@inproceedings{yue-etal-2026-omnithoughtvis,
title = "{O}mni{T}hought{V}is: A Scalable Distillation Pipeline for Deployable Multimodal Reasoning Models",
author = "Yue, Yuanhao and
Wang, Chengyu and
Lyu, Yuanjie and
Shen, Lei and
Huang, Jun",
editor = "Atanasova, Pepa and
Haffari, Gholamreza and
Wang, Minghan",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-industry.15/",
pages = "225--239",
ISBN = "979-8-89176-460-6",
abstract = "Production multimodal systems must serve complex vision-language tasks under strict latency, memory, and cost constraints. Frontier multimodal large language models (MLLMs) deliver strong reasoning but are prohibitively expensive for online serving; compact alternatives fit deployment budgets but lack structured reasoning ability. We present OmniThoughtVis, a scalable distillation pipeline that converts open-source multimodal corpora into deployment-ready chain-of-thought (CoT) supervision for compact MLLMs. Starting from a diverse seed pool, the pipeline generates structured reasoning traces from a high-capacity teacher, jointly annotates reasoning difficulty, answer quality, and semantic task tags, and applies difficulty-aware filtering with tag-based diversity sampling to produce a reusable 1.8M-sample candidate pool. Practitioners construct budget-specific training subsets from this pool without re-running distillation. A key practical finding is that reasoning difficulty is a far more effective data selection signal than answer quality, which becomes uninformative when teacher outputs are already strong. Distilled Qwen3-VL models (2B-8B) consistently improve over undistilled baselines across nine benchmarks, with the distilled 4B model matching or exceeding the undistilled 8B on multiple tasks. In production, replacing a large-model serving pipeline with the distilled compact model reduces end-to-end response latency by 79{\%} while maintaining comparable downstream quality."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yue-etal-2026-omnithoughtvis">
<titleInfo>
<title>OmniThoughtVis: A Scalable Distillation Pipeline for Deployable Multimodal Reasoning Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yuanhao</namePart>
<namePart type="family">Yue</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chengyu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuanjie</namePart>
<namePart type="family">Lyu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lei</namePart>
<namePart type="family">Shen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pepa</namePart>
<namePart type="family">Atanasova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gholamreza</namePart>
<namePart type="family">Haffari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Minghan</namePart>
<namePart type="family">Wang</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">979-8-89176-460-6</identifier>
</relatedItem>
<abstract>Production multimodal systems must serve complex vision-language tasks under strict latency, memory, and cost constraints. Frontier multimodal large language models (MLLMs) deliver strong reasoning but are prohibitively expensive for online serving; compact alternatives fit deployment budgets but lack structured reasoning ability. We present OmniThoughtVis, a scalable distillation pipeline that converts open-source multimodal corpora into deployment-ready chain-of-thought (CoT) supervision for compact MLLMs. Starting from a diverse seed pool, the pipeline generates structured reasoning traces from a high-capacity teacher, jointly annotates reasoning difficulty, answer quality, and semantic task tags, and applies difficulty-aware filtering with tag-based diversity sampling to produce a reusable 1.8M-sample candidate pool. Practitioners construct budget-specific training subsets from this pool without re-running distillation. A key practical finding is that reasoning difficulty is a far more effective data selection signal than answer quality, which becomes uninformative when teacher outputs are already strong. Distilled Qwen3-VL models (2B-8B) consistently improve over undistilled baselines across nine benchmarks, with the distilled 4B model matching or exceeding the undistilled 8B on multiple tasks. In production, replacing a large-model serving pipeline with the distilled compact model reduces end-to-end response latency by 79% while maintaining comparable downstream quality.</abstract>
<identifier type="citekey">yue-etal-2026-omnithoughtvis</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-industry.15/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>225</start>
<end>239</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T OmniThoughtVis: A Scalable Distillation Pipeline for Deployable Multimodal Reasoning Models
%A Yue, Yuanhao
%A Wang, Chengyu
%A Lyu, Yuanjie
%A Shen, Lei
%A Huang, Jun
%Y Atanasova, Pepa
%Y Haffari, Gholamreza
%Y Wang, Minghan
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-460-6
%F yue-etal-2026-omnithoughtvis
%X Production multimodal systems must serve complex vision-language tasks under strict latency, memory, and cost constraints. Frontier multimodal large language models (MLLMs) deliver strong reasoning but are prohibitively expensive for online serving; compact alternatives fit deployment budgets but lack structured reasoning ability. We present OmniThoughtVis, a scalable distillation pipeline that converts open-source multimodal corpora into deployment-ready chain-of-thought (CoT) supervision for compact MLLMs. Starting from a diverse seed pool, the pipeline generates structured reasoning traces from a high-capacity teacher, jointly annotates reasoning difficulty, answer quality, and semantic task tags, and applies difficulty-aware filtering with tag-based diversity sampling to produce a reusable 1.8M-sample candidate pool. Practitioners construct budget-specific training subsets from this pool without re-running distillation. A key practical finding is that reasoning difficulty is a far more effective data selection signal than answer quality, which becomes uninformative when teacher outputs are already strong. Distilled Qwen3-VL models (2B-8B) consistently improve over undistilled baselines across nine benchmarks, with the distilled 4B model matching or exceeding the undistilled 8B on multiple tasks. In production, replacing a large-model serving pipeline with the distilled compact model reduces end-to-end response latency by 79% while maintaining comparable downstream quality.
%U https://aclanthology.org/2026.emnlp-industry.15/
%P 225-239
Markdown (Informal)
[OmniThoughtVis: A Scalable Distillation Pipeline for Deployable Multimodal Reasoning Models](https://aclanthology.org/2026.emnlp-industry.15/) (Yue et al., EMNLP 2026)
ACL