In this work, we explicitly show that modern LLMs tend to generate correct facts first, then "drift away" and generate incorrect facts later: this was occasionally observed but never properly measured. We develop a semantic drift score that measures the degree of separation between correct and incorrect facts in generated texts and confirm our hypothesis when generating Wikipedia-style biographies. This correct-then-incorrect generation pattern suggests that factual accuracy can be improved by knowing when to stop generation. Therefore, we explore the trade-off between information quantity and factual accuracy for several early stopping methods and manage to improve factuality by a large margin. We further show that reranking with semantic similarity can further improve these results, both compared to the baseline and when combined with early stopping. Finally, we try calling external API to bring the model back to the right generation path, but do not get positive results. Overall, our methods generalize and can be applied to any long-form text generation to produce more reliable information, by balancing trade-offs between factual accuracy, information quantity and computational cost.
相關內容
SGD performs worse than Adam by a significant margin on Transformers, but the reason remains unclear. In this work, we provide an explanation of SGD's bad performance on Transformers through the lens of Hessian: (i) Transformers are "heterogeneous": the Hessian spectrum across parameter blocks vary dramatically, a phenomenon we call "block heterogeneity"; (ii) Heterogeneity hampers SGD: SGD performs badly on problems with block heterogeneity. To validate that heterogeneity hampers SGD, we check various Transformers, CNNs, MLPs, and quadratic problems, and find that SGD works well on problems without block heterogeneity but performs badly when the heterogeneity exists. Our initial theoretical analysis indicates that SGD performs poorly because it applies one single learning rate to all blocks, which cannot handle the heterogeneity among blocks. This limitation could be ameliorated if we use coordinate-wise learning rates, as designed in Adam.
Fairness-aware Graph Neural Networks (GNNs) often face a challenging trade-off, where prioritizing fairness may require compromising utility. In this work, we re-examine fairness through the lens of spectral graph theory, aiming to reconcile fairness and utility within the framework of spectral graph learning. We explore the correlation between sensitive features and spectrum in GNNs, using theoretical analysis to delineate the similarity between original sensitive features and those after convolution under different spectrum. Our analysis reveals a reduction in the impact of similarity when the eigenvectors associated with the largest magnitude eigenvalue exhibit directional similarity. Based on these theoretical insights, we propose FUGNN, a novel spectral graph learning approach that harmonizes the conflict between fairness and utility. FUGNN ensures algorithmic fairness and utility by truncating the spectrum and optimizing eigenvector distribution during the encoding process. The fairness-aware eigenvector selection reduces the impact of convolution on sensitive features while concurrently minimizing the sacrifice of utility. FUGNN further optimizes the distribution of eigenvectors through a transformer architecture. By incorporating the optimized spectrum into the graph convolution network, FUGNN effectively learns node representations. Experiments on six real-world datasets demonstrate the superiority of FUGNN over baseline methods. The codes are available at //github.com/yushuowiki/FUGNN.
Collaborative software development happens in teams, that cooperate on shared artefacts, and discuss development on online platforms. Due to the complexity of development and the variety of teams, software components often act as effective containers for parallel work and teams. Past research has shown how communication between team members, especially in an open-source environment, can become extremely toxic, and lead to members leaving the development team. This has a direct effect on the evolution and maintenance of the project in which the former members were active in. The purpose of our study is two-fold: first, we propose an approach to evaluate, at a finer granularity, the positive and negative emotions in the communication between developers; and second, we aim to characterise a project's development paths, or components, as more or less impacted by the emotions. Our analysis evaluates single sentences rather than whole messages as the finest granularity of communication. The previous study found that the high positivity or negativity at the sentence level may indirectly impact the writer him/herself, or the reader. In this way, we could highlight specific paths of Gentoo as the most affected by negative emotions, and show how negative emotions have evolved and changed along the same paths. By joining the analysis of the mailing lists, from which we derive the sentiment of the developers, with the information derived from the development logs, we obtained a longitudinal picture of how development paths have been historically affected by positive or negative emotions. Our study shows that, in recent years, negative emotions have generally decreased in the communication between Gentoo developers. We also show how file paths, as collaborative software development artefacts, were more or less impacted by the emotions of the developers.
