raises average accuracy from 47.8% to 53.6% while reducing the average response length by 39.0%.</abstract>
<identifier type="citekey">sun-etal-2026-binned</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1965/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>41903</start>
<end>41918</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Binned Relative Policy Optimization towards Efficient Thinking of VLMs
%A Sun, ShuaiHang
%A Li, Zongzhao
%A Shi, Kai
%A Yang, Jun
%A Yang, Zhenyu
%A Lu, Haonan
%A Huang, Wenbing
%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 sun-etal-2026-binned
%X Recent VLMs improve multimodal reasoning by generating explicit chains of thought, but often overthink with redundant reasoning. Although conventional length penalties mitigate this redundancy, they rely on coarse global rewards that fail to discern intermediate process quality or adapt to task complexity. To address this dilemma, we propose Binned Relative Policy Optimization (BRPO), a training-efficient framework that enables non-trivial process-level evaluation without introducing costly process reward models. Specifically, BRPO leverages high entropy lexical markers to segment reasoning trajectories into discrete intermediate prefixes, evaluating their problem-solving potential through forced answer completions. Building on this, we introduce Bin Normalization along with a length-aware adjustment to map prefixes into process bins, dynamically calibrating rewards and regulating length constraints based on reasoning stages and visual complexity. Experiments on six multimodal math benchmarks show that our method improves both accuracy and efficiency: on Qwen2.5-VL-7B, it raises average accuracy from 47.8% to 53.6% while reducing the average response length by 39.0%.
%U https://aclanthology.org/2026.findings-emnlp.1965/
%P 41903-41918
Markdown (Informal)
[Binned Relative Policy Optimization towards Efficient Thinking of VLMs](https://aclanthology.org/2026.findings-emnlp.1965/) (Sun et al., Findings 2026)
ACL