In this paper, we introduce SUTRA, multilingual Large Language Model architecture capable of understanding, reasoning, and generating text in over 50 languages. SUTRA's design uniquely decouples core conceptual understanding from language-specific processing, which facilitates scalable and efficient multilingual alignment and learning. Employing a Mixture of Experts framework both in language and concept processing, SUTRA demonstrates both computational efficiency and responsiveness. Through extensive evaluations, SUTRA is demonstrated to surpass existing models like GPT-3.5, Llama2 by 20-30% on leading Massive Multitask Language Understanding (MMLU) benchmarks for multilingual tasks. SUTRA models are also online LLMs that can use knowledge from the internet to provide hallucination-free, factual and up-to-date responses while retaining their multilingual capabilities. Furthermore, we explore the broader implications of its architecture for the future of multilingual AI, highlighting its potential to democratize access to AI technology globally and to improve the equity and utility of AI in regions with predominantly non-English languages. Our findings suggest that SUTRA not only fills pivotal gaps in multilingual model capabilities but also establishes a new benchmark for operational efficiency and scalability in AI applications.
Large Language Models (LLMs) have made great strides in areas such as language processing and computer vision. Despite the emergence of diverse techniques to improve few-shot learning capacity, current LLMs fall short in handling the languages in biology and chemistry. For example, they are struggling to capture the relationship between molecule structure and pharmacochemical properties. Consequently, the few-shot learning capacity of small-molecule drug modification remains impeded. In this work, we introduced DrugLLM, a LLM tailored for drug design. During the training process, we employed Group-based Molecular Representation (GMR) to represent molecules, arranging them in sequences that reflect modifications aimed at enhancing specific molecular properties. DrugLLM learns how to modify molecules in drug discovery by predicting the next molecule based on past modifications. Extensive computational experiments demonstrate that DrugLLM can generate new molecules with expected properties based on limited examples, presenting a powerful few-shot molecule generation capacity.
Social media aids disaster response but suffers from noise, hindering accurate impact assessment and decision making for resilient cities, which few studies considered. To address the problem, this study proposes the first domain-specific LLM model and an integrated method for rapid earthquake impact assessment. First, a few categories are introduced to classify and filter microblogs considering their relationship to the physical and social impacts of earthquakes, and a dataset comprising 7282 earthquake-related microblogs from twenty earthquakes in different locations is developed as well. Then, with a systematic analysis of various influential factors, QuakeBERT, a domain-specific large language model (LLM), is developed and fine-tuned for accurate classification and filtering of microblogs. Meanwhile, an integrated method integrating public opinion trend analysis, sentiment analysis, and keyword-based physical impact quantification is introduced to assess both the physical and social impacts of earthquakes based on social media texts. Experiments show that data diversity and data volume dominate the performance of QuakeBERT and increase the macro average F1 score by 27%, while the best classification model QuakeBERT outperforms the CNN- or RNN-based models by improving the macro average F1 score from 60.87% to 84.33%. Finally, the proposed approach is applied to assess two earthquakes with the same magnitude and focal depth. Results show that the proposed approach can effectively enhance the impact assessment process by accurate detection of noisy microblogs, which enables effective post-disaster emergency responses to create more resilient cities.
This paper presents the use of Retrieval Augmented Generation (RAG) to improve the feedback generated by Large Language Models for programming tasks. For this purpose, corresponding lecture recordings were transcribed and made available to the Large Language Model GPT-4 as external knowledge source together with timestamps as metainformation by using RAG. The purpose of this is to prevent hallucinations and to enforce the use of the technical terms and phrases from the lecture. In an exercise platform developed to solve programming problems for an introductory programming lecture, students can request feedback on their solutions generated by GPT-4. For this task GPT-4 receives the students' code solution, the compiler output, the result of unit tests and the relevant passages from the lecture notes available through the use of RAG as additional context. The feedback generated by GPT-4 should guide students to solve problems independently and link to the lecture content, using the time stamps of the transcript as meta-information. In this way, the corresponding lecture videos can be viewed immediately at the corresponding positions. For the evaluation, students worked with the tool in a workshop and decided for each feedback whether it should be extended by RAG or not. First results based on a questionnaire and the collected usage data show that the use of RAG can improve feedback generation and is preferred by students in some situations. Due to the slower speed of feedback generation, the benefits are situation dependent.
