# Yuhang Zang

> Young researcher at Shanghai AI Laboratory focusing on post-training for multimodal LLMs and vision-language pre-training. Area Chair for NeurIPS, ICLR, CVPR, AAAI, and COLM.

Yuhang Zang (臧宇航, Zang Yuhang) is a Researcher at Shanghai AI Laboratory (https://www.shlab.org.cn/). Research areas: Multimodal Large Language Models; Vision-Language Models; Reinforcement Learning from Human Feedback; Machine Learning; Computer Vision; Artificial Intelligence. This site is the canonical, author-maintained record of Yuhang Zang's publications: every paper page carries the author abstract, a one-sentence takeaway, verified results with table or section locators in the source paper, and a BibTeX entry checked against the published record.

Publications listed: 66. Topics: Multimodal Large Language Models; AIGC; Image Understanding; Reinforcement Learning from Human Feedback; Vision-Language Models.

## Author identifiers

- ORCID: https://orcid.org/0000-0003-1110-5062
- Google Scholar: https://scholar.google.com/citations?user=hW23VKIAAAAJ
- DBLP: https://dblp.org/pid/230/4433
- OpenAlex: https://openalex.org/A5005200501
- Semantic Scholar: https://www.semanticscholar.org/author/12862495
- GitHub: https://github.com/yuhangzang
- Hugging Face: https://huggingface.co/yuhangzang
- X (Twitter): https://twitter.com/yuhangzang
- LinkedIn: https://www.linkedin.com/in/yuhang-zang/
- Homepage: https://yuhangzang.github.io/

## Machine-readable data

- [All publications (JSON)](https://yuhangzang.github.io/data/publications.json): every record on this site with abstracts, takeaways, verified results, source locators and citation metadata
- [All publications (BibTeX)](https://yuhangzang.github.io/publications.bib): one file with every citation entry
- [Full text of every paper page](https://yuhangzang.github.io/llms-full.txt): the complete content of all paper pages in Markdown
- [Sitemap](https://yuhangzang.github.io/sitemap.xml): all pages with last-modified dates
- Each paper page also has a BibTeX file (`.bib`) at the same address as its `.html` page; its full record is the matching entry in the JSON file above.

