In recent years, the field of Natural Language Generation (NLG) has been boosted by the recent advances in deep learning technologies. Nonetheless, these new data-intensive methods introduce language-dependent disparities in NLG as the main training data sets are in English. Also, most neural NLG systems use decoder-only (causal) transformer language models, which work well for English, but were not designed with other languages in mind. In this work we depart from the hypothesis that they may introduce generation bias in target languages with less rigid word ordering, subject omission, or different attachment preferences for relative clauses, so that for these target languages other language generation strategies may be more desirable. This paper first compares causal and non-causal language modeling for English and Spanish, two languages with different grammatical structures and over 1.5 billion and 0.5 billion speakers, respectively. For this purpose, we define a novel metric of average causal and non-causal context-conditioned entropy of the grammatical category distribution for both languages as an information-theoretic a priori approach. The evaluation of natural text sources (such as training data) in both languages reveals lower average non-causal conditional entropy in Spanish and lower causal conditional entropy in English. According to this experiment, Spanish is more predictable than English given a non-causal context. Then, by applying a conditional relative entropy metric to text generation experiments, we obtain as insights that the best performance is respectively achieved with causal NLG in English, and with non-causal NLG in Spanish. These insights support further research in NLG in Spanish using bidirectional transformer language models.
Large language models (LLMs) have shown potential in biomedical applications, leading to efforts to fine-tune them on domain-specific data. However, the effectiveness of this approach remains unclear. This study evaluates the performance of biomedically fine-tuned LLMs against their general-purpose counterparts on a variety of clinical tasks. We evaluated their performance on clinical case challenges from the New England Journal of Medicine (NEJM) and the Journal of the American Medical Association (JAMA) and on several clinical tasks (e.g., information extraction, document summarization, and clinical coding). Using benchmarks specifically chosen to be likely outside the fine-tuning datasets of biomedical models, we found that biomedical LLMs mostly perform inferior to their general-purpose counterparts, especially on tasks not focused on medical knowledge. While larger models showed similar performance on case tasks (e.g., OpenBioLLM-70B: 66.4% vs. Llama-3-70B-Instruct: 65% on JAMA cases), smaller biomedical models showed more pronounced underperformance (e.g., OpenBioLLM-8B: 30% vs. Llama-3-8B-Instruct: 64.3% on NEJM cases). Similar trends were observed across the CLUE (Clinical Language Understanding Evaluation) benchmark tasks, with general-purpose models often performing better on text generation, question answering, and coding tasks. Our results suggest that fine-tuning LLMs to biomedical data may not provide the expected benefits and may potentially lead to reduced performance, challenging prevailing assumptions about domain-specific adaptation of LLMs and highlighting the need for more rigorous evaluation frameworks in healthcare AI. Alternative approaches, such as retrieval-augmented generation, may be more effective in enhancing the biomedical capabilities of LLMs without compromising their general knowledge.
Large Language Models (LLMs) are increasingly serving as evaluators in Natural Language Generation (NLG) tasks. However, the capabilities of LLMs in scoring NLG quality remain inadequately explored. Current studies depend on human assessments and simple metrics that fail to capture the discernment of LLMs across diverse NLG tasks. To address this gap, we propose the Discernment of Hierarchical Perturbation (DHP) benchmarking framework, which provides quantitative discernment scores for LLMs utilizing hierarchically perturbed text data and statistical tests to measure the NLG evaluation capabilities of LLMs systematically. We have re-established six evaluation datasets for this benchmark, covering four NLG tasks: Summarization, Story Completion, Question Answering, and Translation. Our comprehensive benchmarking of five major LLM series provides critical insight into their strengths and limitations as NLG evaluators.
Sampling-based decoding strategies have been widely adopted for Large Language Models (LLMs) in numerous applications, targeting a balance between diversity and quality via temperature tuning and tail truncation. Considering the strong dependency of the candidate next tokens on different prefixes, recent studies propose to adaptively truncate the tail of LLMs' predicted distribution. Although improved results have been reported with these methods on open-ended text generation tasks, the results are highly dependent on the curated parameters and the limited exemplar text. In this paper, we propose a systematic way to estimate the capacity of a truncation sampling method by considering the trade-off between diversity and risk at each decoding step, based on our collected prefix tree which preserves the context of a full sentence. Our work offers a comprehensive comparison of existing truncation sampling methods and serves as a practical user guideline for their parameter selection.