Child-directed speech (CDS) is a particular type of speech that adults use when addressing young children. Its properties also change as a function of extralinguistic factors, such as age of the child being addressed. Access to large amounts of representative and varied CDS would be useful for child language research, as this would enable controlled computational modeling experiments of infant language acquisition with realistic input in terms of quality and quantity. In this study, we describe an approach to model age-dependent linguistic properties of CDS using a language model (LM) trained on CDS transcripts and ages of the recipient children, as obtained from North American English corpora of the CHILDES database. The created LM can then be used to stochastically generate synthetic CDS transcripts in an age-appropriate manner, thereby scaling beyond the original datasets in size. We compare characteristics of the generated CDS against the real speech addressed at children of different ages, showing that the LM manages to capture age-dependent changes in CDS, except for a slight difference in the effective vocabulary size. As a side product, we also provide a systematic characterization of age-dependent linguistic properties of CDS in CHILDES, illustrating how all measured aspects of the CDS change with children's age.
Large language models (LLMs) are good knowledge bases but struggle to perform equally well for all classes in text classification. This paper investigates the case of class accuracy imbalance in LLMs, where deeply entangled pretraining biases and prompt-specific cues contribute to the imbalance. To overcome the difficulty in bias identification and inaccessibility of retraining, we post-hoc balance class accuracy using only output probabilities. This is enabled by reformulating debiasing as a combinatorial optimization problem. In details, we first motivate a post-hoc bias metric, the Contextual Oddity Bias (COBias), to quantify the over-/under-prediction (a tendency to over-predict some classes while under-predicting others) in LLMs. We then propose the Debiasing as Nonlinear Integer Programming (DNIP) method to reweight LLM output class probabilities towards minimizing COBias and max12 imizing overall accuracy, without being constrained by bias sources or updating LLM parameters. Since the DNIP model contains non-differentiable elements, we use simulated annealing to efficiently solve it. Evaluations on five LLMs across NLP classification benchmarks show that DNIP simultaneously achieves signifi16 cant COBias reduction (61% relative reduction) and accuracy improvement (18% relative increase) under different LLM prompting setups.
Fact-checking is essential due to the explosion of misinformation in the media ecosystem. Although false information exists in every language and country, most research to solve the problem mainly concentrated on huge communities like English and Chinese. Low-resource languages like Vietnamese are necessary to explore corpora and models for fact verification. To bridge this gap, we construct ViWikiFC, the first manual annotated open-domain corpus for Vietnamese Wikipedia Fact Checking more than 20K claims generated by converting evidence sentences extracted from Wikipedia articles. We analyze our corpus through many linguistic aspects, from the new dependency rate, the new n-gram rate, and the new word rate. We conducted various experiments for Vietnamese fact-checking, including evidence retrieval and verdict prediction. BM25 and InfoXLM (Large) achieved the best results in two tasks, with BM25 achieving an accuracy of 88.30% for SUPPORTS, 86.93% for REFUTES, and only 56.67% for the NEI label in the evidence retrieval task, InfoXLM (Large) achieved an F1 score of 86.51%. Furthermore, we also conducted a pipeline approach, which only achieved a strict accuracy of 67.00% when using InfoXLM (Large) and BM25. These results demonstrate that our dataset is challenging for the Vietnamese language model in fact-checking tasks.
While nationality is a pivotal demographic element that enhances the performance of language models, it has received far less scrutiny regarding inherent biases. This study investigates nationality bias in ChatGPT (GPT-3.5), a large language model (LLM) designed for text generation. The research covers 195 countries, 4 temperature settings, and 3 distinct prompt types, generating 4,680 discourses about nationality descriptions in Chinese and English. Automated metrics were used to analyze the nationality bias, and expert annotators alongside ChatGPT itself evaluated the perceived bias. The results show that ChatGPT's generated discourses are predominantly positive, especially compared to its predecessor, GPT-2. However, when prompted with negative inclinations, it occasionally produces negative content. Despite ChatGPT considering its generated text as neutral, it shows consistent self-awareness about nationality bias when subjected to the same pair-wise comparison annotation framework used by human annotators. In conclusion, while ChatGPT's generated texts seem friendly and positive, they reflect the inherent nationality biases in the real world. This bias may vary across different language versions of ChatGPT, indicating diverse cultural perspectives. The study highlights the subtle and pervasive nature of biases within LLMs, emphasizing the need for further scrutiny.
