Elliott Wen

Shamane Siriwardhana, Rivindu Weerasekera, Elliott Wen et al.

Retrieval Augment Generation (RAG) is a recent advancement in Open-Domain Question Answering (ODQA). RAG has only been trained and explored with a Wikipedia-based external knowledge base and is not optimized for use in other specialized domains such as healthcare and news. In this paper, we evaluate the impact of joint training of the retriever and generator components of RAG for the task of domain adaptation in ODQA. We propose \textit{RAG-end2end}, an extension to RAG, that can adapt to a domain-specific knowledge base by updating all components of the external knowledge base during training. In addition, we introduce an auxiliary training signal to inject more domain-specific knowledge. This auxiliary signal forces \textit{RAG-end2end} to reconstruct a given sentence by accessing the relevant information from the external knowledge base. Our novel contribution is unlike RAG, RAG-end2end does joint training of the retriever and generator for the end QA task and domain adaptation. We evaluate our approach with datasets from three domains: COVID-19, News, and Conversations, and achieve significant performance improvements compared to the original RAG model. Our work has been open-sourced through the Huggingface Transformers library, attesting to our work's credibility and technical consistency.

Elliott Wen, Chitralekha Gupta, Prasanth Sasikumar et al.

Researchers have used machine learning approaches to identify motion sickness in VR experience. These approaches demand an accurately-labeled, real-world, and diverse dataset for high accuracy and generalizability. As a starting point to address this need, we introduce `VR.net', a dataset offering approximately 12-hour gameplay videos from ten real-world games in 10 diverse genres. For each video frame, a rich set of motion sickness-related labels, such as camera/object movement, depth field, and motion flow, are accurately assigned. Building such a dataset is challenging since manual labeling would require an infeasible amount of time. Instead, we utilize a tool to automatically and precisely extract ground truth data from 3D engines' rendering pipelines without accessing VR games' source code. We illustrate the utility of VR.net through several applications, such as risk factor detection and sickness level prediction. We continuously expand VR.net and envision its next version offering 10X more data than the current form. We believe that the scale, accuracy, and diversity of VR.net can offer unparalleled opportunities for VR motion sickness research and beyond.

Zushuai Zhang, Elliott Wen, Ewan Tempero

Background: Large Language Models (LLMs) are increasingly used for code generation. However, their ability to generate multi-class projects that require object-oriented design (OOD) remains unclear, especially relative to projects developed with human involvement. Aims: The primary objective of this study is to compare OOD quality in projects from three authorship conditions: PreAI (human-involved projects produced before widespread LLM use), PostAI (human-involved projects produced after widespread LLM use), and PureAI (projects generated end-to-end by contemporary LLMs). Method: We conducted a comparative case study on a postgraduate Java assignment. Two offerings of the same assignment were selected as the PreAI and PostAI datasets. PureAI projects were generated using three contemporary LLMs. We analyzed OOD quality using project-level OOD metrics, code smell density, and domain modeling. Results: Relative to human-involved projects, PureAI projects show lower code smell density and generally appear simpler in terms of total size, complexity, and coupling. However, this is consistent with oversimplification, as it is associated with missing abstractions and weaker responsibility separation. PostAI is closer to PureAI than PreAI on many OOD measures and also shows tendencies toward oversimplification. Conclusions: Our findings indicate that appropriate human guidance on object-oriented decomposition and responsibility assignment remains important when LLMs are used for object-oriented design.

Letian Cheng, Junyan Wang, Yan Gao et al.

Perplexity is a widely adopted metric for assessing the predictive quality of large language models (LLMs) and often serves as a reference metric for downstream evaluations. However, recent evidence shows that perplexity can be unreliable, especially when irrelevant long inputs are used, raising concerns for both benchmarking and system deployment. While prior efforts have employed selective input filtering and curated datasets, the impact of input length on perplexity has not been systematically studied from a systems perspective and input length has rarely been treated as a first-class system variable affecting both fairness and efficiency. In this work, we close this gap by introducing LengthBenchmark, a system-conscious evaluation framework that explicitly integrates input length, evaluation protocol design, and system-level costs, evaluating representative LLMs under two scoring protocols (direct accumulation and fixed window sliding) across varying context lengths. Unlike prior work that focuses solely on accuracy-oriented metrics, LengthBenchmark additionally measures latency, memory footprint, and evaluation cost, thereby linking predictive metrics to deployment realities. We further incorporate quantized variants not as a main contribution, but as robustness checks, showing that length-induced biases persist across both full-precision and compressed models. This design disentangles the effects of evaluation logic, quantization, and input length, and demonstrates that length bias is a general phenomenon that undermines fair cross-model comparison. Our analysis yields two key observations: (i) sliding window evaluation consistently inflates performance on short inputs, and (ii) both full-precision and quantized models appear to realise gains as the evaluated segment length grows.