## Selected publications

- [EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models](https://yuhangzang.github.io/papers/arxiv-2603.12252.html): European Conference on Computer Vision (ECCV), 2026. EndoCoT couples iterative latent reasoning with diffusion generation, achieving 92.1% average accuracy under task-specific training across the visual reasoning settings in Table 1, compared with 83.8% for DiffThinker. arXiv:2603.12252. BibTeX: https://yuhangzang.github.io/papers/arxiv-2603.12252.bib
- [Unified Multimodal Chain-of-Thought Reward Model through Reinforcement Fine-Tuning](https://yuhangzang.github.io/papers/arxiv-2505.03318.html): Neural Information Processing Systems (NeurIPS), 2025. UnifiedReward-Think strengthens multimodal reward judgments through explicit multi-step reasoning and reinforcement fine-tuning across visual understanding and generation tasks. arXiv:2505.03318, DOI:10.52202/085713-5315. BibTeX: https://yuhangzang.github.io/papers/arxiv-2505.03318.bib
- [Visual-RFT: Visual Reinforcement Fine-Tuning](https://yuhangzang.github.io/papers/arxiv-2503.01785.html): IEEE International Conference on Computer Vision (ICCV), 2025. Visual-RFT shows that GRPO with task-specific verifiable rewards improves classification, detection, and grounding in large vision-language models trained with limited labeled data. arXiv:2503.01785. BibTeX: https://yuhangzang.github.io/papers/arxiv-2503.01785.bib
- [InternLM-XComposer2.5-Reward: A Simple Yet Effective Multi-Modal Reward Model](https://yuhangzang.github.io/papers/arxiv-2501.12368.html): Findings of the Association for Computational Linguistics (ACL), 2025. IXC-2.5-Reward provides a multimodal preference signal that supports reinforcement learning, best-response selection and instruction-data filtering across text, image and video inputs. arXiv:2501.12368, DOI:10.18653/v1/2025.findings-acl.340. BibTeX: https://yuhangzang.github.io/papers/arxiv-2501.12368.bib
- [VideoRoPE: What Makes for Good Video Rotary Position Embedding?](https://yuhangzang.github.io/papers/arxiv-2502.05173.html): International Conference on Machine Learning (ICML), 2025 (Oral). VideoRoPE adapts rotary position embeddings to video through low-frequency temporal allocation, a spatially symmetric diagonal layout and adjustable temporal spacing. arXiv:2502.05173. BibTeX: https://yuhangzang.github.io/papers/arxiv-2502.05173.bib
- [InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD](https://yuhangzang.github.io/papers/arxiv-2404.06512.html): Neural Information Processing Systems (NeurIPS), 2024. InternLM-XComposer2-4KHD uses dynamic patch layouts to support inputs from 336 pixels to 4K HD, improving access to fine-grained visual details across varied resolutions. arXiv:2404.06512, DOI:10.52202/079017-1348. BibTeX: https://yuhangzang.github.io/papers/arxiv-2404.06512.bib
- [MMLongBench-Doc: Benchmarking Long-context Document Understanding with Visualizations](https://yuhangzang.github.io/papers/arxiv-2407.01523.html): Neural Information Processing Systems (NeurIPS), 2024 (Datasets and Benchmarks Track) (Spotlight). MMLongBench-Doc exposes a substantial gap in long-document visual understanding through questions requiring diverse evidence, cross-page reasoning and recognition of unanswerable queries. arXiv:2407.01523, DOI:10.52202/079017-3041. BibTeX: https://yuhangzang.github.io/papers/arxiv-2407.01523.bib
- [Overcoming the Pitfalls of Vision-Language Model Finetuning for OOD Generalization](https://yuhangzang.github.io/papers/arxiv-2401.15914.html): International Conference on Learning Representations (ICLR), 2024. OGEN improves out-of-distribution generalization during vision-language fine-tuning by synthesizing features for unknown classes and using adaptive self-distillation to limit overfitting. arXiv:2401.15914. BibTeX: https://yuhangzang.github.io/papers/arxiv-2401.15914.bib
- [Contextual Object Detection with Multimodal Large Language Models](https://yuhangzang.github.io/papers/arxiv-2305.18279.html): International Journal of Computer Vision (IJCV), 2025. ContextDET grounds objects within language interaction by generating context-relevant object words and then detecting their image locations in an end-to-end multimodal model. arXiv:2305.18279, DOI:10.1007/s11263-024-02214-4. BibTeX: https://yuhangzang.github.io/papers/arxiv-2305.18279.bib
- [Unified Vision and Language Prompt Learning](https://yuhangzang.github.io/papers/arxiv-2210.07225.html): arXiv 2022. Unified Prompt Tuning jointly optimizes visual and textual prompts, combining complementary strengths that neither unimodal prompting method consistently provides alone. arXiv:2210.07225. BibTeX: https://yuhangzang.github.io/papers/arxiv-2210.07225.bib
- [Semi-Supervised and Long-Tailed Object Detection with CascadeMatch](https://yuhangzang.github.io/papers/arxiv-2305.14813.html): International Journal of Computer Vision (IJCV), 2023. CascadeMatch improves semi-supervised long-tailed detection through progressive cascade thresholds, adaptive pseudo-label mining and ensemble supervision that reduces confirmation bias. arXiv:2305.14813, DOI:10.1007/s11263-022-01738-x. BibTeX: https://yuhangzang.github.io/papers/arxiv-2305.14813.bib
- [Open-Vocabulary DETR with Conditional Matching](https://yuhangzang.github.io/papers/arxiv-2203.11876.html): European Conference on Computer Vision (ECCV), 2022 (Oral). OV-DETR reformulates detection as conditional binary matching so a transformer detector can localize objects specified by either a class name or an exemplar image. arXiv:2203.11876, DOI:10.1007/978-3-031-20077-9_7. BibTeX: https://yuhangzang.github.io/papers/arxiv-2203.11876.bib
- [FASA: Feature Augmentation and Sampling Adaptation for Long-Tailed Instance Segmentation](https://yuhangzang.github.io/papers/arxiv-2102.12867.html): IEEE International Conference on Computer Vision (ICCV), 2021. FASA addresses rare-class data scarcity by generating virtual features from observed class statistics and adapting their sampling to the model's training losses. arXiv:2102.12867. BibTeX: https://yuhangzang.github.io/papers/arxiv-2102.12867.bib

