Large Language Models (LLMs) have become foundational in modern language-driven applications, profoundly influencing daily life. A critical technique in leveraging their potential is role-playing, where LLMs simulate diverse roles to enhance their real-world utility. However, while research has highlighted the presence of social biases in LLM outputs, it remains unclear whether and to what extent these biases emerge during role-playing scenarios. In this paper, we introduce BiasLens, a fairness testing framework designed to systematically expose biases in LLMs during role-playing. Our approach uses LLMs to generate 550 social roles across a comprehensive set of 11 demographic attributes, producing 33,000 role-specific questions targeting various forms of bias. These questions, spanning Yes/No, multiple-choice, and open-ended formats, are designed to prompt LLMs to adopt specific roles and respond accordingly. We employ a combination of rule-based and LLM-based strategies to identify biased responses, rigorously validated through human evaluation. Using the generated questions as the benchmark, we conduct extensive evaluations of six advanced LLMs released by OpenAI, Mistral AI, Meta, Alibaba, and DeepSeek. Our benchmark reveals 72,716 biased responses across the studied LLMs, with individual models yielding between 7,754 and 16,963 biased responses, underscoring the prevalence of bias in role-playing contexts. To support future research, we have publicly released the benchmark, along with all scripts and experimental results.

On-device large language models (LLMs) are catalyzing novel mobile applications such as UI task automation and personalized email auto-reply, without giving away users' private data. However, on-device LLMs still suffer from unacceptably long inference latency, especially the time to first token (prefill stage) due to the need of long context for accurate, personalized content generation, as well as the lack of parallel computing capacity of mobile CPU/GPU. To enable practical on-device LLM, we present mllm-NPU, the first-of-its-kind LLM inference system that efficiently leverages on-device Neural Processing Unit (NPU) offloading. Essentially, mllm-NPU is an algorithm-system co-design that tackles a few semantic gaps between the LLM architecture and contemporary NPU design. Specifically, it re-constructs the prompt and model in three levels: (1) At prompt level, it divides variable-length prompts into multiple fixed-sized chunks while maintaining data dependencies; (2) At tensor level, it identifies and extracts significant outliers to run on the CPU/GPU in parallel with minimal overhead; (3) At block level, it schedules Transformer blocks in an out-of-order manner to the CPU/GPU and NPU based on their hardware affinity and sensitivity to accuracy. Compared to competitive baselines, mllm-NPU achieves 22.4x faster prefill speed and 30.7x energy savings on average, and up to 32.8x speedup in an end-to-end real-world application. For the first time, mllm-NPU achieves more than 1,000 tokens/sec prefilling for a billion-sized model (Qwen1.5-1.8B), paving the way towards practical on-device LLM.

Retrieval-Augmented Generation (RAG) has shown significant improvements in various natural language processing tasks by integrating the strengths of large language models (LLMs) and external knowledge databases. However, RAG introduces long sequence generation and leads to high computation and memory costs. We propose RAGCache, a novel multilevel dynamic caching system tailored for RAG. Our analysis benchmarks current RAG systems, pinpointing the performance bottleneck (i.e., long sequence due to knowledge injection) and optimization opportunities (i.e., caching knowledge's intermediate states). Based on these insights, we design RAGCache, which organizes the intermediate states of retrieved knowledge in a knowledge tree and caches them in the GPU and host memory hierarchy. RAGCache proposes a replacement policy that is aware of LLM inference characteristics and RAG retrieval patterns. It also dynamically overlaps the retrieval and inference steps to minimize the end-to-end latency. We implement RAGCache and evaluate it on vLLM, a state-of-the-art LLM inference system and Faiss, a state-of-the-art vector database. The experimental results show that RAGCache reduces the time to first token (TTFT) by up to 4x and improves the throughput by up to 2.1x compared to vLLM integrated with Faiss.

