Multimodal Continual Instruction Tuning (MCIT) enables Multimodal Large Language Models (MLLMs) to meet continuously emerging requirements without expensive retraining. MCIT faces two major obstacles: catastrophic forgetting (where old knowledge is forgotten) and negative forward transfer (where the performance of future tasks is degraded). Although existing methods have greatly alleviated catastrophic forgetting, they still suffer from negative forward transfer. By performing singular value decomposition (SVD) on input embeddings, we discover a large discrepancy in different input embeddings. The discrepancy results in the model learning irrelevant information for old and pre-trained tasks, which leads to catastrophic forgetting and negative forward transfer. To address these issues, we propose Fwd-Prompt, a prompt-based method projecting prompt gradient to the residual space to minimize the interference between tasks and to the pre-trained subspace for reusing pre-trained knowledge. Our experiments demonstrate that Fwd-Prompt achieves state-of-the-art performance while updating fewer parameters and requiring no old samples. Our research sheds light on the potential of continuously adapting MLLMs to new tasks under the instruction tuning paradigm and encourages future studies to explore MCIT. The code will soon be publicly available.

The progress introduced by pre-trained language models and their fine-tuning has resulted in significant improvements in most downstream NLP tasks. The unsupervised training of a language model combined with further target task fine-tuning has become the standard QA fine-tuning procedure. In this work, we demonstrate that this strategy is sub-optimal for fine-tuning QA models, especially under a low QA annotation budget, which is a usual setting in practice due to the extractive QA labeling cost. We draw our conclusions by conducting an exhaustive analysis of the performance of the alternatives of the sequential fine-tuning strategy on different QA datasets. Based on the experiments performed, we observed that the best strategy to fine-tune the QA model in low-budget settings is taking a pre-trained language model (PLM) and then fine-tuning PLM with a dataset composed of the target dataset and SQuAD dataset. With zero extra annotation effort, the best strategy outperforms the standard strategy by 2.28% to 6.48%. Our experiments provide one of the first investigations on how to best fine-tune a QA system under a low budget and are therefore of the utmost practical interest to the QA practitioners.

In the contemporary information era, significantly accelerated by the advent of Large-scale Language Models, the proliferation of scientific literature is reaching unprecedented levels. Researchers urgently require efficient tools for reading and summarizing academic papers, uncovering significant scientific literature, and employing diverse interpretative methodologies. To address this burgeoning demand, the role of automated scientific literature interpretation systems has become paramount. However, prevailing models, both commercial and open-source, confront notable challenges: they often overlook multimodal data, grapple with summarizing over-length texts, and lack diverse user interfaces. In response, we introduce an open-source multi-modal automated academic paper interpretation system (MMAPIS) with three-step process stages, incorporating LLMs to augment its functionality. Our system first employs the hybrid modality preprocessing and alignment module to extract plain text, and tables or figures from documents separately. It then aligns this information based on the section names they belong to, ensuring that data with identical section names are categorized under the same section. Following this, we introduce a hierarchical discourse-aware summarization method. It utilizes the extracted section names to divide the article into shorter text segments, facilitating specific summarizations both within and between sections via LLMs with specific prompts. Finally, we have designed four types of diversified user interfaces, including paper recommendation, multimodal Q\&A, audio broadcasting, and interpretation blog, which can be widely applied across various scenarios. Our qualitative and quantitative evaluations underscore the system's superiority, especially in scientific summarization, where it outperforms solutions relying solely on GPT-4.

With the growing popularity of dialogue agents based on large language models (LLMs), urgent attention has been drawn to finding ways to ensure their behaviour is ethical and appropriate. These are largely interpreted in terms of the 'HHH' criteria: making outputs more helpful and honest, and avoiding harmful (biased, toxic, or inaccurate) statements. Whilst this semantic focus is useful from the perspective of viewing LLM agents as mere mediums for information, it fails to account for pragmatic factors that can make the same utterance seem more or less offensive or tactless in different social situations. We propose an approach to ethics that is more centred on relational and situational factors, exploring what it means for a system, as a social actor, to treat an individual respectfully in a (series of) interaction(s). Our work anticipates a set of largely unexplored risks at the level of situated interaction, and offers practical suggestions to help LLM technologies behave as 'good' social actors and treat people respectfully.

Generative artificial intelligence (GenAI) models are a type of deep learning neural network model capable of learning from large datasets and generating new content from a given context. They represent a significant leap in the ability of the artificial intelligence (AI) field to not just interpret data but also to create something new, including text, images, videos, code, and sound [2]. Large language models (LLMs) are a type of GenAI model designed to understand and generate natural language. The market for LLMs is estimated to reach 40.8 billion USD by 2029, up from 10.5 billion USD in 2022 [10]. Organizations are currently competing to develop the most sophisticated LLM capable of mimicking human-like conversations and tasks. This has led to the creation of models such as OpenAI's ChatGPT,