Recent AI systems have shown extremely powerful performance, even surpassing human performance, on various tasks such as information retrieval, language generation, and image generation based on large language models (LLMs). At the same time, there are diverse safety risks that can cause the generation of malicious contents by circumventing the alignment in LLMs, which are often referred to as jailbreaking. However, most of the previous works only focused on the text-based jailbreaking in LLMs, and the jailbreaking of the text-to-image (T2I) generation system has been relatively overlooked. In this paper, we first evaluate the safety of the commercial T2I generation systems, such as ChatGPT, Copilot, and Gemini, on copyright infringement with naive prompts. From this empirical study, we find that Copilot and Gemini block only 12\% and 17\% of the attacks with naive prompts, respectively, while ChatGPT blocks 84\% of them. Then, we further propose a stronger automated jailbreaking pipeline for T2I generation systems, which produces prompts that bypass their safety guards. Our automated jailbreaking framework leverages an LLM optimizer to generate prompts to maximize degree of violation from the generated images without any weight updates or gradient computation. Surprisingly, our simple yet effective approach successfully jailbreaks the ChatGPT with 11.0\% block rate, making it generate copyrighted contents in 76\% of the time. Finally, we explore various defense strategies, such as post-generation filtering and machine unlearning techniques, but found that they were inadequate, which suggests the necessity of stronger defense mechanisms.
In this paper, we make the first attempt to understand and test potential computation efficiency robustness in state-of-the-art LLMs. By analyzing the working mechanism and implementation of 20,543 public-accessible LLMs, we observe a fundamental property in LLMs that could be manipulated in an adversarial manner to reduce computation efficiency significantly. Our key motivation is to generate test inputs that could sufficiently delay the generation of EOS such that LLMs would have to go through enough iterations to satisfy the pre-configured threshold. We present \tool, which can work under both white-box setting and black-box setting. In the white-box scenario, \tool develops a gradient-guided technique that searches for a minimal and unnoticeable perturbation at character-level, token-level, and structure-level. In the black-box scenario, \tool employs a causal inference-based approach to find critical tokens and similarly applies three levels of imperceptible perturbation to them. Both the white-box and black-box settings effectively delay the appearance of EOS, compelling these inputs to reach the naturally-unreachable threshold. To demonstrate the effectiveness of \tool, we conduct a systematic evaluation on nine public-available LLMs: Google T5, AllenAI WMT14, Helsinki-NLP translator, Facebook FairSeq, UNICAMP-DL translator, MarianMT, Google FLAN-T5, MBZUAI LaMini-GPT and Salesforce CodeGen. Experimental results show that \tool can increase on average LLMs' response latency and energy consumption by 325\% to 3244\% and 344\% to 3616\%, respectively, by perturbing just one character or token in the input sentence.
Text generation has become more accessible than ever, and the increasing interest in these systems, especially those using large language models, has spurred an increasing number of related publications. We provide a systematic literature review comprising 244 selected papers between 2017 and 2024. This review categorizes works in text generation into five main tasks: open-ended text generation, summarization, translation, paraphrasing, and question answering. For each task, we review their relevant characteristics, sub-tasks, and specific challenges (e.g., missing datasets for multi-document summarization, coherence in story generation, and complex reasoning for question answering). Additionally, we assess current approaches for evaluating text generation systems and ascertain problems with current metrics. Our investigation shows nine prominent challenges common to all tasks and sub-tasks in recent text generation publications: bias, reasoning, hallucinations, misuse, privacy, interpretability, transparency, datasets, and computing. We provide a detailed analysis of these challenges, their potential solutions, and which gaps still require further engagement from the community. This systematic literature review targets two main audiences: early career researchers in natural language processing looking for an overview of the field and promising research directions, as well as experienced researchers seeking a detailed view of tasks, evaluation methodologies, open challenges, and recent mitigation strategies.