Generative large language models (LLMs) have achieved state-of-the-art results on a wide range of tasks, yet they remain susceptible to backdoor attacks: carefully crafted triggers in the input can manipulate the model to produce adversary-specified outputs. While prior research has predominantly focused on backdoor risks in vision and classification settings, the vulnerability of LLMs in open-ended text generation remains underexplored. To fill this gap, we introduce BackdoorLLM (Our BackdoorLLM benchmark was awarded First Prize in the SafetyBench competition, https://www.mlsafety.org/safebench/winners, organized by the Center for AI Safety, https://safe.ai/.), the first comprehensive benchmark for systematically evaluating backdoor threats in text-generation LLMs. BackdoorLLM provides: (i) a unified repository of benchmarks with a standardized training and evaluation pipeline; (ii) a diverse suite of attack modalities, including data poisoning, weight poisoning, hidden-state manipulation, and chain-of-thought hijacking; (iii) over 200 experiments spanning 8 distinct attack strategies, 7 real-world scenarios, and 6 model architectures; (iv) key insights into the factors that govern backdoor effectiveness and failure modes in LLMs; and (v) a defense toolkit encompassing 7 representative mitigation techniques. Our code and datasets are available at https://github.com/bboylyg/BackdoorLLM. We will continuously incorporate emerging attack and defense methodologies to support the research in advancing the safety and reliability of LLMs.
Speculative decoding (SD), where an extra draft model is employed to provide multiple \textit{draft} tokens first and then the original target model verifies these tokens in parallel, has shown great power for LLM inference acceleration. However, existing SD methods suffer from the mutual waiting problem, i.e., the target model gets stuck when the draft model is \textit{guessing} tokens, and vice versa. This problem is directly incurred by the asynchronous execution of the draft model and the target model, and is exacerbated due to the fixed draft length in speculative decoding. To address these challenges, we propose a conceptually simple, flexible, and general framework to boost speculative decoding, namely \textbf{P}arallel sp\textbf{E}culative decoding with \textbf{A}daptive d\textbf{R}aft \textbf{L}ength (PEARL). Specifically, PEARL proposes \textit{pre-verify} to verify the first draft token in advance during the drafting phase, and \textit{post-verify} to generate more draft tokens during the verification phase. PEARL parallels the drafting phase and the verification phase via applying the two strategies, and achieves adaptive draft length for different scenarios, which effectively alleviates the mutual waiting problem. Moreover, we theoretically demonstrate that the mean accepted tokens of PEARL is more than existing \textit{draft-then-verify} works. Experiments on various text generation benchmarks demonstrate the effectiveness of our \name, leading to a superior speedup performance up to \textbf{3.79$\times$} and \textbf{1.52$\times$}, compared to auto-regressive decoding and vanilla speculative decoding, respectively.
This paper explores the application of Stochastic Differential Equations (SDE) to interpret the text generation process of Large Language Models (LLMs) such as GPT-4. Text generation in LLMs is modeled as a stochastic process where each step depends on previously generated content and model parameters, sampling the next word from a vocabulary distribution. We represent this generation process using SDE to capture both deterministic trends and stochastic perturbations. The drift term describes the deterministic trends in the generation process, while the diffusion term captures the stochastic variations. We fit these functions using neural networks and validate the model on real-world text corpora. Through numerical simulations and comprehensive analyses, including drift and diffusion analysis, stochastic process property evaluation, and phase space exploration, we provide deep insights into the dynamics of text generation. This approach not only enhances the understanding of the inner workings of LLMs but also offers a novel mathematical perspective on language generation, which is crucial for diagnosing, optimizing, and controlling the quality of generated text.