## Publications 2026

- [Beyond the Current Observation: Evaluating Multimodal Large Language Models in Controllable Non-Markov Games](https://yuhangzang.github.io/papers/arxiv-2606.19338.html): Empirical Methods in Natural Language Processing (EMNLP), 2026. RNG-Bench separates memory failures from action-selection errors in interactive visual games, finding that forgotten observations account for most residual errors in the evaluated models. arXiv:2606.19338. BibTeX: https://yuhangzang.github.io/papers/arxiv-2606.19338.bib
- [OVO-S-Bench: A Hierarchical Benchmark for Streaming Spatial Intelligence in Multimodal LLMs](https://yuhangzang.github.io/papers/arxiv-2606.03890.html): Empirical Methods in Natural Language Processing (EMNLP), 2026. OVO-S-Bench shows that streaming spatial reasoning remains difficult even for strong multimodal models, with allocentric mapping a major bottleneck under prefix-only video access. arXiv:2606.03890. BibTeX: https://yuhangzang.github.io/papers/arxiv-2606.03890.bib
- [Demo-ICL: In-Context Learning for Procedural Video Knowledge Acquisition](https://yuhangzang.github.io/papers/arxiv-2602.08439.html): Findings of Empirical Methods in Natural Language Processing (Findings of EMNLP), 2026. Demo-ICL trains video-language models to acquire procedural knowledge from in-context demonstrations, rather than relying only on knowledge already stored in their parameters. arXiv:2602.08439. BibTeX: https://yuhangzang.github.io/papers/arxiv-2602.08439.bib
- [Skill-as-Pseudocode: Refactoring Skill Libraries to Pseudocode for LLM Agents](https://yuhangzang.github.io/papers/arxiv-2605.27955.html): Findings of Empirical Methods in Natural Language Processing (Findings of EMNLP), 2026. Skill-as-Pseudocode improves agent use of procedural libraries by pairing typed contracts with concrete action templates; contracts alone do not deliver the same benefit. arXiv:2605.27955. BibTeX: https://yuhangzang.github.io/papers/arxiv-2605.27955.bib
- [EMemBench: Interactive Benchmarking of Episodic Memory for VLM Agents](https://yuhangzang.github.io/papers/arxiv-2601.16690.html): Findings of Empirical Methods in Natural Language Processing (Findings of EMNLP), 2026. EMemBench finds that induction and spatial reasoning remain persistent weaknesses of episodic-memory agents, especially when memories must be grounded in visual environments. arXiv:2601.16690. BibTeX: https://yuhangzang.github.io/papers/arxiv-2601.16690.bib
- [Visionary-R1: Mitigating Shortcuts in Visual Reasoning with Reinforcement Learning](https://yuhangzang.github.io/papers/arxiv-2505.14677.html): Transactions on Machine Learning Research (TMLR), 2026. Visionary-R1 reduces shortcut learning in visual reinforcement learning by requiring image interpretation before reasoning through a caption–reason–answer output format. arXiv:2505.14677. BibTeX: https://yuhangzang.github.io/papers/arxiv-2505.14677.bib
- [EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models](https://yuhangzang.github.io/papers/arxiv-2603.12252.html): European Conference on Computer Vision (ECCV), 2026. EndoCoT couples iterative latent reasoning with diffusion generation, achieving 92.1% average accuracy under task-specific training across the visual reasoning settings in Table 1, compared with 83.8% for DiffThinker. arXiv:2603.12252. BibTeX: https://yuhangzang.github.io/papers/arxiv-2603.12252.bib
- [From Sparse to Dense: Multi-View GRPO for Flow Models via Augmented Condition Space](https://yuhangzang.github.io/papers/arxiv-2603.12648.html): European Conference on Computer Vision (ECCV), 2026. MV-GRPO obtains richer preference-learning signals by evaluating the same generated samples against multiple related captions, without regenerating those samples. arXiv:2603.12648. BibTeX: https://yuhangzang.github.io/papers/arxiv-2603.12648.bib
- [MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing](https://yuhangzang.github.io/papers/arxiv-2509.22186.html): Association for Computational Linguistics (ACL), Industry Track, 2026. MinerU2.5 decouples low-resolution layout analysis from native-resolution content recognition, preserving document detail while reducing high-resolution processing costs. arXiv:2509.22186, DOI:10.18653/v1/2026.acl-industry.3. BibTeX: https://yuhangzang.github.io/papers/arxiv-2509.22186.bib
- [SEAgent: Self-Evolving Computer Use Agent with Autonomous Learning from Experience](https://yuhangzang.github.io/papers/arxiv-2508.04700.html): International Conference on Machine Learning (ICML), 2026. SEAgent improves computer-use agents on unfamiliar software through autonomous exploration, curriculum generation and policy updates from successful and failed experience. arXiv:2508.04700. BibTeX: https://yuhangzang.github.io/papers/arxiv-2508.04700.bib
- [ARM-Thinker: Reinforcing Multimodal Generative Reward Models with Agentic Tool Use and Visual Reasoning](https://yuhangzang.github.io/papers/arxiv-2512.05111.html): IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2026. ARM-Thinker turns multimodal reward scoring into an evidence-seeking process by learning when to crop images, retrieve document pages and verify reasoning claims. arXiv:2512.05111. BibTeX: https://yuhangzang.github.io/papers/arxiv-2512.05111.bib
- [Think Visually, Reason Textually: Vision-Language Synergy in ARC](https://yuhangzang.github.io/papers/arxiv-2511.15703.html): IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2026. ARC-VL combines visual pattern abstraction and verification with textual rule execution, addressing the complementary strengths of vision and language on ARC-AGI puzzles. arXiv:2511.15703. BibTeX: https://yuhangzang.github.io/papers/arxiv-2511.15703.bib
