Conversational assistants are becoming more and more popular, including in healthcare, partly because of the availability and capabilities of Large Language Models. There is a need for controlled, probing evaluations with real stakeholders which can highlight advantages and disadvantages of more traditional architectures and those based on generative AI. We present a within-group user study to compare two versions of a conversational assistant that allows heart failure patients to ask about salt content in food. One version of the system was developed in-house with a neurosymbolic architecture, and one is based on ChatGPT. The evaluation shows that the in-house system is more accurate, completes more tasks and is less verbose than the one based on ChatGPT; on the other hand, the one based on ChatGPT makes fewer speech errors and requires fewer clarifications to complete the task. Patients show no preference for one over the other.

As large language models (LLMs) scale in size and adoption, their computational and environmental costs continue to rise. Prior benchmarking efforts have primarily focused on latency reduction in idealized settings, often overlooking the diverse real-world inference workloads that shape energy use. In this work, we systematically analyze the energy implications of common inference efficiency optimizations across diverse Natural Language Processing (NLP) and generative Artificial Intelligence (AI) workloads, including conversational AI and code generation. We introduce a modeling approach that approximates real-world LLM workflows through a binning strategy for input-output token distributions and batch size variations. Our empirical analysis spans software frameworks, decoding strategies, GPU architectures, online and offline serving settings, and model parallelism configurations. We show that the effectiveness of inference optimizations is highly sensitive to workload geometry, software stack, and hardware accelerators, demonstrating that naive energy estimates based on FLOPs or theoretical GPU utilization significantly underestimate real-world energy consumption. Our findings reveal that the proper application of relevant inference efficiency optimizations can reduce total energy use by up to 73% from unoptimized baselines. These insights provide a foundation for sustainable LLM deployment and inform energy-efficient design strategies for future AI infrastructure.

Large Language Models (LLMs) increasingly require processing long text sequences, but GPU memory limitations force difficult trade-offs between memory capacity and bandwidth. While HBM-based acceleration offers high bandwidth, its capacity remains constrained. Offloading data to host-side DIMMs improves capacity but introduces costly data swapping overhead. We identify that the critical memory bottleneck lies in the decoding phase of multi-head attention (MHA) exclusively, which demands substantial capacity for storing KV caches and high bandwidth for attention computation. Our key insight reveals this operation uniquely aligns with modern DIMM-based processing-in-memory (PIM) architectures, which offers scalability of both capacity and bandwidth. Based on this observation and insight, we propose L3, a hardware-software co-designed system integrating DIMM-PIM and GPU devices. L3 introduces three innovations: First, hardware redesigns resolve data layout mismatches and computational element mismatches in DIMM-PIM, enhancing LLM inference utilization. Second, communication optimization enables hiding the data transfer overhead with the computation. Third, an adaptive scheduler coordinates GPU-DIMM-PIM operations to maximize parallelism between devices. Evaluations using real-world traces show L3 achieves up to 6.1$\times$ speedup over state-of-the-art HBM-PIM solutions while significantly improving batch sizes.

Large language models (LLMs) often generate responses that deviate from user input or training data, a phenomenon known as "hallucination." These hallucinations undermine user trust and hinder the adoption of generative AI systems. Addressing hallucinations is essential for the advancement of LLMs. This paper introduces a comprehensive hallucination benchmark, incorporating both new extrinsic and existing intrinsic evaluation tasks, built upon clear taxonomy of hallucination. A major challenge in benchmarking hallucinations is the lack of a unified framework due to inconsistent definitions and categorizations. We disentangle LLM hallucination from "factuality," proposing a clear taxonomy that distinguishes between extrinsic and intrinsic hallucinations, to promote consistency and facilitate research. Extrinsic hallucinations, where the generated content is not consistent with the training data, are increasingly important as LLMs evolve. Our benchmark includes dynamic test set generation to mitigate data leakage and ensure robustness against such leakage. We also analyze existing benchmarks, highlighting their limitations and saturation. The work aims to: (1) establish a clear taxonomy of hallucinations, (2) introduce new extrinsic hallucination tasks, with data that can be dynamically regenerated to prevent saturation by leakage, (3) provide a comprehensive analysis of existing benchmarks, distinguishing them from factuality evaluations.

As generative AI models become increasingly integrated into high-stakes domains, the need for robust methods to evaluate their ethical reasoning becomes increasingly important. This paper introduces a five-dimensional audit model -- assessing Analytic Quality, Breadth of Ethical Considerations, Depth of Explanation, Consistency, and Decisiveness -- to evaluate the ethical logic of leading large language models (LLMs). Drawing on traditions from applied ethics and higher-order thinking, we present a multi-battery prompt approach, including novel ethical dilemmas, to probe the models' reasoning across diverse contexts. We benchmark seven major LLMs finding that while models generally converge on ethical decisions, they vary in explanatory rigor and moral prioritization. Chain-of-Thought prompting and reasoning-optimized models significantly enhance performance on our audit metrics. This study introduces a scalable methodology for ethical benchmarking of AI systems and highlights the potential for AI to complement human moral reasoning in complex decision-making contexts.