LLMs are computationally expensive to pre-train due to their large scale. Model growth emerges as a promising approach by leveraging smaller models to accelerate the training of larger ones. However, the viability of these model growth methods in efficient LLM pre-training remains underexplored. This work identifies three critical $\underline{\textit{O}}$bstacles: ($\textit{O}$1) lack of comprehensive evaluation, ($\textit{O}$2) untested viability for scaling, and ($\textit{O}$3) lack of empirical guidelines. To tackle $\textit{O}$1, we summarize existing approaches into four atomic growth operators and systematically evaluate them in a standardized LLM pre-training setting. Our findings reveal that a depthwise stacking operator, called $G_{\text{stack}}$, exhibits remarkable acceleration in training, leading to decreased loss and improved overall performance on eight standard NLP benchmarks compared to strong baselines. Motivated by these promising results, we conduct extensive experiments to delve deeper into $G_{\text{stack}}$ to address $\textit{O}$2 and $\textit{O}$3. For $\textit{O}$2 (untested scalability), our study shows that $G_{\text{stack}}$ is scalable and consistently performs well, with experiments up to 7B LLMs after growth and pre-training LLMs with 750B tokens. For example, compared to a conventionally trained 7B model using 300B tokens, our $G_{\text{stack}}$ model converges to the same loss with 194B tokens, resulting in a 54.6\% speedup. We further address $\textit{O}$3 (lack of empirical guidelines) by formalizing guidelines to determine growth timing and growth factor for $G_{\text{stack}}$, making it practical in general LLM pre-training. We also provide in-depth discussions and comprehensive ablation studies of $G_{\text{stack}}$. Our code and pre-trained model are available at $\href{//llm-stacking.github.io/}{//llm-stacking.github.io/}$.
Training large Deep Neural Network (DNN) models requires thousands of GPUs for days or weeks at a time. At these scales, failures are frequent and can have a big impact on training throughput. Restoring performance using spare GPU servers becomes increasingly expensive as models grow. SlipStream is a system for efficient DNN training in the presence of failures, without using spare servers. It exploits the functional redundancy inherent in distributed training systems -- servers hold the same model parameters across data-parallel groups -- as well as the bubbles in the pipeline schedule within each data-parallel group. SlipStream dynamically re-routes the work of a failed server to its data-parallel peers, ensuring continuous training despite multiple failures. However, re-routing work leads to imbalances across pipeline stages that degrades training throughput. SlipStream introduces two optimizations that allow re-routed work to execute within bubbles of the original pipeline schedule. First, it decouples the backward pass computation into two phases. Second, it staggers the execution of the optimizer step across pipeline stages. Combined, these optimizations enable schedules that minimize or even eliminate training throughput degradation during failures. We describe a prototype for SlipStream and show that it achieves high training throughput under multiple failures, outperforming recent proposals for fault-tolerant training such as Oobleck and Bamboo by up to 1.46x and 1.64x, respectively.
Connecting text and visual modalities plays an essential role in generative intelligence. For this reason, inspired by the success of large language models, significant research efforts are being devoted to the development of Multimodal Large Language Models (MLLMs). These models can seamlessly integrate visual and textual modalities, both as input and output, while providing a dialogue-based interface and instruction-following capabilities. In this paper, we provide a comprehensive review of recent visual-based MLLMs, analyzing their architectural choices, multimodal alignment strategies, and training techniques. We also conduct a detailed analysis of these models across a wide range of tasks, including visual grounding, image generation and editing, visual understanding, and domain-specific applications. Additionally, we compile and describe training datasets and evaluation benchmarks, conducting comparisons among existing models in terms of performance and computational requirements. Overall, this survey offers a comprehensive overview of the current state of the art, laying the groundwork for future MLLMs.
Machine learning techniques have deeply rooted in our everyday life. However, since it is knowledge- and labor-intensive to pursue good learning performance, human experts are heavily involved in every aspect of machine learning. In order to make machine learning techniques easier to apply and reduce the demand for experienced human experts, automated machine learning (AutoML) has emerged as a hot topic with both industrial and academic interest. In this paper, we provide an up to date survey on AutoML. First, we introduce and define the AutoML problem, with inspiration from both realms of automation and machine learning. Then, we propose a general AutoML framework that not only covers most existing approaches to date but also can guide the design for new methods. Subsequently, we categorize and review the existing works from two aspects, i.e., the problem setup and the employed techniques. Finally, we provide a detailed analysis of AutoML approaches and explain the reasons underneath their successful applications. We hope this survey can serve as not only an insightful guideline for AutoML beginners but also an inspiration for future research.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191