Developing new drugs is laborious and costly, demanding extensive time investment. In this paper, we introduce a de-novo drug design strategy, which harnesses the capabilities of language models to devise targeted drugs for specific proteins. Employing a Reinforcement Learning (RL) framework utilizing Proximal Policy Optimization (PPO), we refine the model to acquire a policy for generating drugs tailored to protein targets. The proposed method integrates a composite reward function, combining considerations of drug-target interaction and molecular validity. Following RL fine-tuning, the proposed method demonstrates promising outcomes, yielding notable improvements in molecular validity, interaction efficacy, and critical chemical properties, achieving 65.37 for Quantitative Estimation of Drug-likeness (QED), 321.55 for Molecular Weight (MW), and 4.47 for Octanol-Water Partition Coefficient (logP), respectively. Furthermore, out of the generated drugs, only 0.041% do not exhibit novelty.
Current medical artificial intelligence systems are often limited to narrow applications, hindering their widespread adoption in clinical practice. To address this limitation, we propose MedVersa, a generalist learner that enables flexible learning and tasking for medical image interpretation. By leveraging a large language model as a learnable orchestrator, MedVersa can learn from both visual and linguistic supervision, support multimodal inputs, and perform real-time task specification. This versatility allows MedVersa to adapt to various clinical scenarios and perform multifaceted medical image analysis. We introduce MedInterp, the largest multimodal dataset to date for medical image interpretation, consisting of over 13 million annotated instances spanning 11 tasks across 3 modalities, to support the development of MedVersa. Our experiments demonstrate that MedVersa achieves state-of-the-art performance in 9 tasks, sometimes outperforming specialist counterparts by over 10%. MedVersa is the first to showcase the viability of multimodal generative medical AI in implementing multimodal outputs, inputs, and dynamic task specification, highlighting its potential as a multifunctional system for comprehensive medical image analysis. This generalist approach to medical image interpretation paves the way for more adaptable and efficient AI-assisted clinical decision-making.
Generative Artificial Intelligence (GAI) is taking the world by storm with its unparalleled content creation ability. Large Language Models (LLMs) are at the forefront of this movement. However, the significant resource demands of LLMs often require cloud hosting, which raises issues regarding privacy, latency, and usage limitations. Although edge intelligence has long been utilized to solve these challenges by enabling real-time AI computation on ubiquitous edge resources close to data sources, most research has focused on traditional AI models and has left a gap in addressing the unique characteristics of LLM inference, such as considerable model size, auto-regressive processes, and self-attention mechanisms. In this paper, we present an edge intelligence optimization problem tailored for LLM inference. Specifically, with the deployment of the batching technique and model quantization on resource-limited edge devices, we formulate an inference model for transformer decoder-based LLMs. Furthermore, our approach aims to maximize the inference throughput via batch scheduling and joint allocation of communication and computation resources, while also considering edge resource constraints and varying user requirements of latency and accuracy. To address this NP-hard problem, we develop an optimal Depth-First Tree-Searching algorithm with online tree-Pruning (DFTSP) that operates within a feasible time complexity. Simulation results indicate that DFTSP surpasses other batching benchmarks in throughput across diverse user settings and quantization techniques, and it reduces time complexity by over 45% compared to the brute-force searching method.
Large Language Models (LLMs) have distinguished themselves with outstanding performance in complex language modeling tasks, yet they come with significant computational and storage challenges. This paper explores the potential of quantization to mitigate these challenges. We systematically study the combined application of three well-known post-training techniques, SmoothQuant, AWQ, and GPTQ, and provide a comprehensive analysis of their interactions and implications for advancing LLM quantization. We enhance the versatility of these methods by enabling quantization to microscaling (MX) formats, extending the applicability of these PTQ algorithms beyond their original fixed-point format targets. We show that combining different PTQ methods enables us to quantize models to 4-bit weights and 8-bit activations using the MXINT format with negligible accuracy loss compared to the uncompressed baseline.