In Natural Language Processing (NLP), Large Language Models (LLMs) have demonstrated high text generation quality. However, in real-world applications, LLMs must meet increasingly complex requirements. Beyond avoiding misleading or inappropriate content, LLMs are also expected to cater to specific user needs, such as imitating particular writing styles or generating text with poetic richness. These varied demands have driven the development of Controllable Text Generation (CTG) techniques, which ensure that outputs adhere to predefined control conditions--such as safety, sentiment, thematic consistency, and linguistic style--while maintaining high standards of helpfulness, fluency, and diversity. This paper systematically reviews the latest advancements in CTG for LLMs, offering a comprehensive definition of its core concepts and clarifying the requirements for control conditions and text quality. We categorize CTG tasks into two primary types: content control and attribute control. The key methods are discussed, including model retraining, fine-tuning, reinforcement learning, prompt engineering, latent space manipulation, and decoding-time intervention. We analyze each method's characteristics, advantages, and limitations, providing nuanced insights for achieving generation control. Additionally, we review CTG evaluation methods, summarize its applications across domains, and address key challenges in current research, including reduced fluency and practicality. We also propose several appeals, such as placing greater emphasis on real-world applications in future research. This paper aims to offer valuable guidance to researchers and developers in the field. Our reference list and Chinese version are open-sourced at https://github.com/IAAR-Shanghai/CTGSurvey.
Iterative data generation and model re-training can effectively align large language models(LLMs) to human preferences. The process of data sampling is crucial, as it significantly influences the success of policy improvement. Repeated random sampling is a widely used method that independently queries the model multiple times to generate outputs. In this work, we propose a more effective sampling method, named Preference-Guided Reflective Sampling (PRS). Unlike random sampling, PRS employs a tree-based generation framework to enable more efficient sampling. It leverages adaptive self-refinement techniques to better explore the sampling space. By specifying user preferences in natural language, PRS can further optimize response generation according to these preferences. As a result, PRS can align models to diverse user preferences. Our experiments demonstrate that PRS generates higher-quality responses with significantly higher rewards. On AlpacaEval and Arena-Hard, PRS substantially outperforms repeated random sampling in best-of-$N$ sampling. Moreover, PRS shows strong performance when applied in iterative offline RL training.
The clinical diagnosis of most mental disorders primarily relies on the conversations between psychiatrist and patient. The creation of such diagnostic conversation datasets is promising to boost the AI mental healthcare community. However, directly collecting the conversations in real diagnosis scenarios is near impossible due to stringent privacy and ethical considerations. To address this issue, we seek to synthesize diagnostic conversation by exploiting anonymized patient cases that are easier to access. Specifically, we design a neuro-symbolic multi-agent framework for synthesizing the diagnostic conversation of mental disorders with large language models. It takes patient case as input and is capable of generating multiple diverse conversations with one single patient case. The framework basically involves the interaction between a doctor agent and a patient agent, and generates conversations under symbolic control via a dynamic diagnosis tree. By applying the proposed framework, we develop the largest Chinese mental disorders diagnosis dataset MDD-5k. This dataset is built upon 1000 real, anonymized patient cases by cooperating with Shanghai Mental Health Center and comprises 5000 high-quality long conversations with diagnosis results and treatment opinions as labels. To the best of our knowledge, it's also the first labeled dataset for Chinese mental disorders diagnosis. Human evaluation demonstrates the proposed MDD-5k dataset successfully simulates human-like diagnostic process of mental disorders.
Large language models (LLMs) have exhibited remarkable capabilities in natural language generation, but they have also been observed to magnify societal biases, particularly those related to gender. In response to this issue, several benchmarks have been proposed to assess gender bias in LLMs. However, these benchmarks often lack practical flexibility or inadvertently introduce biases. To address these shortcomings, we introduce GenderCARE, a comprehensive framework that encompasses innovative Criteria, bias Assessment, Reduction techniques, and Evaluation metrics for quantifying and mitigating gender bias in LLMs. To begin, we establish pioneering criteria for gender equality benchmarks, spanning dimensions such as inclusivity, diversity, explainability, objectivity, robustness, and realisticity. Guided by these criteria, we construct GenderPair, a novel pair-based benchmark designed to assess gender bias in LLMs comprehensively. Our benchmark provides standardized and realistic evaluations, including previously overlooked gender groups such as transgender and non-binary individuals. Furthermore, we develop effective debiasing techniques that incorporate counterfactual data augmentation and specialized fine-tuning strategies to reduce gender bias in LLMs without compromising their overall performance. Extensive experiments demonstrate a significant reduction in various gender bias benchmarks, with reductions peaking at over 90% and averaging above 35% across 17 different LLMs. Importantly, these reductions come with minimal variability in mainstream language tasks, remaining below 2%. By offering a realistic assessment and tailored reduction of gender biases, we hope that our GenderCARE can represent a significant step towards achieving fairness and equity in LLMs. More details are available at https://github.com/kstanghere/GenderCARE-ccs24.
