Reliable Reasoning in SVG-LLMs via Multi-Task Multi-Reward Reinforcement Learning

Haomin Wang, Qi Wei, Qianli Ma, Shengyuan Ding, Jinhui Yin, Kai Chen, Hongjie Zhang

PDF | Code | Dataset

CTRL-S is a reinforcement learning framework for reliable SVG generation with explicit chain-of-thought reasoning. It introduces SVG-Sophia, a high-quality dataset covering SVG code refinement, text-to-SVG, and image-to-SVG tasks. By combining structured SVG reasoning, group-level code generation, and multi-reward optimization over visual fidelity, text-image alignment, format validity, and code efficiency, CTRL-S improves both SVG quality and generalization across diverse scenarios.

InternSVG: Towards Unified SVG Tasks with Multimodal Large Language Models

Haomin Wang*, Jinhui Yin*, Qi Wei*, Wenguang Zeng, Lixin Gu, Shenglong Ye, Zhangwei Gao, Yaohui Wang, Yanting Zhang, Yuanqi Li, Yanwen Guo, Wenhai Wang, Kai Chen, Yu Qiao, Hongjie Zhang

PDF | Code | Model | Dataset | Benchmark | Homepage

InternSVG is a unified data–benchmark–model suite for SVG understanding, editing, generation. It introduces SAgoge, a large-scale multi-domain SVG dataset, and SArena, a comprehensive benchmark for evaluating multimodal SVG capabilities. Built on VLMs with SVG-specific tokenization and two stage training, InternSVG-8B model achieves strong performance across diverse SVG tasks.

Point or Line? Using Line-based Representation for Panoptic Symbol Spotting in CAD Drawings

Xingguang Wei*, Haomin Wang*, Shenglong Ye, Ruifeng Luo, Yanting Zhang, Lixin Gu, Jifeng Dai, Yu Qiao, Wenhai Wang, Hongjie Zhang

PDF | Homepage

VecFormer uses a line-based representation of primitives for panoptic symbol spotting in floor plan CAD drawings, achieving a new state-of-the-art in the FloorPlanCAD dataset.

Intern-S1-Pro: Scientific Multimodal Foundation Model at Trillion Scale

Intern-S1-Pro Team, Shanghai AI Laboratory

PDF | Model

Intern-S1-Pro is a trillion-parameter scientific multimodal foundation model that strengthens both general reasoning and domain-specific scientific intelligence. With advanced agent capabilities and support for over 100 specialized tasks across fields such as chemistry, materials, life sciences, and earth sciences, it serves as a specializable generalist for scientific discovery and multimodal understanding.

InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

InternVL Team, Shanghai AI Laboratory

PDF | Code | Models

InternVL 3.5 is a new open-source MLLM family that improves multimodal reasoning, versatility, and inference efficiency. It introduces Cascade RL for stronger coarse-to-fine reasoning, ViR for dynamic visual-token resolution, and DvD for efficient vision-language deployment, achieving major gains in reasoning performance and speed. The model also extends to GUI interaction and embodied agent capabilities, reaching SOTA performance among open-source MLLMs.