Authors: Hansle Gwon, Imjin Ahn, Young-Hak Kim, Sanghyun Park, Tae Joon Jun
Abstract: Recent advancements in Large Language Models (LLMs) have been remarkable, with new models consistently surpassing their predecessors. These advancements are underpinned by extensive research on various training mechanisms. Among these, Preference Optimization has played a significant role in improving the performance of LLMs by incorporating human preferences into the training process. However, constructing preference optimization datasets is challenging and the optimization process is highly sensitive to the dataset quality. In this study, we propose a novel approach to augment Preference Optimization datasets. Additionally, we introduce a Multi-response-based Preference Optimization training method that enables the simultaneous learning of multiple responses.
Authors: Minxue Niu, Yara El-Tawil, Amrit Romana, Emily Mower Provost
Abstract: Modern affective computing systems rely heavily on datasets with human-annotated emotion labels, for training and evaluation. However, human annotations are expensive to obtain, sensitive to study design, and difficult to quality control, because of the subjective nature of emotions. Meanwhile, Large Language Models (LLMs) have shown remarkable performance on many Natural Language Understanding tasks, emerging as a promising tool for text annotation. In this work, we analyze the complexities of emotion annotation in the context of LLMs, focusing on GPT-4 as a leading model. In our experiments, GPT-4 achieves high ratings in a human evaluation study, painting a more positive picture than previous work, in which human labels served as the only ground truth. On the other hand, we observe differences between human and GPT-4 emotion perception, underscoring the importance of human input in annotation studies. To harness GPT-4's strength while preserving human perspective, we explore two ways of integrating GPT-4 into emotion annotation pipelines, showing its potential to flag low-quality labels, reduce the workload of human annotators, and improve downstream model learning performance and efficiency. Together, our findings highlight opportunities for new emotion labeling practices and suggest the use of LLMs as a promising tool to aid human annotation.
Authors: Sagi Shaier
Abstract: This study examines whether large language models (LLMs), such as ChatGPT, specifically the latest GPT-4o-mini, exhibit sensitivity to repeated prompts and whether repeating a question can improve response accuracy. We hypothesize that reiterating a question within a single prompt might enhance the model's focus on key elements of the query. To test this, we evaluate ChatGPT's performance on a large sample of two reading comprehension datasets under both open-book and closed-book settings, varying the repetition of each question to 1, 3, or 5 times per prompt. Our findings indicate that the model does not demonstrate sensitivity to repeated questions, highlighting its robustness and consistency in this context.
Authors: Quinten McNamara, Miguel \'Angel del R\'io Fern\'andez, Nishchal Bhandari, Martin Ratajczak, Danny Chen, Corey Miller, Mig\"uel Jett\'e
Abstract: Word error rate (WER) as a metric has a variety of limitations that have plagued the field of speech recognition. Evaluation datasets suffer from varying style, formality, and inherent ambiguity of the transcription task. In this work, we attempt to mitigate some of these differences by performing style-agnostic evaluation of ASR systems using multiple references transcribed under opposing style parameters. As a result, we find that existing WER reports are likely significantly over-estimating the number of contentful errors made by state-of-the-art ASR systems. In addition, we have found our multireference method to be a useful mechanism for comparing the quality of ASR models that differ in the stylistic makeup of their training data and target task.
Authors: Eric Bigelow, Ari Holtzman, Hidenori Tanaka, Tomer Ullman
Abstract: Estimating uncertainty in Large Language Models (LLMs) is important for properly evaluating LLMs, and ensuring safety for users. However, prior approaches to uncertainty estimation focus on the final answer in generated text, ignoring intermediate steps that might dramatically impact the outcome. We hypothesize that there exist key forking tokens, such that re-sampling the system at those specific tokens, but not others, leads to very different outcomes. To test this empirically, we develop a novel approach to representing uncertainty dynamics across individual tokens of text generation, and applying statistical models to test our hypothesis. Our approach is highly flexible: it can be applied to any dataset and any LLM, without fine tuning or accessing model weights. We use our method to analyze LLM responses on 7 different tasks across 4 domains, spanning a wide range of typical use cases. We find many examples of forking tokens, including surprising ones such as punctuation marks, suggesting that LLMs are often just a single token away from saying something very different.