Recent developments in the field of Natural Language Processing, especially language models such as the transformer have brought state-of-the-art results in language understanding and language generation. In this work, we investigate the use of the transformer model for radiology report generation from chest X-rays. We also highlight limitations in evaluating radiology report generation using only the standard language generation metrics. We then applied a transformer based radiology report generation architecture, and also compare the performance of a transformer based decoder with the recurrence based decoder. Experiments were performed using the IU-CXR dataset, showing superior results to its LSTM counterpart and being significantly faster. Finally, we identify the need of evaluating radiology report generation system using both language generation metrics and classification metrics, which helps to provide robust measure of generated reports in terms of their coherence and diagnostic value.
In the rapidly evolving landscape of Natural Language Processing (NLP) and text generation, the emergence of Retrieval Augmented Generation (RAG) presents a promising avenue for improving the quality and reliability of generated text by leveraging information retrieved from user specified database. Benchmarking is essential to evaluate and compare the performance of the different RAG configurations in terms of retriever and generator, providing insights into their effectiveness, scalability, and suitability for the specific domain and applications. In this paper, we present a comprehensive framework to generate a domain relevant RAG benchmark. Our framework is based on automatic question-answer generation with Human (domain experts)-AI Large Language Model (LLM) teaming. As a case study, we demonstrate the framework by introducing WeQA, a first-of-its-kind benchmark on the wind energy domain which comprises of multiple scientific documents/reports related to environmental impact of wind energy projects. Our framework systematically evaluates RAG performance using diverse metrics and multiple question types with varying complexity level. We also demonstrate the performance of different models on our benchmark.
Large Language Models (LLMs) have shown promise in clinical applications through prompt engineering, allowing flexible clinical predictions. However, they struggle to produce reliable prediction probabilities, which are crucial for transparency and decision-making. While explicit prompts can lead LLMs to generate probability estimates, their numerical reasoning limitations raise concerns about reliability. We compared explicit probabilities from text generation to implicit probabilities derived from the likelihood of predicting the correct label token. Across six advanced open-source LLMs and five medical datasets, explicit probabilities consistently underperformed implicit probabilities in discrimination, precision, and recall. This discrepancy is more pronounced with smaller LLMs and imbalanced datasets, highlighting the need for cautious interpretation, improved probability estimation methods, and further research for clinical use of LLMs.
The fashion domain encompasses a variety of real-world multimodal tasks, including multimodal retrieval and multimodal generation. The rapid advancements in artificial intelligence generated content, particularly in technologies like large language models for text generation and diffusion models for visual generation, have sparked widespread research interest in applying these multimodal models in the fashion domain. However, tasks involving embeddings, such as image-to-text or text-to-image retrieval, have been largely overlooked from this perspective due to the diverse nature of the multimodal fashion domain. And current research on multi-task single models lack focus on image generation. In this work, we present UniFashion, a unified framework that simultaneously tackles the challenges of multimodal generation and retrieval tasks within the fashion domain, integrating image generation with retrieval tasks and text generation tasks. UniFashion unifies embedding and generative tasks by integrating a diffusion model and LLM, enabling controllable and high-fidelity generation. Our model significantly outperforms previous single-task state-of-the-art models across diverse fashion tasks, and can be readily adapted to manage complex vision-language tasks. This work demonstrates the potential learning synergy between multimodal generation and retrieval, offering a promising direction for future research in the fashion domain. The source code is available at https://github.com/xiangyu-mm/UniFashion.
This paper explores the advancements and applications of language models in healthcare, focusing on their clinical use cases. It examines the evolution from early encoder-based systems requiring extensive fine-tuning to state-of-the-art large language and multimodal models capable of integrating text and visual data through in-context learning. The analysis emphasizes locally deployable models, which enhance data privacy and operational autonomy, and their applications in tasks such as text generation, classification, information extraction, and conversational systems. The paper also highlights a structured organization of tasks and a tiered ethical approach, providing a valuable resource for researchers and practitioners, while discussing key challenges related to ethics, evaluation, and implementation.
