OmniThoughtVis: A Scalable Distillation Pipeline for Deployable Multimodal Reasoning Models

• URL: http://arxiv.org/abs/2605.11629v1
• Date: Tue, 12 May 2026 06:54:57 GMT
• Title: OmniThoughtVis: A Scalable Distillation Pipeline for Deployable Multimodal Reasoning Models
• Abstract summary: We present OmniVis, a scalable data curation and distillation pipeline for transferring multimodal reasoning capabilities to smaller, deployment-oriented MLLMs.<n>To maintain data quality at scale, we combine rule-based filtering, difficulty-aware selection, and tag-based diversity sampling.<n>We show consistent gains across model scales, including improvements of up to +16.8 points on MathVerse and +5.6 points on MMMU-Pro for the 4B model.
• Score: 12.869876096473027
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Recent multimodal large language models (MLLMs) have shown strong chain-of-thought (CoT) reasoning ability on vision-language tasks, but their direct deployment in real-world systems is often limited by latency and resource constraints. In practice, smaller MLLMs are preferred for online serving, yet their reasoning performance is bottlenecked by the lack of large-scale, high-quality multimodal CoT supervision. In this paper, we present OmniThoughtVis, a scalable data curation and distillation pipeline for transferring multimodal reasoning capabilities from high-capacity teacher models to smaller, deployment-oriented MLLMs. Starting from a diverse open-source seed pool, our pipeline generates structured CoT traces and performs joint annotation of reasoning difficulty, answer quality, and semantic task tags. To maintain data quality at scale, we combine rule-based filtering, difficulty-aware selection, and tag-based diversity sampling, resulting in a curated corpus of 1.8M samples that supports controllable subset construction for downstream training. We use OmniThoughtVis to distill Qwen3-VL models from 2B to 8B parameters and evaluate them on nine multimodal reasoning benchmarks. The resulting distilled models show consistent gains across model scales, including improvements of up to +16.8 points on MathVerse and +5.6 points on MMMU-Pro for the 4B model. Notably, the distilled 4B model matches or surpasses the undistilled 8B baseline on several tasks, highlighting the practical value of scalable reasoning distillation for deployment-oriented MLLMs.

Related papers

• One LLM to Train Them All: Multi-Task Learning Framework for Fact-Checking [7.856998585396422]
  Large language models (LLMs) are reshaping automated fact-checking (AFC) by enabling unified, end-to-end verification pipelines.<n>We propose textbfmulti-task learning (MTL) as a more efficient alternative that fine-tunes a single model to perform claim detection, evidence ranking, and stance detection jointly.
  arXiv  Detail & Related papers  (2026-01-16T13:44:25Z)
• Think Then Embed: Generative Context Improves Multimodal Embedding [51.76690812535934]
  We propose a Think-Then-Embed (TTE) framework for Universal Multimodal Embeddings (UME), composed of a reasoner and an embedder.<n>By leveraging a powerful MLLM reasoner, we achieve state-of-the-art performance on the MMEB-V2 benchmark, surpassing proprietary models trained on massive in-house datasets.
  arXiv  Detail & Related papers  (2025-10-06T16:53:56Z)
• OptMerge: Unifying Multimodal LLM Capabilities and Modalities via Model Merging [124.91183814854126]
  Model merging seeks to combine multiple expert models into a single model.<n>We introduce a benchmark for model merging research that clearly divides the tasks for MLLM training and evaluation.<n>We find that model merging offers a promising way for building improved MLLMs without requiring training data.
  arXiv  Detail & Related papers  (2025-05-26T12:23:14Z)
• MergeBench: A Benchmark for Merging Domain-Specialized LLMs [25.333088749417414]
  MergeBench is an evaluation suite designed to assess model merging at scale.<n>It builds on state-of-the-art open-source language models, including Llama and Gemma families at 2B to 9B scales.<n>We assess eight representative merging methods across multi-task performance, forgetting and runtime efficiency.
  arXiv  Detail & Related papers  (2025-05-16T04:02:55Z)
• TAMP: Token-Adaptive Layerwise Pruning in Multimodal Large Language Models [23.916205754112774]
  Multimodal Large Language Models (MLLMs) have shown remarkable versatility in understanding diverse multimodal data and tasks.<n>We propose TAMP, a simple yet effective pruning framework tailored for MLLMs.<n>We validate our method on two state-of-the-art MLLMs: LLaVA-NeXT, designed for vision-language tasks, and VideoLLaMA2, capable of processing audio, visual, and language modalities.
  arXiv  Detail & Related papers  (2025-04-14T05:44:38Z)
• SoTA with Less: MCTS-Guided Sample Selection for Data-Efficient Visual Reasoning Self-Improvement [100.85923086072204]
  We introduce ThinkLite-VL, a family of visual reasoning models that achieve state-of-the-art (SoTA) performance using an order of magnitude fewer training samples.<n>We use Monte Carlo Tree Search (MCTS) to measure sample difficulty via the number of reasoning iterations a vision-language model (VLM) requires to solve each instance.<n>ThinkLite-VL-7B and ThinkLite-VL-72B significantly outperform their respective base models across eight visual reasoning benchmarks.
  arXiv  Detail & Related papers  (2025-04-10T17:49:05Z)
• VisualPRM: An Effective Process Reward Model for Multimodal Reasoning [76.35753243272521]
  We introduce VisualPRM, which improves the reasoning abilities of existing Multimodal Large Language Models (MLLMs)<n>Our model achieves a 5.9-point improvement across seven multimodal reasoning benchmarks.<n>For the evaluation of multimodal PRMs, we propose VisualProcessBench, a benchmark with human-annotated step-wise correctness labels.
  arXiv  Detail & Related papers  (2025-03-13T12:03:37Z)
• Enhancing the Reasoning Ability of Multimodal Large Language Models via Mixed Preference Optimization [65.64108848398696]
  We introduce a preference optimization (PO) process to enhance the multimodal reasoning capabilities of MLLMs.<n>Specifically, we design an automated preference data construction pipeline to create MMPR, a high-quality, large-scale multimodal reasoning preference dataset.<n>We explore integrating PO with MLLMs, developing a simple yet effective method, termed Mixed Preference Optimization (MPO), which boosts multimodal CoT performance.
  arXiv  Detail & Related papers  (2024-11-15T18:59:27Z)
• LLaVA-CoT: Let Vision Language Models Reason Step-by-Step [34.32147663809707]
  We introduce LLaVA-CoT, a large Vision-Language Model (VLM) designed to conduct autonomous multistage reasoning.<n>Unlike chain-of-thought prompting, LLaVA-CoT independently engages in sequential stages of summarization, visual interpretation, logical reasoning, and conclusion generation.<n>With only 100k training samples and test-time scaling, LLaVA-CoT not only outperforms its base model by 9.4% on a wide range of multimodal reasoning benchmarks.
  arXiv  Detail & Related papers  (2024-11-15T18:58:31Z)
• LLAVADI: What Matters For Multimodal Large Language Models Distillation [77.73964744238519]
  In this work, we do not propose a new efficient model structure or train small-scale MLLMs from scratch. Our studies involve training strategies, model choices, and distillation algorithms in the knowledge distillation process. By evaluating different benchmarks and proper strategy, even a 2.7B small-scale model can perform on par with larger models with 7B or 13B parameters.
  arXiv  Detail & Related papers  (2024-07-28T06:10:47Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.