Quantization is widely adopted to accelerate inference and reduce memory consumption in large language models (LLMs). While activation-weight joint quantization enables efficient low-precision decoding, it suffers substantial performance degradation on multi-step reasoning tasks. We propose QSPEC, a novel quantization paradigm that decouples efficiency from quality by integrating two complementary schemes via speculative decoding: low-precision joint quantization for fast drafting and high-precision weight-only quantization for accurate verification. QSPEC reuses both weights and KV cache across stages, enabling near-zero-cost switching without retraining or auxiliary models. Compared to high-precision baselines, QSPEC achieves up to 1.64x speedup without quality degradation, and outperforms state-of-the-art speculative decoding methods by up to 1.55x in batched settings. Furthermore, QSPEC supports plug-and-play deployment and generalizes well across model scales, quantization methods, and workloads. These properties make QSPEC a practical and scalable solution for high-fidelity quantized LLM serving under memory-constrained scenarios.

Large vision-language models (LVLMs) have witnessed significant progress on visual understanding tasks. However, they often prioritize language knowledge over image information on visual reasoning tasks, incurring performance degradation. To tackle this issue, we first identify the drawbacks of existing solutions (i.e., limited multi-modal reasoning capacities, and insufficient and irrelevant visual descriptions). We then decompose visual reasoning process into two stages: proactive visual perception (i.e., eyesight) and textual reasoning (i.e., wisdom), and introduce a novel visual reasoning framework named ProReason. This framework features decoupled vision-reasoning capabilities and multi-run proactive perception. Briefly, given a multi-modal question, ProReason iterates proactive information collection and reasoning until the answer can be concluded with necessary and sufficient visual descriptions. Notably, the disassociation of capabilities allows seamless integration of existing large language models (LLMs) to compensate for the reasoning deficits of LVLMs. Our extensive experiments demonstrate that ProReason outperforms existing multi-step reasoning frameworks on various benchmarks for both open-source and closed-source models, with the average performance gain reaching 13.2%. Besides, the integration of LLMs allows ProReason to produce high-quality visual reasoning data, which empowers ProReason-distilled models (i.e., ProReason-VL and ProReason-Q3) to achieve superior performance in downstream tasks. Our insights into existing solutions and the decoupled perspective for feasible integration of LLMs illuminate future research on visual reasoning techniques, especially LLM-assisted ones. The code is available at https://github.com/lian-tian-mo-zun/Pro_Reason.

The rapid evolution of large language models (LLMs) has transformed the competitive landscape in natural language processing (NLP), particularly for English and other data-rich languages. However, underrepresented languages like Cantonese, spoken by over 85 million people, face significant development gaps, which is particularly concerning given the economic significance of the Guangdong-Hong Kong-Macau Greater Bay Area, and in substantial Cantonese-speaking populations in places like Singapore and North America. Despite its wide use, Cantonese has scant representation in NLP research, especially compared to other languages from similarly developed regions. To bridge these gaps, we outline current Cantonese NLP methods and introduce new benchmarks designed to evaluate LLM performance in factual generation, mathematical logic, complex reasoning, and general knowledge in Cantonese, which aim to advance open-source Cantonese LLM technology. We also propose future research directions and recommended models to enhance Cantonese LLM development.

With the rapid scaling of large language models (LLMs), serving numerouslow-rank adaptations (LoRAs) concurrently has become increasingly impractical,leading to unaffordable costs and necessitating more parameter-efficientfinetuning methods. In this work, we introduce Partially Rotation-enhanced Low-Rank Adaptation (PRoLoRA), an intra-layer sharing mechanism comprising fouressential components: broadcast reduction, rotation enhancement,partially-sharing refinement, and rectified initialization strategy. As asuperset of LoRA, PRoLoRA retains its advantages, and effectively circumventthe drawbacks of peer parameter-sharing methods with superior model capacity,practical feasibility, and broad applicability. Empirical experimentsdemonstrate the remarkably higher parameter efficiency of PRoLoRA in bothspecific parameter budget and performance target scenarios, and its scalabilityto larger LLMs. Notably, with one time less trainable parameters, PRoLoRA stilloutperforms LoRA on multiple instruction tuning datasets. Subsequently, anablation study is conducted to validate the necessity of individual componentsand highlight the superiority of PRoLoRA over three potential variants.Hopefully, the conspicuously higher parameter efficiency can establish PRoLoRAas a resource-friendly alternative to LoRA.

With the remarkable capabilities, large language models (LLMs) have emergedas essential elements in numerous NLP applications, while parameter-efficientfinetuning, especially LoRA, has gained popularity as a lightweight approachfor model customization. Meanwhile, various dropout methods, initially designedfor full finetuning with all the parameters updated, alleviates overfittingassociated with excessive parameter redundancy. Hence, a possible contradictionarises from negligible trainable parameters of LoRA and the effectiveness ofprevious dropout methods, which has been largely overlooked. To fill this gap,we first confirm that parameter-efficient LoRA is also overfitting-prone. Wethen revisit transformer-specific dropout methods, and establish theirequivalence and distinctions mathematically and empirically. Building upon thiscomparative analysis, we introduce a unified framework for a comprehensiveinvestigation, which instantiates these methods based on dropping position,structural pattern and compensation measure. Through this framework, we revealthe new preferences and performance comparisons of them when involved withlimited trainable parameters. This framework also allows us to amalgamate themost favorable aspects into a novel dropout method named HiddenKey. Extensiveexperiments verify the remarkable superiority and sufficiency of HiddenKeyacross multiple models and tasks, which highlights it as the preferred approachfor high-performance and parameter-efficient finetuning of LLMs.

When communicating with elders with cognitive impairment, cognitive stimulation (CS) help to maintain the cognitive health of elders. Data sparsity is the main challenge in building CS-based dialogue systems, particularly in the Chinese language. To fill this gap, we construct a Chinese CS conversation (CSConv) dataset, which contains about 2.6K groups of dialogues with therapy principles and emotional support strategy labels. Making chit chat while providing emotional support is overlooked by the majority of existing cognitive dialogue systems. In this paper, we propose a multi-source knowledge fusion method for CS dialogue (CSD), to generate open-ended responses guided by the therapy principle and emotional support strategy. We first use a progressive mask method based on external knowledge to learn encoders as effective classifiers, which is the prerequisite to predict the therapy principle and emotional support strategy of the target response. Then a decoder interacts with the perceived therapy principle and emotional support strategy to generate responses. Extensive experiments conducted on the CSConv dataset demonstrate the effectiveness of the proposed method, while there is still a large space for improvement compared to human performance.

Entities can be found in various text genres, ranging from tweets and web pages to user queries submitted to web search engines. Existing research either considers all entities in the text equally important, or heuristics are used to measure their salience. We believe that a key reason for the relatively limited work on entity salience is the lack of appropriate datasets. To support research on entity salience, we present a new dataset, the WikiNews Salience dataset (WN-Salience), which can be used to benchmark tasks such as entity salience detection and salient entity linking. WN-Salience is built on top of Wikinews, a Wikimedia project whose mission is to present reliable news articles. Entities in Wikinews articles are identified by the authors of the articles and are linked to Wikinews categories when they are salient or to Wikipedia pages otherwise. The dataset is built automatically, and consists of approximately 7,000 news articles, and 90,000 in-text entity annotations. We compare the WN-Salience dataset against existing datasets on the task and analyze their differences. Furthermore, we conduct experiments on entity salience detection; the results demonstrate that WN-Salience is a challenging testbed that is complementary to existing ones.