I am a Master’s student in Artificial Intelligence at the University of Chinese Academy of Sciences (UCAS), Hangzhou Institute for Advanced Study (expected graduation: July 2027). I work with Prof. Jianqiang Huang (黄建强) at the Computer Network Information Center, CAS, on AI for science — multimodal models, post-training, and rigorous evaluation. I received my B.S. in Data Science and Big Data Technology from Nankai University in 2024 (recommended admission).

My research spans LLM post-training (SFT & reinforcement learning), multimodal perception and evaluation, and agent systems (automated evaluation, reward design, and data flywheels). I have 2 CCF-A papers accepted (CVPR 2026, ACM MM 2025) and 4 more under review at top venues, with internships at Alibaba (Qwen Agent), Bilibili, and Shopee validating end-to-end industrial deployment.

Research interests

• Multimodal LLM post-training (GRPO, DAPO, dual-reward RL, lightweight alignment)
• Caption and video evaluation beyond reference-based metrics
• Agent RL: tool-use rewards, LLM-as-judge, and evaluation data flywheels
• Time-series and scientific AI benchmarks (ECG, medical imaging)

News

• 2026.04 – present: Algorithm intern at Alibaba — Qwen shopping-agent post-training & data flywheel
• 2026: CCCaption accepted to CVPR 2026 (first author)
• 2025.12 – 2026.03: Algorithm intern at Bilibili — long-video evaluation & Wan2.1 RL
• 2025: ECG time-series benchmark accepted to ACM MM 2025 (first author)
• 2024: National Scholarship (top 2%) at UCAS

Selected publications

Representative work with method overviews. See the full publication list.

CCCaption: Dual-Reward RL for Complete and Correct Image Captioning

CVPR 2026 (accepted) · First author. Proposes completeness and correctness dual rewards with dynamic query sampling for efficient RL post-training of image captioning. CCCaption-2B achieves capability comparable to Qwen3-VL-32B.

Paper

Long-CODE: Isolating Pure Long-Context as an Orthogonal Dimension in Video Evaluation

ACM MM 2026 (under review) · First author. Decouples long-context video quality from short-video metrics via Dynamic Structure Alignment and MLLM-based semantic scoring. The Long-CODE benchmark isolates narrative and causal consistency with strong human correlation (Spearman 0.76).

Paper

A Comprehensive Benchmark for Electrocardiogram Time-Series

ACM MM 2025 (accepted) · First author. Unified benchmark spanning classification, detection, forecasting, and generation for ECG signals. Introduces the FFD metric and PSSM model, improving over SOTA by +1.6% to +67.2% across tasks.

Paper

Open source

• Awesome-Algorithm — algorithm practice & interview notes

Contact

Email: tang_zhi_jiang@163.com · GitHub: ZhijiangTang