Yuxin Chen (陈禹昕) | Senior Researcher @ Tencent ARCLab

1

MMhops-R1: Multimodal Multi-hop Reasoning

T Zhang, Z Zhang, Z Ma, Y Chen, B Li, C Yuan, G Wang, F Rao, Y Shan et al.

Introduces MMhops, a benchmark for multimodal multi-hop reasoning requiring iterative integration of information across modalities. Proposes MMhops-R1, a dynamic mRAG framework that adaptively plans reasoning hops, outperforming fixed-hop baselines on complex cross-modal QA.

2

How Far are VLMs from Visual Spatial Intelligence? A Benchmark-Driven Perspective

S Yu, Y Chen, H Ju, L Jia, F Zhang, S Huang, Y Wu, R Cui, B Ran et al.

Systematically evaluates state-of-the-art VLMs on a comprehensive visual spatial intelligence benchmark. Identifies critical limitations in depth estimation, spatial relationship understanding, and 3D reasoning, providing a roadmap for future improvements.

3

DOGR: Towards Versatile Visual Document Grounding and Referring

Y Zhou, Y Chen, H Lin, Y Wu, S Yang, Z Qi, C Ma, L Zhu

Proposes DOGR-Engine, a data engine for generating high-quality fine-grained document grounding and referring data. Constructs DOGR-Bench covering 7 tasks across charts, tables, and documents. The resulting model achieves versatile grounding and referring across diverse document types.

4

VisionMath: Vision-Form Mathematical Problem-Solving

Z Ma, Y Chen, Z Zhang, Z Qi, C Yuan, S Zhu, C Zhuo, B Li, Y Liu, Z Li et al.

Addresses mathematical problem-solving where problems are presented in visual form (handwritten or typeset). Introduces a dedicated dataset and a model that jointly understands visual layouts and performs multi-step mathematical reasoning.

5

Mamba-3VL: Taming State Space Model for 3D Vision Language Learning

Y Wang, Y Chen, Z Qi, L Liu, J Jiao, X Feng, Y Liang, Y Shan, Z Zhang

Adapts State Space Models (Mamba) for 3D vision-language tasks. Addresses the challenge of long-range 3D spatial reasoning by exploiting SSM's linear complexity, achieving strong performance on 3D visual grounding and question answering benchmarks.

6

mRAG: Multimodal Retrieval-Reflection-Augmented Generation for Knowledge-Based VQA

T Zhang, Z Zhang, Z Ma, Y Chen, Z Qi, C Yuan, B Li, J Pu, Y Zhao, Z Xie et al.

Introduces a two-stage reflection mechanism for multimodal RAG: Retrieval-Reflection determines whether external knowledge is needed, while Relevance-Reflection identifies beneficial evidence passages. Reduces redundant retrieval and improves answer accuracy on knowledge-intensive VQA.

7

How to Make Cross Encoder a Good Teacher for Efficient Image-Text Retrieval?

Y Chen, Z Ma, Z Zhang, Z Qi, C Yuan, B Li, J Pu, Y Shan, X Qi, W Hu

Proposes a knowledge distillation framework that transfers fine-grained cross-modal interaction knowledge from a cross-encoder to a dual-encoder for efficient retrieval. Introduces a curriculum-based distillation strategy that progressively focuses on hard negatives, significantly boosting dual-encoder retrieval performance.

8

EA-VTR: Event-Aware Video-Text Retrieval

Z Ma, Z Zhang, Y Chen, Z Qi, C Yuan, B Li, Y Luo, X Li, X Qi, Y Shan et al.

Proposes an event-aware approach for video-text retrieval that explicitly models temporal event structures within videos. By aligning textual event descriptions with video event segments, the model achieves more fine-grained cross-modal understanding.

9

VilEM: Visual-Language Error Modeling for Image-Text Retrieval

Y Chen, Z Ma, Z Zhang, Z Qi, C Yuan, Y Shan, B Li, W Hu, X Qie, J Wu

Addresses noisy supervision in image-text retrieval by modeling error distributions at both instance and token levels. Introduces a soft-label mechanism that accounts for semantically similar but non-paired samples, leading to more robust vision-language representations.

10

Open-vocabulary One-stage Detection with Hierarchical Visual-Language Knowledge Distillation

Z Ma, G Luo, J Gao, L Li, Y Chen, S Wang, C Zhang, W Hu

Enables one-stage open-vocabulary object detection by distilling hierarchical knowledge from a pretrained vision-language model. The model learns to detect novel categories not seen during training by leveraging semantic class embeddings as detection anchors.

11

Channel-wise Topology Refinement Graph Convolution for Skeleton-based Action Recognition

Y Chen, Z Zhang, C Yuan, B Li, Y Deng, W Hu

Proposes CTR-GCN, which learns channel-wise topology graphs to capture diverse inter-joint dependencies for skeleton action recognition. By refining topology independently per channel, the model captures both shared structural priors and task-specific relational patterns, achieving state-of-the-art on NTU RGB+D.

12

DetailFusion: A Dual-branch Framework with Detail Enhancement for Composed Image Retrieval

Y Yang, Y Zhou, Y Chen, Z Zhang, Z Ma, C Yuan, B Li, L Song, J Gao, P Li et al.

Addresses composed image retrieval by designing a dual-branch architecture that coordinates global semantics and fine-grained detail. Leverages image editing data to enrich detail-level supervision, improving retrieval accuracy when subtle visual modifications are specified in the query text.

13

TensorAR: Refinement is All You Need in Autoregressive Image Generation

C Cheng, L Song, Y Xiao, Y Chen, X Zhang, H Sun, Y Shan

Reformulates autoregressive image generation from next-token to next-tensor prediction using overlapping windows, enabling refinement of previously generated tokens. This overcomes the fundamental unidirectional limitation of AR models and significantly reduces accumulated prediction errors.

14

Taming Rectified Flow for Inversion and Editing

J Wang, J Pu, Z Qi, J Guo, Y Ma, N Huang, Y Chen, X Li, Y Shan

Proposes RF-Solver, a high-accuracy ODE solver for rectified flow models (e.g., FLUX) that minimizes discretization error during both sampling and inversion. Builds RF-Edit on top, a general feature-sharing framework for image and video editing via inversion-free guidance injection.

15

Mono2Stereo: A Benchmark and Empirical Study for Stereo Conversion

S Yu, Y Chen, Z Qi, Z Xie, Y Wang, L Wang, Y Shan, H Lu

Introduces the Mono2Stereo dataset and benchmark for monocular-to-stereo video conversion, a task critical for 3D content creation at scale. Conducts a systematic empirical study of existing methods and proposes a baseline that leverages depth estimation and stereo synthesis jointly.

16

MindOmni: Unleashing Reasoning Generation in Vision Language Models with RGPO

Y Xiao, L Song, Y Chen, Y Luo, Y Chen, Y Gan, W Huang, X Li, X Qi et al.

Presents MindOmni, a unified multimodal LLM that simultaneously handles multimodal understanding and image generation with chain-of-thought reasoning. Introduces RGPO (Reasoning-Guided Policy Optimization), a three-phase training strategy that incorporates a decoder-only diffusion module and RL-based reasoning alignment.

17

Learning to Reason in 4D: Dynamic Spatial Understanding for Vision Language Models

S Zhou, Y Chen, Y Ge, W Huang, J Lin, Y Shan, X Qi

Extends VLM spatial reasoning from static 3D to dynamic 4D (space + time). Proposes a training paradigm that teaches VLMs to track and reason about object motion and spatial change across video frames, enabling more grounded temporal visual understanding.

Yuxin Chen 陈禹昕