- [Fine-Grained GRPO for Precise Preference Alignment in Flow Models](https://yuhangzang.github.io/papers/arxiv-2510.01982.html): IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2026. G²RPO improves credit assignment in flow-model reinforcement learning by isolating stochastic perturbations and integrating advantages across multiple denoising granularities. arXiv:2510.01982. BibTeX: https://yuhangzang.github.io/papers/arxiv-2510.01982.bib
- [Spatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning](https://yuhangzang.github.io/papers/arxiv-2510.27606.html): IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2026. Spatial-SSRL derives verifiable spatial rewards from ordinary RGB or RGB-D images, improving spatial understanding without human or LVLM annotations for its pretext tasks. arXiv:2510.27606. BibTeX: https://yuhangzang.github.io/papers/arxiv-2510.27606.bib
- [TRivia: Self-supervised Fine-tuning of Vision-Language Models for Table Recognition](https://yuhangzang.github.io/papers/arxiv-2512.01248.html): IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2026. TRivia improves table recognition from unlabeled table images by rewarding whether recognized table content can correctly answer automatically generated questions. arXiv:2512.01248. BibTeX: https://yuhangzang.github.io/papers/arxiv-2512.01248.bib
- [Beyond Fixed: Training-Free Variable-Length Denoising for Diffusion Large Language Models](https://yuhangzang.github.io/papers/arxiv-2508.00819.html): International Conference on Learning Representations (ICLR), 2026. DAEDAL enables training-free variable-length generation in diffusion language models by expanding the response length before and during denoising when internal signals indicate insufficient space. arXiv:2508.00819. BibTeX: https://yuhangzang.github.io/papers/arxiv-2508.00819.bib
- [DiCache: Let Diffusion Model Determine Its Own Cache](https://yuhangzang.github.io/papers/arxiv-2508.17356.html): International Conference on Learning Representations (ICLR), 2026. DiCache lets each diffusion sample determine its caching schedule and cache reuse through shallow-layer probes, reducing dependence on fixed caching rules. arXiv:2508.17356. BibTeX: https://yuhangzang.github.io/papers/arxiv-2508.17356.bib
- [Advancing Complex Video Object Segmentation via Progressive Concept Construction](https://yuhangzang.github.io/papers/arxiv-2507.15852.html): International Conference on Learning Representations (ICLR), 2026. SeC improves video object segmentation through progressively constructed object concepts, helping maintain identity across large appearance changes and scene transitions. arXiv:2507.15852. BibTeX: https://yuhangzang.github.io/papers/arxiv-2507.15852.bib
- [SIM-CoT: Supervised Implicit Chain-of-Thought](https://yuhangzang.github.io/papers/arxiv-2509.20317.html): International Conference on Learning Representations (ICLR), 2026. SIM-CoT stabilizes implicit reasoning by supervising each latent reasoning step with an auxiliary decoder that is removed at inference. arXiv:2509.20317. BibTeX: https://yuhangzang.github.io/papers/arxiv-2509.20317.bib
- [ScaleCap: Inference-Time Scalable Image Captioning via Dual-Modality Debiasing](https://yuhangzang.github.io/papers/arxiv-2506.19848.html): International Conference on Learning Representations (ICLR), 2026. ScaleCap turns additional inference budget into richer, better-grounded image captions by iteratively adding missing visual details and removing hallucinated descriptions. arXiv:2506.19848. BibTeX: https://yuhangzang.github.io/papers/arxiv-2506.19848.bib
- [STAR-Bench: Probing Deep Spatio-Temporal Reasoning as Audio 4D Intelligence](https://yuhangzang.github.io/papers/arxiv-2510.24693.html): International Conference on Learning Representations (ICLR), 2026. STAR-Bench probes audio reasoning that text captions cannot adequately replace, exposing weaknesses in understanding sound dynamics across time and three-dimensional space. arXiv:2510.24693. BibTeX: https://yuhangzang.github.io/papers/arxiv-2510.24693.bib
- [Visual Self-Refine: A Pixel-Guided Paradigm for Accurate Chart Parsing](https://yuhangzang.github.io/papers/arxiv-2602.16455.html): International Conference on Learning Representations (ICLR), 2026. ChartVSR refines pixel-level data-point locations through visual feedback before decoding chart values, giving chart parsing an explicit visual self-correction process. arXiv:2602.16455. BibTeX: https://yuhangzang.github.io/papers/arxiv-2602.16455.bib
- [CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning](https://yuhangzang.github.io/papers/arxiv-2509.22647.html): International Conference on Learning Representations (ICLR), 2026. CapRL trains dense captioners with verifiable question-answering rewards: a useful caption should let a vision-free language model answer questions about the image. arXiv:2509.22647. BibTeX: https://yuhangzang.github.io/papers/arxiv-2509.22647.bib
- [RAR: Retrieving And Ranking Augmented MLLMs for Visual Recognition](https://yuhangzang.github.io/papers/arxiv-2403.13805.html): IEEE Transactions on Image Processing (TIP), 2026. RAR combines CLIP retrieval with MLLM reranking to recognize fine-grained categories without placing an entire large vocabulary into the language model's context. arXiv:2403.13805, DOI:10.1109/tip.2025.3644175. BibTeX: https://yuhangzang.github.io/papers/arxiv-2403.13805.bib

