What happens if we train a new Large Language Model (LLM) using data that are at least partially generated by other LLMs? The explosive success of LLMs means that a substantial amount of content online will be generated by LLMs rather than humans, which will inevitably enter the training datasets of next-generation LLMs. We evaluate the implications of such "regurgitative training" on LLM performance. Through fine-tuning GPT-3.5 with data generated either by itself or by other LLMs in a machine translation task, we find strong evidence that regurgitative training clearly handicaps the performance of LLMs. The same performance loss of regurgitative training is observed on transformer models that we train from scratch. We find suggestive evidence that the performance disadvantage of regurgitative training can be attributed to at least two mechanisms: (1) higher error rates and (2) lower lexical diversity in LLM-generated data as compared to real data. Based on these mechanisms, we propose and evaluate three different strategies to mitigate the performance loss of regurgitative training. First, we devise data-driven metrics to gauge the quality of each LLM-generated data instance, and then carry out an ordered training process where high-quality data are added before low-quality ones. Second, we combine data generated by multiple different LLMs (as an attempt to increase lexical diversity). Third, we train an AI detection classifier to differentiate between LLM- and human-generated data, and include LLM-generated data in the order of resemblance to human-generated data. All three strategies can improve the performance of regurgitative training to some extent but are not always able to fully close the gap from training with real data. Our results highlight the value of real, human-generated data in training LLMs, which cannot be easily substituted by synthetic, LLM-generated data.

The digital revolution and the advent of the world wide web have transformed human communication, notably through the emergence of memes. While memes are a popular and straightforward form of expression, they can also be used to spread misinformation and hate due to their anonymity and ease of use. In response to these challenges, this paper introduces a solution developed by team 'Baseline' for the AI Singapore Online Safety Prize Challenge. Focusing on computational efficiency and feature engineering, the solution achieved an AUROC of 0.76 and an accuracy of 0.69 on the test dataset. As key features, the solution leverages the inherent probabilistic capabilities of large Vision-Language Models (VLMs) to generate task-adapted feature encodings from text, and applies a distilled quantization tailored to the specific cultural nuances present in Singapore. This type of processing and fine-tuning can be adapted to various visual and textual understanding and classification tasks, and even applied on private VLMs such as OpenAI's GPT. Finally it can eliminate the need for extensive model training on large GPUs for resource constrained applications, also offering a solution when little or no data is available.

Event sequences (ESs) arise in many practical domains including finance, retail, social networks, and healthcare. In the context of machine learning, event sequences can be seen as a special type of tabular data with annotated timestamps. Despite the importance of ESs modeling and analysis, little effort was made in adapting large language models (LLMs) to the ESs domain. In this paper, we highlight the common difficulties of ESs processing and propose a novel solution capable of solving multiple downstream tasks with little or no finetuning. In particular, we solve the problem of working with long sequences and improve time and numeric features processing. The resulting method, called ESQA, effectively utilizes the power of LLMs and, according to extensive experiments, achieves state-of-the-art results in the ESs domain.

In this paper, we introduce Dynamic Layer Operations (DLO), a novel approach for vertically scaling transformer-based Large Language Models (LLMs) by dynamically expanding, activating, or skipping layers using a sophisticated routing policy based on layerwise feature similarity. Unlike traditional Mixture-of-Experts (MoE) methods that focus on extending the model width, our approach targets model depth, addressing the redundancy observed across layer representations for various input samples. Our framework is integrated with the Supervised Fine-Tuning (SFT) stage, eliminating the need for resource-intensive Continual Pre-Training (CPT). Experimental results demonstrate that DLO not only outperforms the original unscaled models but also achieves comparable results to densely expanded models with significantly improved efficiency. Our work offers a promising direction for building efficient yet powerful LLMs. We will release our implementation and model weights upon acceptance.

Jailbreaking attacks can enable Large Language Models (LLMs) to bypass the safeguard and generate harmful content. Existing jailbreaking defense methods have failed to address the fundamental issue that harmful knowledge resides within the model, leading to potential jailbreak risks for LLMs. In this paper, we propose a novel defense method called Eraser, which mainly includes three goals: unlearning harmful knowledge, retaining general knowledge, and maintaining safety alignment. The intuition is that if an LLM forgets the specific knowledge required to answer a harmful question, it will no longer have the ability to answer harmful questions. The training of Erase does not actually require the model's own harmful knowledge, and it can benefit from unlearning general answers related to harmful queries, which means it does not need assistance from the red team. The experimental results show that Eraser can significantly reduce the jailbreaking success rate for various attacks without compromising the general capabilities of the model. Our codes are available at https://github.com/ZeroNLP/Eraser.