The significant computational demands of pretrained language models (PLMs), which often require dedicated hardware, present a substantial challenge in serving them efficiently, especially in multi-tenant environments. To address this, we introduce HMI, a Hierarchical knowledge management-based Multi-tenant Inference system, designed to manage tenants with distinct PLMs resource-efficiently. Our approach is three-fold: Firstly, we categorize PLM knowledge into general, domain-specific, and task-specific. Leveraging insights on knowledge acquisition across different model layers, we construct hierarchical PLMs (hPLMs) by extracting and storing knowledge at different levels, significantly reducing GPU memory usage per tenant. Secondly, we establish hierarchical knowledge management for hPLMs generated by various tenants in HMI. We manage domain-specific knowledge with acceptable storage increases by constructing and updating domain-specific knowledge trees based on frequency. We manage task-specific knowledge within limited GPU memory through parameter swapping. Finally, we propose system optimizations to enhance resource utilization and inference throughput. These include fine-grained pipelining via hierarchical knowledge prefetching to overlap CPU and I/O operations with GPU computations, and optimizing parallel implementations with batched matrix multiplications. Our experimental results demonstrate that the proposed HMI can efficiently serve up to 10,000 hPLMs (hBERTs and hGPTs) on a single GPU, with only a negligible compromise in accuracy.

There has been impressive progress in Large Multimodal Models (LMMs). Recent works extend these models to long inputs, including multi-page documents and long videos. However, the model size and performance of these long context models are still limited due to the computational cost in both training and inference. In this work, we explore an orthogonal direction and process long inputs without long context LMMs. We propose Frame Selection Augmented Generation (FRAG), where the model first selects relevant frames within the input, and then only generates the final outputs based on the selected frames. The core of the selection process is done by scoring each frame independently, which does not require long context processing. The frames with the highest scores are then selected by a simple Top-K selection. We show that this frustratingly simple framework is applicable to both long videos and multi-page documents using existing LMMs without any fine-tuning. We consider two models, LLaVA-OneVision and InternVL2, in our experiments and show that FRAG consistently improves the performance and achieves state-of-the-art performances for both long video and long document understanding. For videos, FRAG substantially improves InternVL2-76B by 5.8% on MLVU and 3.7% on Video-MME. For documents, FRAG achieves over 20% improvements on MP-DocVQA compared with recent LMMs specialized in long document understanding. Code is available at: https://github.com/NVlabs/FRAG

Unsupervised machine learning techniques, such as topic modeling and clustering, are often used to identify latent patterns in unstructured text data in fields such as political science and sociology. These methods overcome common concerns about reproducibility and costliness involved in the labor-intensive process of human qualitative analysis. However, two major limitations of topic models are their interpretability and their practicality for answering targeted, domain-specific social science research questions. In this work, we investigate opportunities for using LLM-generated text augmentation to improve the usefulness of topic modeling output. We use a political science case study to evaluate our results in a domain-specific application, and find that topic modeling using GPT-4 augmentations creates highly interpretable categories that can be used to investigate domain-specific research questions with minimal human guidance.

Understanding source code is a topic of great interest in the software engineering community, since it can help programmers in various tasks such as software maintenance and reuse. Recent advances in large language models (LLMs) have demonstrated remarkable program comprehension capabilities, while transformer-based topic modeling techniques offer effective ways to extract semantic information from text. This paper proposes and explores a novel approach that combines these strengths to automatically identify meaningful topics in a corpus of Python programs. Our method consists in applying topic modeling on the descriptions obtained by asking an LLM to summarize the code. To assess the internal consistency of the extracted topics, we compare them against topics inferred from function names alone, and those derived from existing docstrings. Experimental results suggest that leveraging LLM-generated summaries provides interpretable and semantically rich representation of code structure. The promising results suggest that our approach can be fruitfully applied in various software engineering tasks such as automatic documentation and tagging, code search, software reorganization and knowledge discovery in large repositories.

Large language models (LLMs) have demonstrated remarkable capabilities, but they require vast amounts of data and computational resources. In contrast, smaller models (SMs), while less powerful, can be more efficient and tailored to specific domains. In this position paper, we argue that taking a collaborative approach, where large and small models work synergistically, can accelerate the adaptation of LLMs to private domains and unlock new potential in AI. We explore various strategies for model collaboration and identify potential challenges and opportunities. Building upon this, we advocate for industry-driven research that prioritizes multi-objective benchmarks on real-world private datasets and applications.