## Publications 2025

- [Unified Multimodal Chain-of-Thought Reward Model through Reinforcement Fine-Tuning](https://yuhangzang.github.io/papers/arxiv-2505.03318.html): Neural Information Processing Systems (NeurIPS), 2025. UnifiedReward-Think strengthens multimodal reward judgments through explicit multi-step reasoning and reinforcement fine-tuning across visual understanding and generation tasks. arXiv:2505.03318, DOI:10.52202/085713-5315. BibTeX: https://yuhangzang.github.io/papers/arxiv-2505.03318.bib
- [HiFlow: Training-free High-Resolution Image Generation with Flow-Aligned Guidance](https://yuhangzang.github.io/papers/arxiv-2504.06232.html): Neural Information Processing Systems (NeurIPS), 2025. HiFlow improves training-free high-resolution generation by aligning initialization, direction and acceleration with a reference flow derived from lower-resolution generation. arXiv:2504.06232, DOI:10.52202/085713-4831. BibTeX: https://yuhangzang.github.io/papers/arxiv-2504.06232.bib
- [Visual-RFT: Visual Reinforcement Fine-Tuning](https://yuhangzang.github.io/papers/arxiv-2503.01785.html): IEEE International Conference on Computer Vision (ICCV), 2025. Visual-RFT shows that GRPO with task-specific verifiable rewards improves classification, detection, and grounding in large vision-language models trained with limited labeled data. arXiv:2503.01785. BibTeX: https://yuhangzang.github.io/papers/arxiv-2503.01785.bib
- [MM-IFEngine: Towards Multimodal Instruction Following](https://yuhangzang.github.io/papers/arxiv-2504.07957.html): IEEE International Conference on Computer Vision (ICCV), 2025. MM-IFEngine improves multimodal instruction following through constraint-rich training data and evaluation that checks both requested output form and image-grounded requirements. arXiv:2504.07957. BibTeX: https://yuhangzang.github.io/papers/arxiv-2504.07957.bib
- [X-Prompt: Towards Universal In-Context Image Generation in Auto-Regressive Vision Language Foundation Models](https://yuhangzang.github.io/papers/arxiv-2412.01824.html): IEEE International Conference on Computer Vision (ICCV), 2025. X-Prompt extends autoregressive vision-language modeling to in-context image generation, allowing example prompts to specify both familiar and previously unseen generation tasks. arXiv:2412.01824. BibTeX: https://yuhangzang.github.io/papers/arxiv-2412.01824.bib
- [Bootstrap3D: Improving Multi-view Diffusion Model with Synthetic Data](https://yuhangzang.github.io/papers/arxiv-2406.00093.html): IEEE International Conference on Computer Vision (ICCV), 2025. Bootstrap3D uses filtered and recaptioned synthetic multi-view images to improve 3D generation while preserving the strengths of pretrained 2D diffusion models. arXiv:2406.00093. BibTeX: https://yuhangzang.github.io/papers/arxiv-2406.00093.bib
- [Bootstrapping Grounded Chain-of-Thought in Multimodal LLMs for Data-Efficient Model Adaptation](https://yuhangzang.github.io/papers/arxiv-2507.02859.html): IEEE International Conference on Computer Vision (ICCV), 2025 (Highlight). Grounded Chain-of-Thought improves data-efficient adaptation by anchoring distilled reasoning steps to image regions, reducing factual errors in specialized visual tasks. arXiv:2507.02859. BibTeX: https://yuhangzang.github.io/papers/arxiv-2507.02859.bib
- [Light-A-Video: Training-free Video Relighting via Progressive Light Fusion](https://yuhangzang.github.io/papers/arxiv-2502.08590.html): IEEE International Conference on Computer Vision (ICCV), 2025. Light-A-Video adapts image relighting models to temporally coherent video relighting without additional training by combining cross-frame attention and progressive light blending. arXiv:2502.08590. BibTeX: https://yuhangzang.github.io/papers/arxiv-2502.08590.bib
- [Deciphering Cross-Modal Alignment in Large Vision-Language Models with Modality Integration Rate](https://yuhangzang.github.io/papers/arxiv-2410.07167.html): IEEE International Conference on Computer Vision (ICCV), 2025. Modality Integration Rate estimates LVLM pretraining quality from cross-modal distribution distance, helping assess alignment without first running costly supervised fine-tuning. arXiv:2410.07167. BibTeX: https://yuhangzang.github.io/papers/arxiv-2410.07167.bib
- [SAM2Long: Enhancing SAM 2 for Long Video Segmentation with a Training-Free Memory Tree](https://yuhangzang.github.io/papers/arxiv-2410.16268.html): IEEE International Conference on Computer Vision (ICCV), 2025. SAM2Long reduces error accumulation in SAM 2 by retaining multiple segmentation paths and selecting the best video-level trajectory through a constrained memory-tree search. arXiv:2410.16268. BibTeX: https://yuhangzang.github.io/papers/arxiv-2410.16268.bib
- [InternLM-XComposer2.5-Reward: A Simple Yet Effective Multi-Modal Reward Model](https://yuhangzang.github.io/papers/arxiv-2501.12368.html): Findings of the Association for Computational Linguistics (ACL), 2025. IXC-2.5-Reward provides a multimodal preference signal that supports reinforcement learning, best-response selection and instruction-data filtering across text, image and video inputs. arXiv:2501.12368, DOI:10.18653/v1/2025.findings-acl.340. BibTeX: https://yuhangzang.github.io/papers/arxiv-2501.12368.bib