Authors: Tianyi Li, Erenay Dayanik, Shubhi Tyagi, Andrea Pierleoni
Abstract: In this paper, we present HalluCana, a canary lookahead to detect and correct factuality hallucinations of Large Language Models (LLMs) in long-form generation. HalluCana detects and intervenes as soon as traces of hallucination emerge, during and even before generation. To support timely detection, we exploit the internal factuality representation in the LLM hidden space, where we investigate various proxies to the LLMs' factuality self-assessment, and discuss its relation to the models' context familiarity from their pre-training. On biography generation, our method improves generation quality by up to 2.5x, while consuming over 6 times less compute.
Authors: Chung-En Sun, Tuomas Oikarinen, Berk Ustun, Tsui-Wei Weng
Abstract: We introduce the Concept Bottleneck Large Language Model (CB-LLM), a pioneering approach to creating inherently interpretable Large Language Models (LLMs). Unlike traditional black-box LLMs that rely on post-hoc interpretation methods with limited neuron function insights, CB-LLM sets a new standard with its built-in interpretability, scalability, and ability to provide clear, accurate explanations. We investigate two essential tasks in the NLP domain: text classification and text generation. In text classification, CB-LLM narrows the performance gap with traditional black-box models and provides clear interpretability. In text generation, we show how interpretable neurons in CB-LLM can be used for concept detection and steering text generation. Our CB-LLMs enable greater interaction between humans and LLMs across a variety of tasks -- a feature notably absent in existing LLMs. Our code is available at https://github.com/Trustworthy-ML-Lab/CB-LLMs.
Authors: Shengmin Piao, Sanghyun Park
Abstract: Large Language Models exhibit impressive reasoning capabilities across diverse tasks, motivating efforts to distill these capabilities into smaller models through generated reasoning data. However, direct training on such synthesized reasoning data may lead to superficial imitation of reasoning process, rather than fostering a genuine integration of reasoning capabilities with underlying knowledge. To address this, we propose TinyThinker, a framework introducing two novel approaches. First, we introduce a three-stage process that incrementally guides the student model through the reasoning process, progressively refining knowledge from coarse to fine granularity. Second, we develop a two-phase training framework comprising an initial reasoning acquisition phase followed by a self-reflection phase utilizing self-generated data. Experiments on commonsense reasoning benchmarks demonstrate that TinyThinker achieves superior performance compared to baselines. Ablation studies further validate the effectiveness of each component in our framework. TinyThinker is extendable to other knowledge-intensive reasoning tasks, offering an alternative strategy for developing effective reasoning capabilities in smaller language models. Codes are available at https://github.com/shengminp/TinyThinker
Authors: Ao Li, Longwei Xu, Chen Ling, Jinghui Zhang, Pengwei Wang
Abstract: Sentiment analysis and emotion recognition are crucial for applications such as human-computer interaction and depression detection. Traditional unimodal methods often fail to capture the complexity of emotional expressions due to conflicting signals from different modalities. Current Multimodal Large Language Models (MLLMs) also face challenges in detecting subtle facial expressions and addressing a wide range of emotion-related tasks. To tackle these issues, we propose M2SE, a Multistage Multitask Sentiment and Emotion Instruction Tuning Strategy for general-purpose MLLMs. It employs a combined approach to train models on tasks such as multimodal sentiment analysis, emotion recognition, facial expression recognition, emotion reason inference, and emotion cause-pair extraction. We also introduce the Emotion Multitask dataset (EMT), a custom dataset that supports these five tasks. Our model, Emotion Universe (EmoVerse), is built on a basic MLLM framework without modifications, yet it achieves substantial improvements across these tasks when trained with the M2SE strategy. Extensive experiments demonstrate that EmoVerse outperforms existing methods, achieving state-of-the-art results in sentiment and emotion tasks. These results highlight the effectiveness of M2SE in enhancing multimodal emotion perception. The dataset and code are available at https://github.com/xiaoyaoxinyi/M2SE.