Deep Learning (DL) is increasingly being integrated into Web applications through a method known as "in-browser inference", where the DL processes occur directly within Web browsers. However, the actual performance of this method and its effect on user experience quality (QoE) is not well-understood. This gap in knowledge necessitates new forms of QoE measurement, going beyond traditional metrics such as page load time. To address this, we conducted the first extensive performance evaluation of in-browser inference. We introduced new metrics for this purpose: responsiveness, smoothness, and inference accuracy. Our thorough study included 9 widely-used DL models and tested them across 50 popular PC Web browsers. The findings show a significant latency issue with in-browser inference: it's on average 16.9 times slower on CPU and 4.9 times slower on GPU than native inference methods. Several factors contribute to this latency, including underused hardware instruction sets, inherent delays in the runtime environment, resource competition within the browser, and inefficiencies in software libraries and GPU abstractions. Moreover, in-browser inference demands a lot of memory, sometimes up to 334.6 times more than the size of the DL models themselves. This excessive memory usage is partly due to suboptimal memory management. Additionally, we noticed that in-browser inference increases the time it takes for graphical user interface (GUI) components to load in web browsers by a significant 67.2\%, which severely impacts the overall QoE for users of web applications that depend on this technology.

In the rapidly evolving field of artificial intelligence (AI), a paradigm shift is underway. We are witnessing the transition from specialized, fragmented deep learning models to versatile, one-size-fits-all foundation models. These advanced AI systems are capable of operating in an open-world context, interacting with open vocabularies and image pixels for unseen AI tasks, i.e., zero-shot abilities. They are exemplified by (1) Large Language Models (LLMs) such as GPTs [39] that can ingest almost every NLP task in the form as a prompt; (2) Vision Transformers Models (ViTs) such as Masked Autoencoder [133] that can handle various downstream vision tasks; (3) Latent Diffusion Models (LDMs) such as Stable Diffusion [310] that generate high-quality images with arbitrary text-based prompts; (4) Multimodal models such as CLIP [296] and ImageBind [116] that map different modal data into the same latent space and are widely used as backbone for cross-modality tasks like image retrieval/search and visual-question answering. Such flexibility and generality marks a significant departure from the earlier era of AI, setting a new standard for how AI interfaces with the world. The success of these foundation models is deeply rooted in their scalability: unlike their predecessors, these models' accuracy and generalization ability can continuously expand with more data or parameters, without altering the underlying simple algorithms and architectures.

Generative tasks, such as text generation and question answering, hold a crucial position in the realm of mobile applications. Due to their sensitivity to privacy concerns, there is a growing demand for their execution directly on mobile devices. Currently, the execution of these generative tasks heavily depends on Large Language Models (LLMs). Nevertheless, the limited memory capacity of these devices presents a formidable challenge to the scalability of such models. In our research, we introduce LLMCad, an innovative on-device inference engine specifically designed for efficient generative Natural Language Processing (NLP) tasks. The core idea behind LLMCad revolves around model collaboration: a compact LLM, residing in memory, takes charge of generating the most straightforward tokens, while a high-precision LLM steps in to validate these tokens and rectify any identified errors. LLMCad incorporates three novel techniques: (1) Instead of generating candidate tokens in a sequential manner, LLMCad employs the smaller LLM to construct a token tree, encompassing a wider range of plausible token pathways. Subsequently, the larger LLM can efficiently validate all of these pathways simultaneously. (2) It employs a self-adjusting fallback strategy, swiftly initiating the verification process whenever the smaller LLM generates an erroneous token. (3) To ensure a continuous flow of token generation, LLMCad speculatively generates tokens during the verification process by implementing a compute-IO pipeline. Through an extensive series of experiments, LLMCad showcases an impressive token generation speed, achieving rates up to 9.3x faster than existing inference engines.

This paper conducts fairness testing on automated pedestrian detection, a crucial but under-explored issue in autonomous driving systems. We evaluate eight widely-studied pedestrian detectors across demographic groups on large-scale real-world datasets. To enable thorough fairness testing, we provide extensive annotations for the datasets, resulting in 8,311 images with 16,070 gender labels, 20,115 age labels, and 3,513 skin tone labels. Our findings reveal significant fairness issues related to age and skin tone. The detection accuracy for adults is 19.67% higher compared to children, and there is a 7.52% accuracy disparity between light-skin and dark-skin individuals. Gender, however, shows only a 1.1% difference in detection accuracy. Additionally, we investigate common scenarios explored in the literature on autonomous driving testing, and find that the bias towards dark-skin pedestrians increases significantly under scenarios of low contrast and low brightness. We publicly release the code, data, and results to support future research on fairness in autonomous driving.