Large language models (LLMs) have revolutionized many areas (e.g. natural language processing, software engineering, etc.) by achieving state-of-the-art performance on extensive downstream tasks. Aiming to achieve robust and general artificial intelligence, there has been a surge of interest in investigating the reasoning ability of the LLMs. Whereas the textual and numerical reasoning benchmarks adopted by previous works are rather shallow and simple, it is hard to conclude that the LLMs possess strong reasoning ability by merely achieving positive results on these benchmarks. Recent efforts have demonstrated that the LLMs are poor at solving sequential decision-making problems that require common-sense planning by evaluating their performance on the reinforcement learning benchmarks. In this work, we conduct an in-depth assessment of several state-of-the-art LLMs' reasoning ability based on the inductive logic programming (ILP) benchmark, which is broadly recognized as a representative and challenging measurement for evaluating logic program induction/synthesis systems as it requires inducing strict cause-effect logic to achieve robust deduction on independent and identically distributed (IID) and out-of-distribution (OOD) test samples. Our evaluations illustrate that compared with the neural program induction systems which are much smaller in model size, the state-of-the-art LLMs are much poorer in terms of reasoning ability by achieving much lower performance and generalization using either natural language prompting or truth-value matrix prompting.

In our research, we pioneer a novel approach to evaluate the effectiveness of jailbreak attacks on Large Language Models (LLMs), such as GPT-4 and LLaMa2, diverging from traditional robustness-focused binary evaluations. Our study introduces two distinct evaluation frameworks: a coarse-grained evaluation and a fine-grained evaluation. Each framework, using a scoring range from 0 to 1, offers a unique perspective, enabling a more comprehensive and nuanced evaluation of attack effectiveness and empowering attackers to refine their attack prompts with greater understanding. Furthermore, we have developed a comprehensive ground truth dataset specifically tailored for jailbreak tasks. This dataset not only serves as a crucial benchmark for our current study but also establishes a foundational resource for future research, enabling consistent and comparative analyses in this evolving field. Upon meticulous comparison with traditional evaluation methods, we discovered that our evaluation aligns with the baseline's trend while offering a more profound and detailed assessment. We believe that by accurately evaluating the effectiveness of attack prompts in the Jailbreak task, our work lays a solid foundation for assessing a wider array of similar or even more complex tasks in the realm of prompt injection, potentially revolutionizing this field.

The evolution of Neural Machine Translation (NMT) has been significantly influenced by six core challenges (Koehn and Knowles, 2017), which have acted as benchmarks for progress in this field. This study revisits these challenges, offering insights into their ongoing relevance in the context of advanced Large Language Models (LLMs): domain mismatch, amount of parallel data, rare word prediction, translation of long sentences, attention model as word alignment, and sub-optimal beam search. Our empirical findings indicate that LLMs effectively lessen the reliance on parallel data for major languages in the pretraining phase. Additionally, the LLM-based translation system significantly enhances the translation of long sentences that contain approximately 80 words and shows the capability to translate documents of up to 512 words. However, despite these significant improvements, the challenges of domain mismatch and prediction of rare words persist. While the challenges of word alignment and beam search, specifically associated with NMT, may not apply to LLMs, we identify three new challenges for LLMs in translation tasks: inference efficiency, translation of low-resource languages in the pretraining phase, and human-aligned evaluation. The datasets and models are released at https://github.com/pangjh3/LLM4MT.

Addition is perhaps one of the simplest arithmetic tasks one can think of and is usually performed using the carrying over algorithm. This algorithm consists of two tasks: adding digits in the same position and carrying over a one whenever necessary. We study how transformer models implement this algorithm and how the two aforementioned tasks are allocated to different parts of the network. We first focus on two-layer encoder-only models and show that the carrying over algorithm is implemented in a modular fashion. The first layer is mostly responsible for adding digits in the same position. The second layer first decides, in the attention, which positions need a carried one or not, and then performs the carrying of the one in the final MLP. We provide a simple way of precisely identifying which neurons are responsible for that task. This implementation of the carrying over algorithm occurs across a range of hyperparameters for two as well as three-layer models. For small decoder-only models, we observe the same implementation and provide suggestive evidence for its existence in three 7B large language models.

Instruction-tuned large language models (LLMs) excel at many tasks, and will even provide explanations for their behavior. Since these models are directly accessible to the public, there is a risk that convincing and wrong explanations can lead to unsupported confidence in LLMs. Therefore, interpretability-faithfulness of self-explanations is an important consideration for AI Safety. Assessing the interpretability-faithfulness of these explanations, termed self-explanations, is challenging as the models are too complex for humans to annotate what is a correct explanation. To address this, we propose employing self-consistency checks as a measure of faithfulness. For example, if an LLM says a set of words is important for making a prediction, then it should not be able to make the same prediction without these words. While self-consistency checks are a common approach to faithfulness, they have not previously been applied to LLM's self-explanations. We apply self-consistency checks to three types of self-explanations: counterfactuals, importance measures, and redactions. Our work demonstrate that faithfulness is both task and model dependent, e.g., for sentiment classification, counterfactual explanations are more faithful for Llama2, importance measures for Mistral, and redaction for Falcon 40B. Finally, our findings are robust to prompt-variations.