Authors: Ashutosh Bajpai, Tanmoy Chakraborty
Abstract: The unwavering disparity in labeled resources between resource-rich languages and those considered low-resource remains a significant impediment for Large Language Models (LLMs). Recent strides in cross-lingual in-context learning (X-ICL), mainly through semantically aligned examples retrieved from multilingual pre-trained transformers, have shown promise in mitigating this issue. However, our investigation reveals that LLMs intrinsically reward in-language semantically aligned cross-lingual instances over direct cross-lingual semantic alignments, with a pronounced disparity in handling time-sensitive queries in the X-ICL setup. Such queries demand sound temporal reasoning ability from LLMs, yet the advancements have predominantly focused on English. This study aims to bridge this gap by improving temporal reasoning capabilities in low-resource languages. To this end, we introduce mTEMPREASON a temporal reasoning dataset aimed at the varied degrees of low-resource languages and propose Cross-Lingual Time-Sensitive Semantic Alignment (CLiTSSA), a novel method to improve temporal reasoning in these contexts. To facilitate this, we construct an extension of mTEMPREASON comprising pairs of parallel cross-language temporal queries along with their anticipated in-language semantic similarity scores. Our empirical evidence underscores the superior performance of CLiTSSA compared to established baselines across three languages - Romanian, German, and French, encompassing three temporal tasks and including a diverse set of four contemporaneous LLMs. This marks a significant step forward in addressing resource disparity in the context of temporal reasoning across languages.
Authors: Nathana\"el Carraz Rakotonirina, Corentin Kervadec, Francesca Franzon, Marco Baroni
Abstract: It has been widely observed that language models (LMs) respond in predictable ways to algorithmically generated prompts that are seemingly unintelligible. This is both a sign that we lack a full understanding of how LMs work, and a practical challenge, because opaqueness can be exploited for harmful uses of LMs, such as jailbreaking. We present the first thorough analysis of opaque machine-generated prompts, or autoprompts, pertaining to 3 LMs of different sizes and families. We find that machine-generated prompts are characterized by a last token that is often intelligible and strongly affects the generation. A small but consistent proportion of the previous tokens are fillers that probably appear in the prompt as a by-product of the fact that the optimization process fixes the number of tokens. The remaining tokens tend to have at least a loose semantic relation with the generation, although they do not engage in well-formed syntactic relations with it. We find moreover that some of the ablations we applied to machine-generated prompts can also be applied to natural language sequences, leading to similar behavior, suggesting that autoprompts are a direct consequence of the way in which LMs process linguistic inputs in general.
Authors: Anmol Guragain, Nadika Poudel, Rajesh Piryani, Bishesh Khanal
Abstract: This paper explores hate speech detection in Devanagari-scripted languages, focusing on Hindi and Nepali, for Subtask B of the CHIPSAL@COLING 2025 Shared Task. Using a range of transformer-based models such as XLM-RoBERTa, MURIL, and IndicBERT, we examine their effectiveness in navigating the nuanced boundary between hate speech and free expression. Our best performing model, implemented as ensemble of multilingual BERT models achieve Recall of 0.7762 (Rank 3/31 in terms of recall) and F1 score of 0.6914 (Rank 17/31). To address class imbalance, we used backtranslation for data augmentation, and cosine similarity to preserve label consistency after augmentation. This work emphasizes the need for hate speech detection in Devanagari-scripted languages and presents a foundation for further research.
