Haipeng Zhang

Kai Peng, Ying Zhang, Shuai Ling et al.

Celebrities' whereabouts are of pervasive importance. For instance, where politicians go, how often they visit, and who they meet, come with profound geopolitical and economic implications. Although news articles contain travel information of celebrities, it is not possible to perform large-scale and network-wise analysis due to the lack of automatic itinerary detection tools. To design such tools, we have to overcome difficulties from the heterogeneity among news articles: 1)One single article can be noisy, with irrelevant people and locations, especially when the articles are long. 2)Though it may be helpful if we consider multiple articles together to determine a particular trip, the key semantics are still scattered across different articles intertwined with various noises, making it hard to aggregate them effectively. 3)Over 20% of the articles refer to the celebrities' trips indirectly, instead of using the exact celebrity names or location names, leading to large portions of trips escaping regular detecting algorithms. We model text content across articles related to each candidate location as a graph to better associate essential information and cancel out the noises. Besides, we design a special pooling layer based on attention mechanism and node similarity, reducing irrelevant information from longer articles. To make up the missing information resulted from indirect mentions, we construct knowledge sub-graphs for named entities (person, organization, facility, etc.). Specifically, we dynamically update embeddings of event entities like the G7 summit from news descriptions since the properties (date and location) of the event change each time, which is not captured by the pre-trained event representations. The proposed CeleTrip jointly trains these modules, which outperforms all baseline models and achieves 82.53% in the F1 metric.

Xiaocong Du, Haoyu Pei, Haipeng Zhang

Classical Chinese poetry is a vital and enduring part of Chinese literature, conveying profound emotional resonance. Existing studies analyze sentiment based on textual meanings, overlooking the unique rhythmic and visual features inherent in poetry,especially since it is often recited and accompanied by Chinese paintings. In this work, we propose a dialect-enhanced multimodal framework for classical Chinese poetry sentiment analysis. We extract sentence-level audio features from the poetry and incorporate audio from multiple dialects,which may retain regional ancient Chinese phonetic features, enriching the phonetic representation. Additionally, we generate sentence-level visual features, and the multimodal features are fused with textual features enhanced by LLM translation through multimodal contrastive representation learning. Our framework outperforms state-of-the-art methods on two public datasets, achieving at least 2.51% improvement in accuracy and 1.63% in macro F1. We open-source the code to facilitate research in this area and provide insights for general multimodal Chinese representation.

Zibo Zhao, Yuanting Zha, Haipeng Zhang et al.

Self-reflection capabilities emerge in Large Language Models after RL post-training, with multi-turn RL achieving substantial gains over SFT counterparts. Yet the mechanism of how a unified optimization objective gives rise to functionally distinct capabilities of generating solutions and evaluating when to revise them remains opaque. To address this question, we introduce the Gradient Attribution Property to characterize how reward gradients distribute across policy components, formalized through the Two-Stage Decision-Sampling (DS) Hypothesis, which decomposes the policy into sampling ($π_{sample}$) for generation and decision ($π_{d}$) for verification. We prove that surrogate rewards exhibit Balanced Gradient Attribution, while SFT and KL penalties exhibit Unbalanced Gradient Attribution, with length-weighting creating asymmetric regularization that constrains $π_{sample}$ while leaving $π_{d}$ under-optimized, providing an theoretical explanation of why RL succeeds where SFT fails. We also empirically validate our theoretical predictions on arithmetic reasoning demonstrates that RL's superior generalization stems primarily from improved decision-making ($π_{d}$) rather than sampling capabilities, providing a first-principles mechanistic explanation for self-correction in thinking models.