Large Language Models (LLMs) have propelled groundbreaking advancements across several domains and are commonly used for text generation applications. However, the computational demands of these complex models pose significant challenges, requiring efficient hardware acceleration. Benchmarking the performance of LLMs across diverse hardware platforms is crucial to understanding their scalability and throughput characteristics. We introduce LLM-Inference-Bench, a comprehensive benchmarking suite to evaluate the hardware inference performance of LLMs. We thoroughly analyze diverse hardware platforms, including GPUs from Nvidia and AMD and specialized AI accelerators, Intel Habana and SambaNova. Our evaluation includes several LLM inference frameworks and models from LLaMA, Mistral, and Qwen families with 7B and 70B parameters. Our benchmarking results reveal the strengths and limitations of various models, hardware platforms, and inference frameworks. We provide an interactive dashboard to help identify configurations for optimal performance for a given hardware platform.
We introduce DAHL, a benchmark dataset and automated evaluation system designed to assess hallucination in long-form text generation, specifically within the biomedical domain. Our benchmark dataset, meticulously curated from biomedical research papers, consists of 8,573 questions across 29 categories. DAHL evaluates fact-conflicting hallucinations in Large Language Models (LLMs) by deconstructing responses into atomic units, each representing a single piece of information. The accuracy of these responses is averaged to produce the DAHL Score, offering a more in-depth evaluation of hallucinations compared to previous methods that rely on multiple-choice tasks. We conduct experiments with 8 different models, finding that larger models tend to hallucinate less; however, beyond a model size of 7 to 8 billion parameters, further scaling does not significantly improve factual accuracy. The DAHL Score holds potential as an efficient alternative to human-annotated preference labels, being able to be expanded to other specialized domains. We release the dataset and code in public.
Background: Open-Source Pre-Trained Models (PTMs) and datasets provide extensive resources for various Machine Learning (ML) tasks, yet these resources lack a classification tailored to Software Engineering (SE) needs. Aims: We apply an SE-oriented classification to PTMs and datasets on a popular open-source ML repository, Hugging Face (HF), and analyze the evolution of PTMs over time. Method: We conducted a repository mining study. We started with a systematically gathered database of PTMs and datasets from the HF API. Our selection was refined by analyzing model and dataset cards and metadata, such as tags, and confirming SE relevance using Gemini 1.5 Pro. All analyses are replicable, with a publicly accessible replication package. Results: The most common SE task among PTMs and datasets is code generation, with a primary focus on software development and limited attention to software management. Popular PTMs and datasets mainly target software development. Among ML tasks, text generation is the most common in SE PTMs and datasets. There has been a marked increase in PTMs for SE since 2023 Q2. Conclusions: This study underscores the need for broader task coverage to enhance the integration of ML within SE practices.
Specifying all desirable properties of a language model is challenging, but certain requirements seem essential. Given samples from an unknown language, the trained model should produce valid strings not seen in training and be expressive enough to capture the language's full richness. Otherwise, outputting invalid strings constitutes "hallucination," and failing to capture the full range leads to "mode collapse." We ask if a language model can meet both requirements. We investigate this within a statistical language generation setting building on Gold and Angluin. Here, the model receives random samples from a distribution over an unknown language K, which belongs to a possibly infinite collection of languages. The goal is to generate unseen strings from K. We say the model generates from K with consistency and breadth if, as training size increases, its output converges to all unseen strings in K. Kleinberg and Mullainathan [KM24] asked if consistency and breadth in language generation are possible. We answer this negatively: for a large class of language models, including next-token prediction models, this is impossible for most collections of candidate languages. This contrasts with [KM24]'s result, showing consistent generation without breadth is possible for any countable collection of languages. Our finding highlights that generation with breadth fundamentally differs from generation without breadth. As a byproduct, we establish near-tight bounds on the number of samples needed for generation with or without breadth. Finally, our results offer hope: consistent generation with breadth is achievable for any countable collection of languages when negative examples (strings outside K) are available alongside positive ones. This suggests that post-training feedback, which encodes negative examples, can be crucial in reducing hallucinations while limiting mode collapse.