In the pursuit of advancing natural language processing for the Italian language, we introduce a state-of-the-art Large Language Model (LLM) based on the novel Meta LLaMA-3 model: LLaMAntino-3-ANITA-8B-Inst-DPO-ITA. We fine-tuned the original 8B parameters instruction tuned model using the Supervised Fine-tuning (SFT) technique on the English and Italian language datasets in order to improve the original performance. Consequently, a Dynamic Preference Optimization (DPO) process has been used to align preferences, avoid dangerous and inappropriate answers, and limit biases and prejudices. Our model leverages the efficiency of QLoRA to fine-tune the model on a smaller portion of the original model weights and then adapt the model specifically for the Italian linguistic structure, achieving significant improvements in both performance and computational efficiency. Concurrently, DPO is employed to refine the model's output, ensuring that generated content aligns with quality answers. The synergy between SFT, QLoRA's parameter efficiency and DPO's user-centric optimization results in a robust LLM that excels in a variety of tasks, including but not limited to text completion, zero-shot classification, and contextual understanding. The model has been extensively evaluated over standard benchmarks for the Italian and English languages, showing outstanding results. The model is freely available over the HuggingFace hub and, examples of use can be found in our GitHub repository. https://huggingface.co/swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA
Large Language Models (LLMs) enable a new ecosystem with many downstream applications, called LLM applications, with different natural language processing tasks. The functionality and performance of an LLM application highly depend on its system prompt, which instructs the backend LLM on what task to perform. Therefore, an LLM application developer often keeps a system prompt confidential to protect its intellectual property. As a result, a natural attack, called prompt leaking, is to steal the system prompt from an LLM application, which compromises the developer's intellectual property. Existing prompt leaking attacks primarily rely on manually crafted queries, and thus achieve limited effectiveness. In this paper, we design a novel, closed-box prompt leaking attack framework, called PLeak, to optimize an adversarial query such that when the attacker sends it to a target LLM application, its response reveals its own system prompt. We formulate finding such an adversarial query as an optimization problem and solve it with a gradient-based method approximately. Our key idea is to break down the optimization goal by optimizing adversary queries for system prompts incrementally, i.e., starting from the first few tokens of each system prompt step by step until the entire length of the system prompt. We evaluate PLeak in both offline settings and for real-world LLM applications, e.g., those on Poe, a popular platform hosting such applications. Our results show that PLeak can effectively leak system prompts and significantly outperforms not only baselines that manually curate queries but also baselines with optimized queries that are modified and adapted from existing jailbreaking attacks. We responsibly reported the issues to Poe and are still waiting for their response. Our implementation is available at this repository: https://github.com/BHui97/PLeak.
This study explores the innovative use of Large Language Models (LLMs) as analytical tools for interpreting complex financial regulations. The primary objective is to design effective prompts that guide LLMs in distilling verbose and intricate regulatory texts, such as the Basel III capital requirement regulations, into a concise mathematical framework that can be subsequently translated into actionable code. This novel approach aims to streamline the implementation of regulatory mandates within the financial reporting and risk management systems of global banking institutions. A case study was conducted to assess the performance of various LLMs, demonstrating that GPT-4 outperforms other models in processing and collecting necessary information, as well as executing mathematical calculations. The case study utilized numerical simulations with asset holdings -- including fixed income, equities, currency pairs, and commodities -- to demonstrate how LLMs can effectively implement the Basel III capital adequacy requirements. Keywords: Large Language Models, Prompt Engineering, LLMs in Finance, Basel III, Minimum Capital Requirements, LLM Ethics
In today's digital landscape, where cyber attacks have become the norm, the detection of cyber attacks and threats is critically imperative across diverse domains. Our research presents a new empirical framework for cyber threat modeling, adept at parsing and categorizing cyber-related information from news articles, enhancing real-time vigilance for market stakeholders. At the core of this framework is a fine-tuned BERT model, which we call CANAL - Cyber Activity News Alerting Language Model, tailored for cyber categorization using a novel silver labeling approach powered by Random Forest. We benchmark CANAL against larger, costlier LLMs, including GPT-4, LLaMA, and Zephyr, highlighting their zero to few-shot learning in cyber news classification. CANAL demonstrates superior performance by outperforming all other LLM counterparts in both accuracy and cost-effectiveness. Furthermore, we introduce the Cyber Signal Discovery module, a strategic component designed to efficiently detect emerging cyber signals from news articles. Collectively, CANAL and Cyber Signal Discovery module equip our framework to provide a robust and cost-effective solution for businesses that require agile responses to cyber intelligence.