Commentary provides readers with a deep understanding of events by presenting diverse arguments and evidence. However, creating commentary is a time-consuming task, even for skilled commentators. Large language models (LLMs) have simplified the process of natural language generation, but their direct application in commentary creation still faces challenges due to unique task requirements. These requirements can be categorized into two levels: 1) fundamental requirements, which include creating well-structured and logically consistent narratives, and 2) advanced requirements, which involve generating quality arguments and providing convincing evidence. In this paper, we introduce Xinyu, an efficient LLM-based system designed to assist commentators in generating Chinese commentaries. To meet the fundamental requirements, we deconstruct the generation process into sequential steps, proposing targeted strategies and supervised fine-tuning (SFT) for each step. To address the advanced requirements, we present an argument ranking model for arguments and establish a comprehensive evidence database that includes up-to-date events and classic books, thereby strengthening the substantiation of the evidence with retrieval augmented generation (RAG) technology. To evaluate the generated commentaries more fairly, corresponding to the two-level requirements, we introduce a comprehensive evaluation metric that considers five distinct perspectives in commentary generation. Our experiments confirm the effectiveness of our proposed system. We also observe a significant increase in the efficiency of commentators in real-world scenarios, with the average time spent on creating a commentary dropping from 4 hours to 20 minutes. Importantly, such an increase in efficiency does not compromise the quality of the commentaries.
In text generation, hallucinations refer to the generation of seemingly coherent text that contradicts established knowledge. One compelling hypothesis is that hallucinations occur when a language model is given a generation task outside its parametric knowledge (due to rarity, recency, domain, etc.). A common strategy to address this limitation is to infuse the language models with retrieval mechanisms, providing the model with relevant knowledge for the task. In this paper, we leverage the planning capabilities of instruction-tuned LLMs and analyze how planning can be used to guide retrieval to further reduce the frequency of hallucinations. We empirically evaluate several variations of our proposed approach on long-form text generation tasks. By improving the coverage of relevant facts, plan-guided retrieval and generation can produce more informative responses while providing a higher rate of attribution to source documents.
For many low-resource languages, the only available language models are large multilingual models trained on many languages simultaneously. However, using FLORES perplexity as a metric, we find that these models perform worse than bigrams for many languages (e.g. 24% of languages in XGLM 4.5B; 43% in BLOOM 7.1B). To facilitate research that focuses on low-resource languages, we pre-train and release Goldfish, a suite of monolingual autoregressive Transformer language models up to 125M parameters for 350 languages. The Goldfish reach lower FLORES perplexities than BLOOM, XGLM, and MaLA-500 on 98 of 204 FLORES languages, despite each Goldfish model being over 10x smaller. However, the Goldfish significantly underperform larger multilingual models on reasoning benchmarks, suggesting that for low-resource languages, multilinguality primarily improves general reasoning abilities rather than basic text generation. We release models trained on 5MB (350 languages), 10MB (288 languages), 100MB (166 languages), and 1GB (83 languages) of text data where available. The Goldfish models are available as baselines, fine-tuning sources, or augmentations to existing models in low-resource NLP research, and they are further useful for crosslinguistic studies requiring maximally comparable models across languages.
Although instruction tuning is widely used to adjust behavior in Large Language Models (LLMs), extensive empirical evidence and research indicates that it is primarily a process where the model fits to specific task formats, rather than acquiring new knowledge or capabilities. We propose that this limitation stems from biased features learned during instruction tuning, which differ from ideal task-specfic features, leading to learn less underlying semantics in downstream tasks. However, ideal features are unknown and incalculable, constraining past work to rely on prior knowledge to assist reasoning or training, which limits LLMs' capabilities to the developers' abilities, rather than data-driven scalable learning. In our paper, through our novel data synthesis method, DELIA (Diversity-Enhanced Learning for Instruction Adaptation), we leverage the buffering effect of extensive diverse data in LLMs training to transform biased features in instruction tuning into approximations of ideal features, without explicit prior ideal features. Experiments show DELIA's better performance compared to common instruction tuning and other baselines. It outperforms common instruction tuning by 17.07%-33.41% on Icelandic-English translation bleurt score (WMT-21 dataset, gemma-7b-it) and improves accuracy by 36.1% on formatted text generation (Llama2-7b-chat). Notably, among knowledge injection methods we've known, DELIA uniquely align the internal representations of new special tokens with their prior semantics.