The integration of advanced technologies into telecommunication networks complicates troubleshooting, posing challenges for manual error identification in Packet Capture (PCAP) data. This manual approach, requiring substantial resources, becomes impractical at larger scales. Machine learning (ML) methods offer alternatives, but the scarcity of labeled data limits accuracy. In this study, we propose a self-supervised, large language model-based (LLMcap) method for PCAP failure detection. LLMcap leverages language-learning abilities and employs masked language modeling to learn grammar, context, and structure. Tested rigorously on various PCAPs, it demonstrates high accuracy despite the absence of labeled data during training, presenting a promising solution for efficient network analysis. Index Terms: Network troubleshooting, Packet Capture Analysis, Self-Supervised Learning, Large Language Model, Network Quality of Service, Network Performance.

We introduce MIA-Bench, a new benchmark designed to evaluate multimodal large language models (MLLMs) on their ability to strictly adhere to complex instructions. Our benchmark comprises a diverse set of 400 image-prompt pairs, each crafted to challenge the models' compliance with layered instructions in generating accurate responses that satisfy specific requested patterns. Evaluation results from a wide array of state-of-the-art MLLMs reveal significant variations in performance, highlighting areas for improvement in instruction fidelity. Additionally, we create extra training data and explore supervised fine-tuning to enhance the models' ability to strictly follow instructions without compromising performance on other tasks. We hope this benchmark not only serves as a tool for measuring MLLM adherence to instructions, but also guides future developments in MLLM training methods.

Large language models (LLMs) have emerged as powerful tools for tackling complex tasks across diverse domains, but they also raise privacy concerns when fine-tuned on sensitive data due to potential memorization. While differential privacy (DP) offers a promising solution by ensuring models are 'almost indistinguishable' with or without any particular privacy unit, current evaluations on LLMs mostly treat each example (text record) as the privacy unit. This leads to uneven user privacy guarantees when contributions per user vary. We therefore study user-level DP motivated by applications where it necessary to ensure uniform privacy protection across users. We present a systematic evaluation of user-level DP for LLM fine-tuning on natural language generation tasks. Focusing on two mechanisms for achieving user-level DP guarantees, Group Privacy and User-wise DP-SGD, we investigate design choices like data selection strategies and parameter tuning for the best privacy-utility tradeoff.

Hereby, we investigate over 100 hand-crafted and synthetically generated prompts and find that the Memorization in Large Language Models (LLMs) correct PII is extracted in less than 1% of cases. In has recently enjoyed a surge of interest (Hartmann contrast, using the true prefix of the target PII as et al., 2023) ranging from memorization localization a single query yields extraction rates of up to 6%. (Maini et al., 2023), quantification (Carlini Second, we propose PII-Compass, a novel method et al., 2022) to controlling (Ozdayi et al., 2023) and that achieves a substantially higher extraction rate auditing (Zhang et al., 2023a). The major reason than simple adversarial prompts. Our approach is for this is the risk of training data extraction (Carlini based on the intuition that querying the model with et al., 2021; Ishihara, 2023). To assess this risk, a prompt that has a close embedding to the embedding various methods have been proposed in prior work of the target piece of data, i.e., the PII and its (Yu et al., 2023; Zhang et al., 2023b; Panda et al., prefix, should increase the likelihood of extracting 2024; Wang et al., 2024). In this work, we aim to the PII. We do this by prepending the hand-crafted assess the privacy leakage risk of a subclass of training prompt with a true prefix of a different data subject data, namely personal identifiable information than the targeted data subject.

Rapidly increasing model scales coupled with steering methods such as chain-of-thought prompting have led to drastic improvements in language model reasoning. At the same time, models struggle with compositional generalization and are far from human performance on many reasoning-based benchmarks. Leveraging the success of chain-of-thought prompting, and also taking inspiration from context-aware decoding (CAD), we explore input-based contrasting methods to further encourage the type of reasoning induced by chain-of-thought prompting. While work remains to stabilize these results across datasets and models, the improvements we find warrant further investigation into input-based steering methods for context-aware reasoning.