Category fluency is a widely studied cognitive phenomenon, yet two conflicting accounts have been proposed as the underlying retrieval mechanism -- an optimal foraging process deliberately searching through memory (Hills et al., 2012) and a random walk sampling from a semantic network (Abbott et al., 2015). Evidence for both accounts has centered around predicting human patch switches, where both existing models of category fluency produce paradoxically identical results. We begin by peeling back the assumptions made by existing models, namely that each named example only depends on the previous example, by (i) adding an additional bias to model the category transition probability directly and (ii) relying on a large language model to predict based on the entire existing sequence. Then, we present evidence towards resolving the disagreement between each account of foraging by reformulating models as sequence generators. To evaluate, we compare generated category fluency runs to a bank of human-written sequences by proposing a metric based on n-gram overlap. We find category switch predictors do not necessarily produce human-like sequences, in fact the additional biases used by the Hills et al. (2012) model are required to improve generation quality, which are later improved by our category modification. Even generating exclusively with an LLM requires an additional global cue to trigger the patch switching behavior during production. Further tests on only the search process on top of the semantic network highlight the importance of deterministic search to replicate human behavior.
We present the workflow of Online Iterative Reinforcement Learning from Human Feedback (RLHF) in this technical report, which is widely reported to outperform its offline counterpart by a large margin in the recent large language model (LLM) literature. However, existing open-source RLHF projects are still largely confined to the offline learning setting. In this technical report, we aim to fill in this gap and provide a detailed recipe that is easy to reproduce for online iterative RLHF. In particular, since online human feedback is usually infeasible for open-source communities with limited resources, we start by constructing preference models using a diverse set of open-source datasets and use the constructed proxy preference model to approximate human feedback. Then, we discuss the theoretical insights and algorithmic principles behind online iterative RLHF, followed by a detailed practical implementation. Our trained LLM achieves impressive performance on LLM chatbot benchmarks, including AlpacaEval-2, Arena-Hard, and MT-Bench, as well as other academic benchmarks such as HumanEval and TruthfulQA. We have shown that supervised fine-tuning (SFT) and iterative RLHF can obtain state-of-the-art performance with fully open-source datasets. Further, we have made our models, curated datasets, and comprehensive step-by-step code guidebooks publicly available. Please refer to https://github.com/RLHFlow/RLHF-Reward-Modeling and https://github.com/RLHFlow/Online-RLHF for more detailed information.
Linking (aligning) biomedical concepts across diverse data sources enables various integrative analyses, but it is challenging due to the discrepancies in concept naming conventions. Various strategies have been developed to overcome this challenge, such as those based on string-matching rules, manually crafted thesauri, and machine learning models. However, these methods are constrained by limited prior biomedical knowledge and can hardly generalize beyond the limited amounts of rules, thesauri, or training samples. Recently, large language models (LLMs) have exhibited impressive results in diverse biomedical NLP tasks due to their unprecedentedly rich prior knowledge and strong zero-shot prediction abilities. However, LLMs suffer from issues including high costs, limited context length, and unreliable predictions. In this research, we propose PromptLink, a novel biomedical concept linking framework that leverages LLMs. It first employs a biomedical-specialized pre-trained language model to generate candidate concepts that can fit in the LLM context windows. Then it utilizes an LLM to link concepts through two-stage prompts, where the first-stage prompt aims to elicit the biomedical prior knowledge from the LLM for the concept linking task and the second-stage prompt enforces the LLM to reflect on its own predictions to further enhance their reliability. Empirical results on the concept linking task between two EHR datasets and an external biomedical KG demonstrate the effectiveness of PromptLink. Furthermore, PromptLink is a generic framework without reliance on additional prior knowledge, context, or training data, making it well-suited for concept linking across various types of data sources. The source code is available at https://github.com/constantjxyz/PromptLink.
The rapid advancement of Large Language Models (LLMs) has improved text understanding and generation but poses challenges in computational resources. This study proposes a curriculum learning-inspired, data-centric training strategy that begins with simpler tasks and progresses to more complex ones, using criteria such as prompt length, attention scores, and loss values to structure the training data. Experiments with Mistral-7B (Jiang et al., 2023) and Gemma-7B (Team et al., 2024) models demonstrate that curriculum learning slightly improves performance compared to traditional random data shuffling. Notably, we observed that sorting data based on our proposed attention criteria generally led to better performance. This approach offers a sustainable method to enhance LLM performance without increasing model size or dataset volume, addressing scalability challenges in LLM training.