This work explores expanding the capabilities of large language models (LLMs) pretrained on text to generate 3D meshes within a unified model. This offers key advantages of (1) leveraging spatial knowledge already embedded in LLMs, derived from textual sources like 3D tutorials, and (2) enabling conversational 3D generation and mesh understanding. A primary challenge is effectively tokenizing 3D mesh data into discrete tokens that LLMs can process seamlessly. To address this, we introduce LLaMA-Mesh, a novel approach that represents the vertex coordinates and face definitions of 3D meshes as plain text, allowing direct integration with LLMs without expanding the vocabulary. We construct a supervised fine-tuning (SFT) dataset enabling pretrained LLMs to (1) generate 3D meshes from text prompts, (2) produce interleaved text and 3D mesh outputs as required, and (3) understand and interpret 3D meshes. Our work is the first to demonstrate that LLMs can be fine-tuned to acquire complex spatial knowledge for 3D mesh generation in a text-based format, effectively unifying the 3D and text modalities. LLaMA-Mesh achieves mesh generation quality on par with models trained from scratch while maintaining strong text generation performance.
Although people are impressed by the content generation skills of large language models, the use of LLMs, such as ChatGPT, is limited by the domain grounding of the content. The correctness and groundedness of the generated content need to be based on a verified context, such as results from Retrieval-Augmented Generation (RAG). One important issue when adapting LLMs to a customized domain is that the generated responses are often incomplete, or the additions are not verified and may even be hallucinated. Prior studies on hallucination detection have focused on evaluation metrics, which are not easily adaptable to dynamic domains and can be vulnerable to attacks like jail-breaking. In this work, we propose 1) a post-processing algorithm that leverages knowledge triplets in RAG context to correct hallucinations and 2) a dual-decoder model that fuses RAG context to guide the generation process.
Significant advances have been made in natural language processing in recent years. However, our current deep learning approach to language modeling requires substantial resources in terms of data and computation. One of the side effects of this data-hungry paradigm is the current schism between languages, separating those considered high-resource, where most of the development happens and resources are available, and the low-resource ones, which struggle to attain the same level of performance and autonomy. This study aims to introduce a new set of resources to stimulate the future development of neural text generation in Portuguese. In this work, we document the development of GigaVerbo, a concatenation of deduplicated Portuguese text corpora amounting to 200 billion tokens. Via this corpus, we trained a series of decoder-transformers named Tucano. Our models perform equal or superior to other Portuguese and multilingual language models of similar size in several Portuguese benchmarks. The evaluation of our models also reveals that model performance on many currently available benchmarks used by the Portuguese NLP community has little to no correlation with the scaling of token ingestion during training, highlighting the limitations of such evaluations when it comes to the assessment of Portuguese generative language models. All derivatives of our study are openly released on GitHub and Hugging Face. See https://nkluge-correa.github.io/Tucano/
Large Language Models (LLMs) are increasingly adopted for complex scientific text generation tasks, yet they often suffer from limitations in accuracy, consistency, and hallucination control. This thesis introduces a Parameter-Efficient Fine-Tuning (PEFT) approach tailored for GPT-like models, aiming to mitigate hallucinations and enhance reproducibility, particularly in the computational domain of mass spectrometry. We implemented Low-Rank Adaptation (LoRA) adapters to refine GPT-2, termed MS-GPT, using a specialized corpus of mass spectrometry literature. Through novel evaluation methods applied to LLMs, including BLEU, ROUGE, and Perplexity scores, the fine-tuned MS-GPT model demonstrated superior text coherence and reproducibility compared to the baseline GPT-2, confirmed through statistical analysis with the Wilcoxon rank-sum test. Further, we propose a reproducibility metric based on cosine similarity of model outputs under controlled prompts, showcasing MS-GPT's enhanced stability. This research highlights PEFT's potential to optimize LLMs for scientific contexts, reducing computational costs while improving model reliability.