Large language models (LLMs) power a new generation of interactive AI applications exemplified by ChatGPT. The interactive nature of these applications demand low job completion time (JCT) for model inference. Existing LLM serving systems use run-to-completion processing for inference jobs, which suffers from head-of-line blocking and long JCT. We present FastServe, a distributed inference serving system for LLMs. FastServe exploits the autoregressive pattern of LLM inference to enable preemption at the granularity of each output token. FastServe uses preemptive scheduling to minimize JCT with a novel skip-join Multi-Level Feedback Queue scheduler. Based on the new semi information-agnostic setting of LLM inference, the scheduler leverages the input length information to assign an appropriate initial queue for each arrival job to join. The higher priority queues than the joined queue are skipped to reduce demotions. We design an efficient GPU memory management mechanism that proactively offloads and uploads intermediate states between GPU memory and host memory for LLM inference. We build a system prototype of FastServe based on NVIDIA FasterTransformer. Experimental results show that compared to the state-of-the-art solution Orca, FastServe improves the average and tail JCT by up to 5.1$\times$ and 6.4$\times$, respectively.

Deep Learning (DL) is finding its way into a growing number of mobile software applications. These software applications, named as DL based mobile applications (abbreviated as mobile DL apps) integrate DL models trained using large-scale data with DL programs. A DL program encodes the structure of a desirable DL model and the process by which the model is trained using training data. Due to the increasing dependency of current mobile apps on DL, software engineering (SE) for mobile DL apps has become important. However, existing efforts in SE research community mainly focus on the development of DL models and extensively analyze faults in DL programs. In contrast, faults related to the deployment of DL models on mobile devices (named as deployment faults of mobile DL apps) have not been well studied. Since mobile DL apps have been used by billions of end users daily for various purposes including for safety-critical scenarios, characterizing their deployment faults is of enormous importance. To fill the knowledge gap, this paper presents the first comprehensive study on the deployment faults of mobile DL apps. We identify 304 real deployment faults from Stack Overflow and GitHub, two commonly used data sources for studying software faults. Based on the identified faults, we construct a fine-granularity taxonomy consisting of 23 categories regarding to fault symptoms and distill common fix strategies for different fault types. Furthermore, we suggest actionable implications and research avenues that could further facilitate the deployment of DL models on mobile devices.

The Spatio-Temporal Crowd Flow Prediction (STCFP) problem is a classical problem with plenty of prior research efforts that benefit from traditional statistical learning and recent deep learning approaches. While STCFP can refer to many real-world problems, most existing studies focus on quite specific applications, such as the prediction of taxi demand, ridesharing order, and so on. This hinders the STCFP research as the approaches designed for different applications are hardly comparable, and thus how an applicationdriven approach can be generalized to other scenarios is unclear. To fill in this gap, this paper makes two efforts: (i) we propose an analytic framework, called STAnalytic, to qualitatively investigate STCFP approaches regarding their design considerations on various spatial and temporal factors, aiming to make different application-driven approaches comparable; (ii) we construct an extensively large-scale STCFP benchmark datasets with four different scenarios (including ridesharing, bikesharing, metro, and electrical vehicle charging) with up to hundreds of millions of flow records, to quantitatively measure the generalizability of STCFP approaches. Furthermore, to elaborate the effectiveness of STAnalytic in helping design generalizable STCFP approaches, we propose a spatio-temporal meta-model, called STMeta, by integrating generalizable temporal and spatial knowledge identified by STAnalytic. We implement three variants of STMeta with different deep learning techniques. With the datasets, we demonstrate that STMeta variants can outperform state-of-the-art STCFP approaches by 5%.