- [Towards Storage-Efficient Visual Document Retrieval: An Empirical Study on Reducing Patch-Level Embeddings](https://yuhangzang.github.io/papers/arxiv-2506.04997.html): Findings of the Association for Computational Linguistics (ACL), 2025. Light-ColPali/ColQwen2 retains 98.2% of the original retrieval performance with 11.8% of the embedding memory by merging document patches instead of pruning them. arXiv:2506.04997, DOI:10.18653/v1/2025.findings-acl.1003. BibTeX: https://yuhangzang.github.io/papers/arxiv-2506.04997.bib
- [VideoRoPE: What Makes for Good Video Rotary Position Embedding?](https://yuhangzang.github.io/papers/arxiv-2502.05173.html): International Conference on Machine Learning (ICML), 2025 (Oral). VideoRoPE adapts rotary position embeddings to video through low-frequency temporal allocation, a spatially symmetric diagonal layout and adjustable temporal spacing. arXiv:2502.05173. BibTeX: https://yuhangzang.github.io/papers/arxiv-2502.05173.bib
- [SongGen: A Single Stage Auto-regressive Transformer for Text-to-Song Generation](https://yuhangzang.github.io/papers/arxiv-2502.13128.html): International Conference on Machine Learning (ICML), 2025. SongGen generates controllable vocals and accompaniment in a single autoregressive stage, with either mixed audio output or separate vocal and accompaniment tracks. arXiv:2502.13128. BibTeX: https://yuhangzang.github.io/papers/arxiv-2502.13128.bib
- [ByTheWay: Boost Your Text-to-Video Generation Model to Higher Quality in a Training-free Way](https://yuhangzang.github.io/papers/arxiv-2410.06241.html): IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. ByTheWay improves video consistency and motion without training by reducing disagreement between temporal attention maps and amplifying their motion-related frequency content. arXiv:2410.06241. BibTeX: https://yuhangzang.github.io/papers/arxiv-2410.06241.bib
- [OVO-Bench: How Far is Your Video-LLMs from Real-World Online Video Understanding?](https://yuhangzang.github.io/papers/arxiv-2501.05510.html): IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. OVO-Bench evaluates whether video models respond appropriately at a specified moment, distinguishing backward tracing, real-time understanding and waiting for future evidence. arXiv:2501.05510. BibTeX: https://yuhangzang.github.io/papers/arxiv-2501.05510.bib
- [Dispider: Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction](https://yuhangzang.github.io/papers/arxiv-2501.03218.html): IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. Dispider separates perception, decision and response so a video assistant can keep monitoring a stream while asynchronously generating an interaction. arXiv:2501.03218. BibTeX: https://yuhangzang.github.io/papers/arxiv-2501.03218.bib
- [PyramidDrop: Accelerating Your Large Vision-Language Models via Pyramid Visual Redundancy Reduction](https://yuhangzang.github.io/papers/arxiv-2410.17247.html): IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. PyramidDrop retains visual detail in shallow LVLM layers and progressively removes redundant tokens in deeper layers to reduce training and inference cost. arXiv:2410.17247. BibTeX: https://yuhangzang.github.io/papers/arxiv-2410.17247.bib
- [WildAvatar: Learning In-the-wild 3D Avatars from the Web](https://yuhangzang.github.io/papers/arxiv-2407.02165.html): IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. WildAvatar scales 3D human avatar data beyond laboratory capture by automatically annotating and filtering web videos with more than 10,000 subjects and scenes. arXiv:2407.02165. BibTeX: https://yuhangzang.github.io/papers/arxiv-2407.02165.bib
- [MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models](https://yuhangzang.github.io/papers/arxiv-2410.17637.html): International Conference on Learning Representations (ICLR), 2025. MIA-DPO creates multi-image preference-training pairs from single-image data and attention signals, reducing the need for new human annotations or external judge models. arXiv:2410.17637. BibTeX: https://yuhangzang.github.io/papers/arxiv-2410.17637.bib
- [MotionClone: Training-Free Motion Cloning for Controllable Video Generation](https://yuhangzang.github.io/papers/arxiv-2406.05338.html): International Conference on Learning Representations (ICLR), 2025. MotionClone transfers motion from a reference video without training by using sparse temporal attention as guidance for text-to-video and image-to-video generation. arXiv:2406.05338. BibTeX: https://yuhangzang.github.io/papers/arxiv-2406.05338.bib
- [Contextual Object Detection with Multimodal Large Language Models](https://yuhangzang.github.io/papers/arxiv-2305.18279.html): International Journal of Computer Vision (IJCV), 2025. ContextDET grounds objects within language interaction by generating context-relevant object words and then detecting their image locations in an end-to-end multimodal model. arXiv:2305.18279, DOI:10.1007/s11263-024-02214-4. BibTeX: https://yuhangzang.github.io/papers/arxiv-2305.18279.bib

