Papers by Zijun Wu
Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures (2026.acl-long)
Copied to clipboard
Yi Hu, Jiaqi Gu, Ruxin Wang, Zijun Yao, Hao Peng, Xiaobao Wu, Jianhui Chen, Muhan Zhang, Liangming Pan
| Challenge: | Recent research has shown that reinforcement learning can elicit intriguing emergent reasoning behaviors. |
| Approach: | They propose a comprehensive survey of the mechanistic understanding of large reasoning models . they organize findings into three core dimensions: 1) training dynamics, 2) reasoning mechanisms, and 3) unintended behaviors. |
| Outcome: | This paper synthesizes the mechanistic understanding of large reasoning models into three dimensions . authors outline a roadmap for future studies including improved interpretability and methodologies . |
Modeling Discriminative Representations for Out-of-Domain Detection with Supervised Contrastive Learning (2021.acl-short)
Copied to clipboard
| Challenge: | Existing methods of OOD detection only focus on whether a sample is correctly classified . lack of real OOD examples leads to poor prior knowledge about these unknown intents . |
| Approach: | They propose a supervised contrastive learning objective to minimize intra-class variance . they employ an adversarial augmentation mechanism to obtain pseudo diverse views . |
| Outcome: | The proposed method minimizes intra-class variance by pulling together in-domain intents belonging to the same class and maximizes inter-class variation by pushing apart samples from different classes. |
Multi-Persona Thinking for Bias Mitigation in Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models exhibit social biases, which can lead to harmful stereotypes and unfair outcomes. |
| Approach: | They propose a simple inference-time framework that encourages reasoning from multiple perspectives. |
| Outcome: | The proposed framework reduces bias by encouraging reasoning from multiple perspectives. |
ChineseBERT: Chinese Pretraining Enhanced by Glyph and Pinyin Information (2021.acl-long)
Copied to clipboard
| Challenge: | ChineseBERT model incorporates glyph and pinyin information of Chinese characters into pretraining . proposed model achieves new performance boost over baseline models with fewer training steps . |
| Approach: | They propose a ChineseBERT model that incorporates glyph and pinyin information into pretraining . the glyph embedding is obtained based on different fonts of a character, and the pinyink embeddment characterizes the pronunciation of Chinese characters. |
| Outcome: | The proposed model achieves new performance boosts over baseline models with fewer training steps. |
Ultra-Low-Dimensional Prompt Tuning via Random Projection (2026.eacl-long)
Copied to clipboard
| Challenge: | Prompt tuning addresses parameter-efficiency by learning embeddings, but these embeddements are typically tied to the model’s hidden dimensionality, limiting parameter saving. |
| Approach: | They propose a parameter-efficient method that learns prompt embeddings exclusively in the input layer of the model and uses a frozen random matrix for up-projection. |
| Outcome: | The proposed method outperforms previous methods using significantly fewer parameters while maintaining performance. |