Large language models (LLMs) are often equipped with multi-sample decoding strategies. An LLM implicitly defines an arithmetic code book, facilitating efficient and embarrassingly parallelizable \textbf{arithmetic sampling} to produce multiple samples using quasi-random codes. Traditional text generation methods, such as beam search and sampling-based techniques, have notable limitations: they lack parallelizability or diversity of sampled sequences. This study explores the potential of arithmetic sampling, contrasting it with ancestral sampling across two decoding tasks that employ multi-sample inference: chain-of-thought reasoning with self-consistency and machine translation with minimum Bayes risk decoding. Our results demonstrate that arithmetic sampling produces more diverse samples, significantly improving reasoning and translation performance as the sample size increases. We observe a $\mathbf{3\text{-}5\%}$ point increase in accuracy on the GSM8K dataset and a $\mathbf{0.45\text{-}0.89\%}$ point increment in COMET score for WMT19 tasks using arithmetic sampling without any significant computational overhead.
Large Language Models (LLMs) have demonstrated remarkable capabilities in generating human-like text, but their output may not be aligned with the user or even produce harmful content. This paper presents a novel approach to detect and steer concepts such as toxicity before generation. We introduce the Sparse Conditioned Autoencoder (SCAR), a single trained module that extends the otherwise untouched LLM. SCAR ensures full steerability, towards and away from concepts (e.g., toxic content), without compromising the quality of the model's text generation on standard evaluation benchmarks. We demonstrate the effective application of our approach through a variety of concepts, including toxicity, safety, and writing style alignment. As such, this work establishes a robust framework for controlling LLM generations, ensuring their ethical and safe deployment in real-world applications.
We introduce the GEM (Generative Estimator for Mutual Information), an evaluation metric for assessing language generation by Large Language Models (LLMs), particularly in generating informative judgments, without the need for a gold standard reference. GEM broadens the scenarios where we can benchmark LLM generation performance-from traditional ones, like machine translation and summarization, where gold standard references are readily available, to subjective tasks without clear gold standards, such as academic peer review. GEM uses a generative model to estimate mutual information between candidate and reference responses, without requiring the reference to be a gold standard. In experiments on a human-annotated dataset, GEM demonstrates competitive correlations with human scores compared to the state-of-the-art GPT-4o Examiner, and outperforms all other baselines. Additionally, GEM is more robust against strategic manipulations, such as rephrasing or elongation, which can artificially inflate scores under a GPT-4o Examiner. We also present GRE-bench (Generating Review Evaluation Benchmark) which evaluates LLMs based on how well they can generate high-quality peer reviews for academic research papers. Because GRE-bench is based upon GEM, it inherits its robustness properties. Additionally, GRE-bench circumvents data contamination problems (or data leakage) by using the continuous influx of new open-access research papers and peer reviews each year. We show GRE-bench results of various popular LLMs on their peer review capabilities using the ICLR2023 dataset.
Large Language Models (LLMs) have shown impressive capability in language generation and understanding, but their tendency to hallucinate and produce factually incorrect information remains a key limitation. To verify LLM-generated contents and claims from other sources, traditional verification approaches often rely on holistic models that assign a single factuality label to complex claims, potentially obscuring nuanced errors. In this paper, we advocate for a shift toward fine-grained verification, where complex claims are broken down into smaller sub-claims for individual verification, allowing for more precise identification of inaccuracies, improved transparency, and reduced ambiguity in evidence retrieval. However, generating sub-claims poses challenges, such as maintaining context and ensuring semantic equivalence with respect to the original claim. We introduce FactLens, a benchmark for evaluating fine-grained fact verification, with metrics and automated evaluators of sub-claim quality. The benchmark data is manually curated to ensure high-quality ground truth. Our results show alignment between automated FactLens evaluators and human judgments, and we discuss the impact of sub-claim characteristics on the overall verification performance.