## Publications 2024

- [VLMEvalKit: An Open-Source Toolkit for Evaluating Large Multi-Modality Models](https://yuhangzang.github.io/papers/arxiv-2407.11691.html): ACM Multimedia (ACM MM), 2024 (Open Source Software Competition). VLMEvalKit standardizes multimodal evaluation behind a common interface that handles data preparation, distributed inference, post-processing and metrics for reproducible comparisons. arXiv:2407.11691, DOI:10.1145/3664647.3685520. BibTeX: https://yuhangzang.github.io/papers/arxiv-2407.11691.bib
- [MMLongBench-Doc: Benchmarking Long-context Document Understanding with Visualizations](https://yuhangzang.github.io/papers/arxiv-2407.01523.html): Neural Information Processing Systems (NeurIPS), 2024 (Datasets and Benchmarks Track) (Spotlight). MMLongBench-Doc exposes a substantial gap in long-document visual understanding through questions requiring diverse evidence, cross-page reasoning and recognition of unanswerable queries. arXiv:2407.01523, DOI:10.52202/079017-3041. BibTeX: https://yuhangzang.github.io/papers/arxiv-2407.01523.bib
- [MMDU: A Multi-Turn Multi-Image Dialog Understanding Benchmark and Instruction-Tuning Dataset for LVLMs](https://yuhangzang.github.io/papers/arxiv-2406.11833.html): Neural Information Processing Systems (NeurIPS), 2024 (Datasets and Benchmarks Track). MMDU evaluates sustained multi-image dialogue, while MMDU-45k provides instruction data that helps open LVLMs close the conversational gap identified by the benchmark. arXiv:2406.11833, DOI:10.52202/079017-0278. BibTeX: https://yuhangzang.github.io/papers/arxiv-2406.11833.bib
- [ShareGPT4Video: Improving Video Understanding and Generation with Better Captions](https://yuhangzang.github.io/papers/arxiv-2406.04325.html): Neural Information Processing Systems (NeurIPS), 2024 (Datasets and Benchmarks Track). ShareGPT4Video improves video supervision with dense captions that describe both frame-level detail and precise temporal changes, then scales annotation through a dedicated video captioner. arXiv:2406.04325, DOI:10.52202/079017-0614. BibTeX: https://yuhangzang.github.io/papers/arxiv-2406.04325.bib
- [Streaming Long Video Understanding with Large Language Models](https://yuhangzang.github.io/papers/arxiv-2405.16009.html): Neural Information Processing Systems (NeurIPS), 2024. VideoStreaming combines propagated clip memory with question-conditioned memory selection so an LLM can answer about long videos using a fixed number of selected video tokens. arXiv:2405.16009, DOI:10.52202/079017-3792. BibTeX: https://yuhangzang.github.io/papers/arxiv-2405.16009.bib
- [Are We on the Right Way for Evaluating Large Vision-Language Models?](https://yuhangzang.github.io/papers/arxiv-2403.20330.html): Neural Information Processing Systems (NeurIPS), 2024. MMStar tests whether multimodal gains actually depend on visual input, using visually necessary questions and metrics designed to reveal language shortcuts and possible data leakage. arXiv:2403.20330, DOI:10.52202/079017-0850. BibTeX: https://yuhangzang.github.io/papers/arxiv-2403.20330.bib
- [InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD](https://yuhangzang.github.io/papers/arxiv-2404.06512.html): Neural Information Processing Systems (NeurIPS), 2024. InternLM-XComposer2-4KHD uses dynamic patch layouts to support inputs from 336 pixels to 4K HD, improving access to fine-grained visual details across varied resolutions. arXiv:2404.06512, DOI:10.52202/079017-1348. BibTeX: https://yuhangzang.github.io/papers/arxiv-2404.06512.bib
- [MVSGaussian: Fast Generalizable Gaussian Splatting Reconstruction from Multi-View Stereo](https://yuhangzang.github.io/papers/arxiv-2405.12218.html): European Conference on Computer Vision (ECCV), 2024. MVSGaussian derives geometry-aware Gaussian representations from multi-view stereo, enabling generalizable scene reconstruction, real-time rendering and fast per-scene refinement. arXiv:2405.12218, DOI:10.1007/978-3-031-72649-1_3. BibTeX: https://yuhangzang.github.io/papers/arxiv-2405.12218.bib
- [Long-CLIP: Unlocking the Long-Text Capability of CLIP](https://yuhangzang.github.io/papers/arxiv-2403.15378.html): European Conference on Computer Vision (ECCV), 2024. Long-CLIP extends CLIP to detailed text descriptions through efficient fine-tuning while preserving the embedding alignment needed for plug-and-play downstream use. arXiv:2403.15378, DOI:10.1007/978-3-031-72983-6_18. BibTeX: https://yuhangzang.github.io/papers/arxiv-2403.15378.bib
- [Alpha-CLIP: A CLIP Model Focusing on Wherever You Want](https://yuhangzang.github.io/papers/arxiv-2312.03818.html): IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. Alpha-CLIP adds an alpha channel to focus CLIP on a selected image region while retaining whole-image recognition and supporting region-aware understanding and generation. arXiv:2312.03818. BibTeX: https://yuhangzang.github.io/papers/arxiv-2312.03818.bib
- [Overcoming the Pitfalls of Vision-Language Model Finetuning for OOD Generalization](https://yuhangzang.github.io/papers/arxiv-2401.15914.html): International Conference on Learning Representations (ICLR), 2024. OGEN improves out-of-distribution generalization during vision-language fine-tuning by synthesizing features for unknown classes and using adaptive self-distillation to limit overfitting. arXiv:2401.15914. BibTeX: https://yuhangzang.github.io/papers/arxiv-2401.15914.bib

