My research focuses on reducing intelligent models’ dependence on human-generated data. Weakly supervised learning provided the initial point of departure: how can effective learning signals be derived from limited annotations? As this research progressed, the question extended from the data space to the weight space. Through model merging, multi-task fusion, weight generation, and model editing, my work explores how to reuse, combine, and update the knowledge and capabilities already encoded in models. This trajectory has naturally expanded toward large language models and foundation models.
This line of research further points toward reinforcement learning-based paradigms for autonomous learning. Through exploration, interaction, and feedback, models can generate learning signals beyond existing human corpora and annotations, gradually shifting from imitating established knowledge to acquiring and creating knowledge autonomously. This may offer a promising path toward artificial general intelligence.
If you are interested in my current research or technical blog, please feel free to contact me by email.
Research Interests
- Reinforcement Learning for Autonomous Intelligence
- Weight Generation and Model Editing
- Weight-Space Learning and Model Fusion
- Data-Efficient and Weakly Supervised Learning
Current Projects
- Reinforcement Learning with Verifiable Rewards and Post-Training active
Explore how constructing verifiable rewards and encouraging models to sample and explore over a broader space can enhance reinforcement learning, thereby improving its effectiveness and applicability in post-training.
- Scholarly: A Self-Evolving, End-to-End Research Engine active
An entrepreneurial project focused on building fully automated research environments, developing and deploying research agents, and exploring how they can drive the training of next-generation foundation models for scientific research.
- Interpretability and Editability of Large Language Model Weight Spaces active
This research focuses on editing knowledge and experience encoded in large language model weights, as well as weight-space-based meta-learning. It explores how a deeper understanding of neural network weight spaces can guide the continual updating of model knowledge and behavior.
Selected Publications
- AnyEdit++: Adaptive Long-Form Knowledge Editing via Bayesian Surprise
Bowen Tian, Caixue He, Jiemin Wu, Jingying Wang, Wenshuo Chen, Zexi Li, Yutao Yue
ICML 2026 · CCF-A arXivcode - Text2Weight: Bridging Natural Language and Neural Network Weight Spaces
Bowen Tian, Wenshuo Chen, Zexi Li, Songning Lai, Jiemin Wu, Yutao Yue
ACM MM 2025 · CCF-A arXivcode - PEPL: Precision-Enhanced Pseudo-Labeling for Fine-Grained Image Classification in Semi-Supervised Learning
Bowen Tian, Songning Lai, Lujundong Li, Zhihao Shuai, Runwei Guan, Tian Wu, Yutao Yue
IEEE ICASSP 2025 · CCF-B arXiv - Beyond Task Vectors: Selective Task Arithmetic Based on Importance Metrics
Bowen Tian, Songning Lai, Jiemin Wu, Zhihao Shuai, Shiming Ge, Yutao Yue
arXiv 2024 arXiv