🎙️ News

• Oct 10, 2025. We release our paper in arxiv.
• Feb 23, 2026. Our paper has been accepted by CVPR Findings 2026!

Dataset

We release three training datasets covering mathematical reasoning, OCR, and multimodal structured data understanding.

Dataset	Link	Description
Math		Mathematical reasoning dataset used for training the math expert model.
OCR		OCR-oriented dataset for text-rich visual understanding and reasoning.
StructureQA		Structured data QA dataset covering charts, tables, documents, posters, and scientific images.

StructureQA Overview

The StructureQA dataset contains chart, poster, document, table, and scientific image QA data collected and processed from multiple public sources.

Category	File	Source	Description
Chart / Poster	chart_galaxy_full_4k.parquet	ChartGalaxy/ChartGalaxy	QA data selected from chart, icon, and poster samples.
Scientific Plot	plotqa_sample_8k.parquet	NiteshMethani/PlotQA	A sampled subset of PlotQA. The original dataset contains 100k samples across four categories; this version keeps 8k samples by randomly selecting 2k from each category. Many samples are generated from Python plotting tools such as Matplotlib.
Document	tatdqa_full_11k.parquet	NExTplusplus/TAT-DQA	Reformatted document QA data from TAT-DQA. The documents include tables and support more complex document-table reasoning.
Table	the_cauldron_robut_wtq_38k.parquet	HuggingFaceM4/the_cauldron	Cleaned and converted from the robut_wtq subset.
Scientific Image	the_cauldron_tqa_7k.parquet	HuggingFaceM4/the_cauldron	Cleaned and converted from the tqa subset.
Table	wikitablequestions_9k.parquet	WTQ official dataset	Converted from the default training split of WTQ. The original data only contains HTML and CSV tables without images, so we render them into images with randomized borders and background colors while keeping icon consistency during conversion.

💡 About Tiny-R1V

Although Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities across diverse tasks, they encounter challenges in terms of reasoning efficiency, large model size and overthinking. However, existing lightweight MLLMs lack the capability to balance high efficiency and performance at a small scale. To this end, we propose Tiny-R1V, a novel lightweight 3B model that achieves faster inference and higher accuracy via a two-stage optimization, while unifying multimodal reasoning across multiple tasks with fewer inference tokens. In the first stage, Tiny-R1V introduces Length-Informed Relative Policy Optimization (LIPO), a new reinforcement learning method, to train each reasoning model, including mathematical reasoning, chart reasoning, and OCR capability. The LIPO dynamically adjusts the advantages of responses within groups by prioritizing concise yet high-quality responses to encourage the generation of shorter and more accurate responses. In the second stage, we propose Adaptive Model Merging (AMM), a training-free model merging method that merges multiple specialist models into a unified architecture. Specifically, AMM adaptively adjusts the weights of task vectors via a novel gradient projection regularization loss function, thus mitigating redundant conflicts between them. Extensive evaluations on ten widely-used reasoning benchmarks covering mathematics, structured data (charts, tables, documents), OCR, and general capabilities showcase the superior performance of Tiny-R1V, enabling lightweight models to excel in diverse multimodal reasoning tasks.

(a)Two-stage framework for training lightweight MLLMs for unified reasoning tasks. (b) The average response length of GRPO, DAPO and LIPO(Ours) on the training set during the RL training process. (c) Tiny-R1V achieves the state-of-the-art performance on a broad range of multimodal reasoning tasks compared with other open source models. Tiny-R1V employs Length-Informed Relative Policy Optimization (LIPO) and Adaptive Model Merging (AMM). In the first stage, Tiny-R1V trains three expert models separately using LIPO, which dynamically adjusts the advantages between groups to minimize the number of response tokens while ensuring the accuracy of the answer. In the second stage, Tiny-R1V merges the three models using AMM, determines the dynamic weights of model parameters, and reduces parameter conflicts, resulting in the final Tiny-R1V-3B model.

🙏 Acknowledgment

Our work is primarily based on the following codebases. We are sincerely grateful for their work.

• EasyR1: We use EasyR1 to fine-tune R1-ShareVL Models.
• VLMEvalKit: We use VLMEvalKit for evaluation.