Learning after pre-training
Mid-Training, SFT, reinforcement learning, and on-policy distillation for capable, reliable foundation models.
UCAS / CASIA / BEIJING
Wulin Xie
Master’s student · Class of 2028
I am a master’s student in Pattern Recognition and Intelligent Systems at the University of Chinese Academy of Sciences and the Institute of Automation, Chinese Academy of Sciences.
My research focuses on LLM mid/post-training, agents, and multimodal reinforcement learning. I am interested in how models learn to reason, use tools, and connect understanding with generation.
I am currently a research intern at Meituan LongCat Interaction through the Beidou Talent Program, following research internships at Alibaba Amap and vivo.
300+Google Scholar citations
4Accepted first / co-first papers
3Industry research internships
As of Sep. 2026
01Research interests
Agents that learn by doing
Tool use, long-horizon interaction, and efficient learning from multi-turn agent experience.
Understanding × generation
Unified multimodal evaluation, video understanding, visual reasoning, and preference alignment.
Recent highlights
- VI-Bench accepted to NeurIPS 2026 as a first-author paper.
- Joined Meituan LongCat Interaction through the Beidou Talent Program.
- MME-Unify accepted to ICLR 2026 as a co-first-author paper.
02Selected publications
All on Scholar ↗A selection of my work on learning, alignment, and evaluation. First / co-first-author work is highlighted.
7 publications
VI-Bench: Benchmarking Prompt Inversion from AIGC Videos
Can a model recover the prompt behind a generated video? A closed-loop benchmark with 900 verified videos and 18 VLMs connects video understanding to generation control.
MME-Unify: A Comprehensive Benchmark for Unified Multimodal Understanding and Generation Models
Evaluating understanding and generation together: 4,104 examples across 30 subtasks reveal the gap between strong individual capabilities and truly unified multimodal reasoning.
Q-Hawkeye: Reliable Visual Policy Optimization for Image Quality Assessment
Making visual reinforcement learning more reliable with uncertainty-aware optimization and perceptual supervision. KADID PLCC / SRCC improves from 0.723 / 0.719 to 0.779 / 0.775 over VisualQuality-R1.
MM-RLHF: The Next Step Forward in Multimodal LLM Alignment
Multimodal preference alignment with 120K human-annotated preference pairs, critique-based reward modeling, and dynamic preference optimization, evaluated on 27 benchmarks.
MME-VideoOCR: Evaluating OCR-Based Capabilities of Multimodal LLMs in Video Scenarios
Beyond reading a single frame: 2,000 human-annotated questions probe text recognition, cross-frame integration, and spatiotemporal reasoning in video.
Multi-View Factorizing and Disentangling: A Novel Framework for Incomplete Multi-View Multi-Label Classification
Learning shared and view-specific representations under missing views and labels. Evaluated against 10 methods on five datasets, with stronger average precision under 50% missingness.
Uncertainty-Aware Pseudo-Labeling and Dual Graph Driven Network for Incomplete Multi-View Multi-Label Classification
Turning incomplete supervision into reliable learning signals through uncertainty-aware pseudo-labeling and label-guided dual graph constraints, validated on six datasets.
Additional research & journal articles
- A Survey of Unified Multimodal Understanding and Generation: Advances and Challenges
Y. Yang, H. Tian, Y. Shi, Wulin Xie, et al. · TPAMI · Under Review
- Confidence-Enhanced Dual-Space Semantic Alignment for Partial Multi-View Incomplete Multi-Label Classification
J. Chen, Wulin Xie, et al. · Knowledge-Based Systems
- Cross-modal mapping: Mitigating the modality gap for few-shot classification
X. Yang, Wulin Xie, et al. · Pattern Recognition
03Research experience
Meituan · LongCat Interaction
2026.04 — PresentBeidou Talent Program · Research Intern
- Working on Agentic Mid-Training and foundation-model post-training, spanning data feedback, SFT, domain-specific RL, and on-policy distillation.
- Contributed to a 35B-A3B customer-service foundation model serving approximately 80 million calls per day.
- DivOPD: improving learning efficiency for multi-turn agents. Across six settings, mean Final-5 success rate improves by 7.1 percentage points, with 1.87× learner GPU efficiency at matched target success rates. ICLR 2027, under review.
Alibaba · Amap
2025.08 — 2026.03Algorithm Research Intern
- Developed Q-Hawkeye for reliable reinforcement learning in image quality assessment, addressing noisy optimization signals and weak visual grounding. Evaluated on eight IQA benchmarks.
- First-author submission to AAAI 2027; code released through AMAP-ML.
vivo
2024.12 — 2025.04Algorithm Research Intern
- Led benchmark construction for MME-Unify, bringing multimodal understanding and generation into a unified evaluation protocol: 4,104 examples, 30 subtasks, 22 evaluated models.
- ICLR 2026 · Co-first author.
Institute of Automation, CAS
2024.12 — 2025.08Research Intern
- Research on multimodal evaluation and alignment, including MM-RLHF and MME-VideoOCR. Contributions span preference learning, reward modeling, and video understanding benchmarks.
- ICML 2025 / NeurIPS 2025 · Core contributor. Continuing research includes VI-Bench (NeurIPS 2026, first author).
Harbin Institute of Technology, Shenzhen
2023.08 — 2024.07Research Intern
- Studied incomplete multi-view multi-label learning and cross-modal few-shot classification, focusing on reliable pseudo-labels, disentangled representations, and image–text alignment.
- ACM MM 2024 / WACV 2025 Oral · First author; Pattern Recognition · Second author.
04Education
2025 — 2028
University of Chinese Academy of Sciences
Institute of Automation, Chinese Academy of Sciences
M.Sc. in Pattern Recognition and Intelligent Systems · Expected Jul. 2028
2021 — 2025
Guizhou University
B.Sc. in Information Management and Information Systems
GPA: 4.25 / 5.0
LET’S CONNECT
Building models that do more.
I’m looking for research internships in foundation models, agents, and multimodal learning. If our interests overlap, I’d love to hear from you.
xwl1085930920@gmail.com