Authors: Thibault Prouteau, Nicolas Dugu\'e, Simon Guillot
Abstract: Modelling information from complex systems such as humans social interaction or words co-occurrences in our languages can help to understand how these systems are organized and function. Such systems can be modelled by networks, and network theory provides a useful set of methods to analyze them. Among these methods, graph embedding is a powerful tool to summarize the interactions and topology of a network in a vectorized feature space. When used in input of machine learning algorithms, embedding vectors help with common graph problems such as link prediction, graph matching, etc. Word embedding has the goal of representing the sense of words, extracting it from large text corpora. Despite differences in the structure of information in input of embedding algorithms, many graph embedding approaches are adapted and inspired from methods in NLP. Limits of these methods are observed in both domains. Most of these methods require long and resource greedy training. Another downside to most methods is that they are black-box, from which understanding how the information is structured is rather complex. Interpretability of a model allows understanding how the vector space is structured without the need for external information, and thus can be audited more easily. With both these limitations in mind, we propose a novel framework to efficiently embed network vertices in an interpretable vector space. Our Lower Dimension Bipartite Framework (LDBGF) leverages the bipartite projection of a network using cliques to reduce dimensionality. Along with LDBGF, we introduce two implementations of this framework that rely on communities instead of cliques: SINr-NR and SINr-MF. We show that SINr-MF can perform well on classical graphs and SINr-NR can produce high-quality graph and word embeddings that are interpretable and stable across runs.
Authors: Phan Phuong Mai Chau, Souhail Bakkali, Antoine Doucet
Abstract: Abstractive summarization has made significant strides in condensing and rephrasing large volumes of text into coherent summaries. However, summarizing administrative documents presents unique challenges due to domain-specific terminology, OCR-generated errors, and the scarcity of annotated datasets for model fine-tuning. Existing models often struggle to adapt to the intricate structure and specialized content of such documents. To address these limitations, we introduce DocSum, a domain-adaptive abstractive summarization framework tailored for administrative documents. Leveraging pre-training on OCR-transcribed text and fine-tuning with an innovative integration of question-answer pairs, DocSum enhances summary accuracy and relevance. This approach tackles the complexities inherent in administrative content, ensuring outputs that align with real-world business needs. To evaluate its capabilities, we define a novel downstream task setting-Document Abstractive Summarization-which reflects the practical requirements of business and organizational settings. Comprehensive experiments demonstrate DocSum's effectiveness in producing high-quality summaries, showcasing its potential to improve decision-making and operational workflows across the public and private sectors.
Authors: Jiacheng Hu, Runyuan Bao, Yang Lin, Hanchao Zhang, Yanlin Xiang
Abstract: This study evaluated the effect of BioBERT in medical text processing for the task of medical named entity recognition. Through comparative experiments with models such as BERT, ClinicalBERT, SciBERT, and BlueBERT, the results showed that BioBERT achieved the best performance in both precision and F1 score, verifying its applicability and superiority in the medical field. BioBERT enhances its ability to understand professional terms and complex medical texts through pre-training on biomedical data, providing a powerful tool for medical information extraction and clinical decision support. The study also explored the privacy and compliance challenges of BioBERT when processing medical data, and proposed future research directions for combining other medical-specific models to improve generalization and robustness. With the development of deep learning technology, the potential of BioBERT in application fields such as intelligent medicine, personalized treatment, and disease prediction will be further expanded. Future research can focus on the real-time and interpretability of the model to promote its widespread application in the medical field.
Authors: Pascal Tilli, Ngoc Thang Vu
Abstract: Explainable artificial intelligence (XAI) aims to make machine learning models more transparent. While many approaches focus on generating explanations post-hoc, interpretable approaches, which generate the explanations intrinsically alongside the predictions, are relatively rare. In this work, we integrate different discrete subset sampling methods into a graph-based visual question answering system to compare their effectiveness in generating interpretable explanatory subgraphs intrinsically. We evaluate the methods on the GQA dataset and show that the integrated methods effectively mitigate the performance trade-off between interpretability and answer accuracy, while also achieving strong co-occurrences between answer and question tokens. Furthermore, we conduct a human evaluation to assess the interpretability of the generated subgraphs using a comparative setting with the extended Bradley-Terry model, showing that the answer and question token co-occurrence metrics strongly correlate with human preferences. Our source code is publicly available.