## Publications 2023

- [Real-World Object Detection](https://yuhangzang.github.io/papers/scholar-hW23VKIAAAAJ-eQOLeE2rZwMC.html): PhD Thesis, Nanyang Technological University, 2023. This thesis addresses two obstacles to real-world object detection—long-tailed data and unseen categories—through adaptive training, vision-language transfer and context-aware grounding. DOI:10.32657/10356/171489. BibTeX: https://yuhangzang.github.io/papers/scholar-hW23VKIAAAAJ-eQOLeE2rZwMC.bib
- [Semi-Supervised and Long-Tailed Object Detection with CascadeMatch](https://yuhangzang.github.io/papers/arxiv-2305.14813.html): International Journal of Computer Vision (IJCV), 2023. CascadeMatch improves semi-supervised long-tailed detection through progressive cascade thresholds, adaptive pseudo-label mining and ensemble supervision that reduces confirmation bias. arXiv:2305.14813, DOI:10.1007/s11263-022-01738-x. BibTeX: https://yuhangzang.github.io/papers/arxiv-2305.14813.bib

## Publications 2022

- [Unified Vision and Language Prompt Learning](https://yuhangzang.github.io/papers/arxiv-2210.07225.html): arXiv 2022. Unified Prompt Tuning jointly optimizes visual and textual prompts, combining complementary strengths that neither unimodal prompting method consistently provides alone. arXiv:2210.07225. BibTeX: https://yuhangzang.github.io/papers/arxiv-2210.07225.bib
- [Open-Vocabulary DETR with Conditional Matching](https://yuhangzang.github.io/papers/arxiv-2203.11876.html): European Conference on Computer Vision (ECCV), 2022 (Oral). OV-DETR reformulates detection as conditional binary matching so a transformer detector can localize objects specified by either a class name or an exemplar image. arXiv:2203.11876, DOI:10.1007/978-3-031-20077-9_7. BibTeX: https://yuhangzang.github.io/papers/arxiv-2203.11876.bib

## Publications 2021

- [FASA: Feature Augmentation and Sampling Adaptation for Long-Tailed Instance Segmentation](https://yuhangzang.github.io/papers/arxiv-2102.12867.html): IEEE International Conference on Computer Vision (ICCV), 2021. FASA addresses rare-class data scarcity by generating virtual features from observed class statistics and adapting their sampling to the model's training losses. arXiv:2102.12867. BibTeX: https://yuhangzang.github.io/papers/arxiv-2102.12867.bib
- [Seesaw Loss for Long-Tailed Instance Segmentation](https://yuhangzang.github.io/papers/arxiv-2008.10032.html): IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. Seesaw Loss balances rare-class learning by reducing overwhelming negative gradients while restoring penalties for misclassified examples to control false positives. arXiv:2008.10032. BibTeX: https://yuhangzang.github.io/papers/arxiv-2008.10032.bib

## Publications 2020

- [KPNet: Towards Minimal Face Detector](https://yuhangzang.github.io/papers/arxiv-2003.07543.html): AAAI Conference on Artificial Intelligence (AAAI), 2020. KPNet builds a compact face detector by predicting facial keypoints first and deriving bounding boxes from them, avoiding complex anchor and detection-head designs. arXiv:2003.07543, DOI:10.1609/aaai.v34i07.6878. BibTeX: https://yuhangzang.github.io/papers/arxiv-2003.07543.bib

## Publications 2019

- [Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network](https://yuhangzang.github.io/papers/arxiv-1908.05900.html): IEEE International Conference on Computer Vision (ICCV), 2019. Pixel Aggregation Network detects arbitrary-shaped scene text efficiently by combining a lightweight segmentation head with learned pixel grouping. arXiv:1908.05900. BibTeX: https://yuhangzang.github.io/papers/arxiv-1908.05900.bib
- [Scene Text Detection with Supervised Pyramid Context Network](https://yuhangzang.github.io/papers/arxiv-1811.08605.html): AAAI Conference on Artificial Intelligence (AAAI), 2019. SPCNET uses supervised semantic context within a feature-pyramid instance-segmentation model to localize scene text and suppress false positives with little extra computation. arXiv:1811.08605, DOI:10.1609/aaai.v33i01.33019038. BibTeX: https://yuhangzang.github.io/papers/arxiv-1811.08605.bib

## Pages

- [Home](https://yuhangzang.github.io/): biography, news and selected publications
- [All publications](https://yuhangzang.github.io/research.html): the complete list with filters by year, topic and venue