Language models (LMs) are a central component of modern AI systems, and diffusion-based language models (DLMs) have recently emerged as a competitive alternative. Both paradigms rely on word embeddings not only to represent the input sentence, but also to represent the target sentence that backbone models are trained to predict. We argue that such static embedding of the target word is insensitive to neighboring words, encouraging locally accurate word prediction while neglecting global structure across the target sentence. To address this limitation, we propose a continuous sentence representation, termed sentence curve, defined as a spline curve whose control points affect multiple words in the sentence. Based on this representation, we introduce sentence curve language model (SCLM), which extends DLMs to predict sentence curves instead of the static word embeddings. We theoretically show that sentence curve prediction induces a regularization effect that promotes global structure modeling, and characterize how different sentence curve types affect this behavior. Empirically, SCLM achieves SOTA performance among DLMs on IWSLT14 and WMT14, shows stable training without burdensome knowledge distillation, and demonstrates promising potential compared to discrete DLMs on LM1B.

It is challenging to understand Literary Sinitic text from the Joseon dynasty, since there is a lack of explicit word separators, which creates significant semantic ambiguity. To address this, both sentence segmentation and named entity recognition (NER) are essential. We propose a Transformer-based analyzer that performs these two tasks simultaneously. Trained on a labeled corpus from the Seungjeongwon Ilgi, our model effectively segments sentences and identifies named entities, thereby significantly improving the understanding of sentence structure and overall context.

This paper proposes a novel reinforcement learning (RL) approach for internet network management (NM) that facilitates the RL agent can handle dynamic preference scenario. Traditional RL-based NM methods typically use fixed preferences to optimize multiple objectives like quality of service (QoS) and computing resource usage. However, in real-world scenarios, preferences of the multiple objectives can be changed dynamically due to factors such as network overloads or server failures. We present a method that allows RL agents to adapt to dynamic preferences during testing. Our experiments show that the proposed method significantly improves generalizability across various network preferences compared to previous RL methods, offering a more efficient and flexible solution for NM.