As large language models (LLMs) evolve from conversational assistants into agents capable of handling complex tasks, they are increasingly deployed in high-risk domains. However, existing benchmarks largely rely on mixed queries and posterior evaluation, output-level scoring, which quantifies hallucination severity but offers limited insight into where and why hallucinations arise in the generation pipeline. We therefore reformulate hallucination evaluation as a diagnostic problem and propose PRISM, a controlled benchmark that disentangles hallucinations into four dimensions: knowledge missing, knowledge errors, reasoning errors, and instruction-following errors, grounded in three stages of generation (memory, instruction, and reasoning). PRISM contains 9,448 instances across 65 tasks and supports fine-grained, stage-aware diagnostic evaluation. Evaluating 24 mainstream open-source and proprietary LLMs, we uncover consistent trade-offs across instruction following, memory retrieval, and logical reasoning, showing that mitigation strategies often improve specific dimensions at the expense of others.We hope PRISM provides a framework for understanding the specific mechanisms behind LLMs hallucinations, ultimately accelerating the development of trustworthy large language models.

Large Language Models (LLMs) inherently use autoregressive decoding, which lacks parallelism in inference and results in significantly slow inference speed. While methods such as Medusa constructs parallelized heads, they lack adequate information interaction across different prediction positions. To overcome this limitation, we introduce Amphista, an enhanced speculative decoding framework that builds upon Medusa. Specifically, Amphista models an *Auto-embedding Block* capable of parallel inference, incorporating bi-directional attention to enable interaction between different drafting heads. Additionally, Amphista integrates *Staged Adaptation Layers*, which ensure a seamless transition of semantic information from the target model’s autoregressive inference to the drafting heads’ non-autoregressive inference, effectively achieving paradigm shift and feature fusion. Experimental results on Vicuna models using MT-Bench and Spec-Bench demonstrate that Amphista achieves substantial acceleration while maintaining generation quality. On MT-Bench, Amphista delivers up to **2.75×** speedup over vanilla autoregressive decoding and **1.40×** over Medusa on Vicuna 33B in wall-clock time.

Word-level information is important in natural language processing (NLP), especially for the Chinese language due to its high linguistic complexity. Chinese word segmentation (CWS) is an essential task for Chinese downstream NLP tasks. Existing methods have already achieved a competitive performance for CWS on large-scale annotated corpora. However, the accuracy of the method will drop dramatically when it handles an unsegmented text with lots of out-of-vocabulary (OOV) words. In addition, there are many different segmentation criteria for addressing different requirements of downstream NLP tasks. Excessive amounts of models with saving different criteria will generate the explosive growth of the total parameters. To this end, we propose a joint multiple criteria model that shares all parameters to integrate different segmentation criteria into one model. Besides, we utilize a transfer learning method to improve the performance of OOV words. Our proposed method is evaluated by designing comprehensive experiments on multiple benchmark datasets (e.g., Bakeoff 2005, Bakeoff 2008 and SIGHAN 2010). Our method achieves the state-of-the-art performances on all datasets. Importantly, our method also shows a competitive practicability and generalization ability for the CWS task.

In medical domain, given a medical question, it is difficult to manually select the most relevant information from a large number of search results. BioNLP 2019 proposes Question Answering (QA) task, which encourages the use of text mining technology to automatically judge whether a search result is an answer to the medical question. The main challenge of QA task is how to mine the semantic relation between question and answer. We propose BioBERT Transformer model to tackle this challenge, which applies Transformers to extract semantic relation between different words in questions and answers. Furthermore, BioBERT is utilized to encode medical domain-specific contextualized word representations. Our method has reached the accuracy of 76.24% and spearman of 17.12% on the BioNLP 2019 QA task.