@inproceedings{fan-etal-2026-ada,
title = "{A}da-{D}e{VL}: Adaptive Decoupling of Visual and Logical Reasoning Pathways for Efficient {MLLM}s",
author = "Fan, Haoyuan and
Zhao, Jie and
Shi, Yufeng and
Zhou, Wengang and
Li, Houqiang",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.417/",
pages = "9540--9562",
ISBN = "979-8-89176-459-0",
abstract = "Scaling test-time computation in Multimodal Large Language Models (MLLMs) enforces heavy visual analysis and long logical reasoning uniformly across queries, leading to severe token waste and overthinking. We propose **Ada-DeVL**, a framework that decouples visual perception depth from logical reasoning intensity, allowing the model to adaptively select one of four visuo-reasoning pathways according to problem complexity. To learn this adaptive behavior, we first apply a Self-Distillation Supervised Fine-Tuning (SFT) stage that generates difficulty-aligned, hallucination-reduced data while avoiding the exposure bias of cross-model distillation. We then use Group Relative Policy Optimization (GRPO) with a composite reward, incorporating diversity and format-adherence constraints to prevent mode collapse and shortcut exploitation. Notably, on the MMMU benchmark, the Qwen2.5-VL-7B based Ada-DeVL achieves **54.0{\%}** accuracy, surpassing Vision-SR1 while saving **39.6{\%}** tokens; the Qwen3-VL-8B based model reaches **64.0{\%}**, outperforming its vanilla counterpart and reducing token consumption by **44.6{\%}**, and achieves competitive accuracy against the strong Qwen3-VL-8B-Thinking baseline with significantly lower token costs (averaging 1560.8 vs. 4889.0 tokens)."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="fan-etal-2026-ada">
<titleInfo>
<title>Ada-DeVL: Adaptive Decoupling of Visual and Logical Reasoning Pathways for Efficient MLLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Haoyuan</namePart>
<namePart type="family">Fan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jie</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yufeng</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wengang</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Houqiang</namePart>
<namePart type="family">Li</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Scaling test-time computation in Multimodal Large Language Models (MLLMs) enforces heavy visual analysis and long logical reasoning uniformly across queries, leading to severe token waste and overthinking. We propose **Ada-DeVL**, a framework that decouples visual perception depth from logical reasoning intensity, allowing the model to adaptively select one of four visuo-reasoning pathways according to problem complexity. To learn this adaptive behavior, we first apply a Self-Distillation Supervised Fine-Tuning (SFT) stage that generates difficulty-aligned, hallucination-reduced data while avoiding the exposure bias of cross-model distillation. We then use Group Relative Policy Optimization (GRPO) with a composite reward, incorporating diversity and format-adherence constraints to prevent mode collapse and shortcut exploitation. Notably, on the MMMU benchmark, the Qwen2.5-VL-7B based Ada-DeVL achieves **54.0%** accuracy, surpassing Vision-SR1 while saving **39.6%** tokens; the Qwen3-VL-8B based model reaches **64.0%**, outperforming its vanilla counterpart and reducing token consumption by **44.6%**, and achieves competitive accuracy against the strong Qwen3-VL-8B-Thinking baseline with significantly lower token costs (averaging 1560.8 vs. 4889.0 tokens).</abstract>
<identifier type="citekey">fan-etal-2026-ada</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.417/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>9540</start>
<end>9562</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Ada-DeVL: Adaptive Decoupling of Visual and Logical Reasoning Pathways for Efficient MLLMs
%A Fan, Haoyuan
%A Zhao, Jie
%A Shi, Yufeng
%A Zhou, Wengang
%A Li, Houqiang
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F fan-etal-2026-ada
%X Scaling test-time computation in Multimodal Large Language Models (MLLMs) enforces heavy visual analysis and long logical reasoning uniformly across queries, leading to severe token waste and overthinking. We propose **Ada-DeVL**, a framework that decouples visual perception depth from logical reasoning intensity, allowing the model to adaptively select one of four visuo-reasoning pathways according to problem complexity. To learn this adaptive behavior, we first apply a Self-Distillation Supervised Fine-Tuning (SFT) stage that generates difficulty-aligned, hallucination-reduced data while avoiding the exposure bias of cross-model distillation. We then use Group Relative Policy Optimization (GRPO) with a composite reward, incorporating diversity and format-adherence constraints to prevent mode collapse and shortcut exploitation. Notably, on the MMMU benchmark, the Qwen2.5-VL-7B based Ada-DeVL achieves **54.0%** accuracy, surpassing Vision-SR1 while saving **39.6%** tokens; the Qwen3-VL-8B based model reaches **64.0%**, outperforming its vanilla counterpart and reducing token consumption by **44.6%**, and achieves competitive accuracy against the strong Qwen3-VL-8B-Thinking baseline with significantly lower token costs (averaging 1560.8 vs. 4889.0 tokens).
%U https://aclanthology.org/2026.emnlp-main.417/
%P 9540-9562
Markdown (Informal)
[Ada-DeVL: Adaptive Decoupling of Visual and Logical Reasoning Pathways for Efficient MLLMs](https://aclanthology.org/2026.emnlp-main.417/) (Fan et al., EMNLP 2026)
ACL