Since the COVID-19 pandemic, clinicians have seen a large and sustained influx in patient portal messages, significantly contributing to clinician burnout. To the best of our knowledge, there are no large-scale public patient portal messages corpora researchers can use to build tools to optimize clinician portal workflows. Informed by our ongoing work with a regional hospital, this study introduces an LLM-powered framework for configurable and realistic patient portal message generation. Our approach leverages few-shot grounded text generation, requiring only a small number of de-identified patient portal messages to help LLMs better match the true style and tone of real data. Clinical experts in our team deem this framework as HIPAA-friendly, unlike existing privacy-preserving approaches to synthetic text generation which cannot guarantee all sensitive attributes will be protected. Through extensive quantitative and human evaluation, we show that our framework produces data of higher quality than comparable generation methods as well as all related datasets. We believe this work provides a path forward for (i) the release of large-scale synthetic patient message datasets that are stylistically similar to ground-truth samples and (ii) HIPAA-friendly data generation which requires minimal human de-identification efforts.
Large Language Models (LLMs) have shown significant advances in text generation but often lack the reliability needed for autonomous deployment in high-stakes domains like healthcare, law, and finance. Existing approaches rely on external knowledge or human oversight, limiting scalability. We introduce a novel framework that repurposes ensemble methods for content validation through model consensus. In tests across 78 complex cases requiring factual accuracy and causal consistency, our framework improved precision from 73.1% to 93.9% with two models (95% CI: 83.5%-97.9%) and to 95.6% with three models (95% CI: 85.2%-98.8%). Statistical analysis indicates strong inter-model agreement ($\kappa$ > 0.76) while preserving sufficient independence to catch errors through disagreement. We outline a clear pathway to further enhance precision with additional validators and refinements. Although the current approach is constrained by multiple-choice format requirements and processing latency, it offers immediate value for enabling reliable autonomous AI systems in critical applications.
Query-focused summarization (QFS) is the task of generating a summary in response to a user-written query. Despite its user-oriented nature, there has been limited work in QFS in explicitly considering a user's understanding of a generated summary, potentially causing QFS systems to underperform at inference time. In this paper, we adapt the Rational Speech Act (RSA) framework, a model of human communication, to explicitly model a reader's understanding of a query-focused summary and integrate it within the generation method of existing QFS systems. In particular, we introduce the answer reconstruction objective which approximates a reader's understanding of a summary by their ability to use it to reconstruct the answer to their initial query. Using this objective, we are able to re-rank candidate summaries generated by existing QFS systems and select summaries that better align with their corresponding query and reference summary. More generally, our study suggests that a simple and effective way of improving a language generation system designed for a user-centered task may be to explicitly incorporate its user requirements into the system's generation procedure.
Large Language Models (LLMs) face significant challenges at inference time due to their high computational demands. To address this, we present Performance-Guided Knowledge Distillation (PGKD), a cost-effective and high-throughput solution for production text classification applications. PGKD utilizes teacher-student Knowledge Distillation to distill the knowledge of LLMs into smaller, task-specific models. PGKD establishes an active learning routine between the student model and the LLM; the LLM continuously generates new training data leveraging hard-negative mining, student model validation performance, and early-stopping protocols to inform the data generation. By employing a cyclical, performance-aware approach tailored for highly multi-class, sparsely annotated datasets prevalent in industrial text classification, PGKD effectively addresses training challenges and outperforms traditional BERT-base models and other knowledge distillation methods on several multi-class classification datasets. Additionally, cost and latency benchmarking reveals that models fine-tuned with PGKD are up to 130X faster and 25X less expensive than LLMs for inference on the same classification task. While PGKD is showcased for text classification tasks, its versatile framework can be extended to any LLM distillation task, including language generation, making it a powerful tool for optimizing performance across a wide range of AI applications.