Shamane Siriwardhana, Rivindu Weerasekera, Elliott Wen et al.

In this paper, we illustrate how to fine-tune the entire Retrieval Augment Generation (RAG) architecture in an end-to-end manner. We highlighted the main engineering challenges that needed to be addressed to achieve this objective. We also compare how end-to-end RAG architecture outperforms the original RAG architecture for the task of question answering. We have open-sourced our implementation in the HuggingFace Transformers library.

Elliott Wen, Sean Ma, Ewan Tempero et al.

While NVIDIA remains the dominant provider of AI accelerators within cloud data center, emerging vendors such as AMD, Intel, Mac, and Huawei offer cost-effective alternatives with claims of compatibility and performance. This paper presents the first empirical study investigating divergence in machine learning model across heterogeneous AI accelerators. Utilizing an automated pipeline, we synthesize over 100,000 variant models derived from 4,000 real-world models and execute them across five different enterprise-grade accelerators. Our findings suggest that newer AI platforms from Mac and Huawei support at least 17\% fewer operators than NVIDIA. These platforms also exhibit a higher rate of output discrepancies (exceeding 5\%), which stem from differences in operator implementations, handling of exceptional numerical values, and instruction scheduling. They are also more susceptible to failures during model compilation-based acceleration, and in some cases, the compiled models produce outputs that differ noticeably from those generated using the standard execution mode. In addition, we identify 7 implementation flaws in PyTorch and 40 platform-specific issues across vendors. These results underscore the challenges of achieving consistent machine learning behavior in an increasingly diverse hardware ecosystem.

Yurun Wu, Yousong Sun, Burkhard Wunsche et al.

Virtual Reality (VR) has rapidly become a mainstream platform for gaming and interactive experiences, yet ensuring the quality, safety, and appropriateness of VR content remains a pressing challenge. Traditional human-based quality assurance is labor-intensive and cannot scale with the industry's rapid growth. While automated testing has been applied to traditional 2D and 3D games, extending it to VR introduces unique difficulties due to high-dimensional sensory inputs and strict real-time performance requirements. We present VRScout, a deep learning-based agent capable of autonomously navigating VR environments and interacting with virtual objects in a human-like and real-time manner. VRScout learns from human demonstrations using an enhanced Action Chunking Transformer that predicts multi-step action sequences. This enables our agent to capture higher-level strategies and generalize across diverse environments. To balance responsiveness and precision, we introduce a dynamically adjustable sliding horizon that adapts the agent's temporal context at runtime. We evaluate VRScout on commercial VR titles and show that it achieves expert-level performance with only limited training data, while maintaining real-time inference at 60 FPS on consumer-grade hardware. These results position VRScout as a practical and scalable framework for automated VR game testing, with direct applications in both quality assurance and safety auditing.

Victoria Huang, Gang Chen, Qiang Fu et al.

Controller placement problem (CPP) is a key issue for Software-Defined Networking (SDN) with distributed controller architectures. This problem aims to determine a suitable number of controllers deployed in important locations so as to optimize the overall network performance. In comparison to communication delay, existing literature on the CPP assumes that the influence of controller workload distribution on network performance is negligible. In this paper, we tackle the CPP that simultaneously considers the communication delay, the control plane utilization, and the controller workload distribution. Due to this reason, our CPP is intrinsically different from and clearly more difficult than any previously studied CPPs that are NP-hard. To tackle this challenging issue, we develop a new algorithm that seamlessly integrates the genetic algorithm (GA) and the gradient descent (GD) optimization method. Particularly, GA is used to search for suitable CPP solutions. The quality of each solution is further evaluated through GD. Simulation results on two representative network scenarios (small-scale and large-scale) show that our algorithm can effectively strike the trade-off between the control plane utilization and the network response time.