Evaluating the quality of automatically generated image descriptions is a complex task that requires metrics capturing various dimensions, such as grammaticality, coverage, accuracy, and truthfulness. Although human evaluation provides valuable insights, its cost and time-consuming nature pose limitations. Existing automated metrics like BLEU, ROUGE, METEOR, and CIDEr attempt to fill this gap, but they often exhibit weak correlations with human judgment. To address this challenge, we propose a novel evaluation framework called Image2Text2Image, which leverages diffusion models, such as Stable Diffusion or DALL-E, for text-to-image generation. In the Image2Text2Image framework, an input image is first processed by a selected image captioning model, chosen for evaluation, to generate a textual description. Using this generated description, a diffusion model then creates a new image. By comparing features extracted from the original and generated images, we measure their similarity using a designated similarity metric. A high similarity score suggests that the model has produced a faithful textual description, while a low score highlights discrepancies, revealing potential weaknesses in the model's performance. Notably, our framework does not rely on human-annotated reference captions, making it a valuable tool for assessing image captioning models. Extensive experiments and human evaluations validate the efficacy of our proposed Image2Text2Image evaluation framework. The code and dataset will be published to support further research in the community.
Current automated fact-checking (AFC) approaches commonly evaluate evidence either implicitly via the predicted verdicts or by comparing retrieved evidence with a predefined closed knowledge source, such as Wikipedia. However, these methods suffer from limitations, resulting from their reliance on evaluation metrics developed for different purposes and constraints imposed by closed knowledge sources. Recent advances in natural language generation (NLG) evaluation offer new possibilities for evidence assessment. In this work, we introduce Ev2R, an evaluation framework for AFC that comprises three types of approaches for evidence evaluation: reference-based, proxy-reference, and reference-less. We evaluate their effectiveness through agreement with human ratings and adversarial tests, and demonstrate that prompt-based scorers, particularly those leveraging LLMs and reference evidence, outperform traditional evaluation approaches.
Despite significant advancements in automated report generation, the opaqueness of text interpretability continues to cast doubt on the reliability of the content produced. This paper introduces a novel approach to identify specific image features in X-ray images that influence the outputs of report generation models. Specifically, we propose Cyclic Vision-Language Manipulator CVLM, a module to generate a manipulated X-ray from an original X-ray and its report from a designated report generator. The essence of CVLM is that cycling manipulated X-rays to the report generator produces altered reports aligned with the alterations pre-injected into the reports for X-ray generation, achieving the term "cyclic manipulation". This process allows direct comparison between original and manipulated X-rays, clarifying the critical image features driving changes in reports and enabling model users to assess the reliability of the generated texts. Empirical evaluations demonstrate that CVLM can identify more precise and reliable features compared to existing explanation methods, significantly enhancing the transparency and applicability of AI-generated reports.
Recent progress on large language models (LLMs) has enabled dialogue agents to generate highly naturalistic and plausible text. However, current LLM language generation focuses on responding accurately to questions and requests with a single effective response. In reality, many real dialogues are interactive, meaning an agent's utterances will influence their conversational partner, elicit information, or change their opinion. Accounting for how an agent can effectively steer a conversation is a crucial ability in many dialogue tasks, from healthcare to preference elicitation. Existing methods for fine-tuning dialogue agents to accomplish such tasks would rely on curating some amount of expert data. However, doing so often requires understanding the underlying cognitive processes of the conversational partner, which is a skill neither humans nor LLMs trained on human data can reliably do. Our key insight is that while LLMs may not be adept at identifying effective strategies for steering conversations a priori, or in the middle of an ongoing conversation, they can do so post-hoc, or in hindsight, after seeing how their conversational partner responds. We use this fact to rewrite and augment existing suboptimal data, and train via offline reinforcement learning (RL) an agent that outperforms both prompting and learning from unaltered human demonstrations. We apply our approach to two domains that require understanding human mental state, intelligent interaction, and persuasion: mental health support, and soliciting charitable donations. Our results in a user study with real humans show that our approach greatly outperforms existing state-of-the-art dialogue agents.