Authors: Marta R. Costa-juss\`a, Pierre Andrews, Mariano Coria Meglioli, Joy Chen, Joe Chuang, David Dale, Christophe Ropers, Alexandre Mourachko, Eduardo S\'anchez, Holger Schwenk, Tuan Tran, Arina Turkatenko, Carleigh Wood
Abstract: This paper presents the Long Context and Form Output (LCFO) benchmark, a novel evaluation framework for assessing gradual summarization and summary expansion capabilities across diverse domains. LCFO consists of long input documents (5k words average length), each of which comes with three summaries of different lengths (20%, 10%, and 5% of the input text), as well as approximately 15 questions and answers (QA) related to the input content. Notably, LCFO also provides alignments between specific QA pairs and corresponding summaries in 7 domains. The primary motivation behind providing summaries of different lengths is to establish a controllable framework for generating long texts from shorter inputs, i.e. summary expansion. To establish an evaluation metric framework for summarization and summary expansion, we provide human evaluation scores for human-generated outputs, as well as results from various state-of-the-art large language models (LLMs). GPT-4o-mini achieves best human scores among automatic systems in both summarization and summary expansion tasks (~ +10% and +20%, respectively). It even surpasses human output quality in the case of short summaries (~ +7%). Overall automatic metrics achieve low correlations with human evaluation scores (~ 0.4) but moderate correlation on specific evaluation aspects such as fluency and attribution (~ 0.6). The LCFO benchmark offers a standardized platform for evaluating summarization and summary expansion performance, as well as corresponding automatic metrics, thereby providing an important evaluation framework to advance generative AI.
Authors: Marta R. Costa-juss\`a, Bokai Yu, Pierre Andrews, Belen Alastruey, Necati Cihan Camgoz, Joe Chuang, Jean Maillard, Christophe Ropers, Arina Turkantenko, Carleigh Wood
Abstract: We introduce the first highly multilingual speech and American Sign Language (ASL) comprehension dataset by extending BELEBELE. Our dataset covers 74 spoken languages at the intersection of BELEBELE and FLEURS, and one sign language (ASL). We evaluate 2M-BELEBELE dataset for both 5-shot and zero-shot settings and across languages, the speech comprehension accuracy is ~ 8% average lower compared to reading comprehension.
Authors: Marta R. Costa-juss\`a, Joy Chen, Ifeoluwanimi Adebara, Joe Chuang, Christophe Ropers, Eduardo S\'anchez
Abstract: The purpose of this work is to share an English-Yor\`ub\'a evaluation dataset for open-book reading comprehension and text generation to assess the performance of models both in a high- and a low- resource language. The dataset contains 358 questions and answers on 338 English documents and 208 Yor\`ub\'a documents. The average document length is ~ 10k words for English and 430 words for Yor\`ub\'a. Experiments show a consistent disparity in performance between the two languages, with Yor\`ub\'a falling behind English for automatic metrics even if documents are much shorter for this language. For a small set of documents with comparable length, performance of Yor\`ub\'a drops by x2.5 times. When analyzing performance by length, we observe that Yor\`ub\'a decreases performance dramatically for documents that reach 1500 words while English performance is barely affected at that length. Our dataset opens the door to showcasing if English LLM reading comprehension capabilities extend to Yor\`ub\'a, which for the evaluated LLMs is not the case.
Authors: Anindita Mondal, Rangavajjala Sankara Bharadwaj, Jhansi Mallela, Anil Kumar Vuppala, Chiranjeevi Yarra
Abstract: Automatic detection of prominence at the word and syllable-levels is critical for building computer-assisted language learning systems. It has been shown that prosody embeddings learned by the current state-of-the-art (SOTA) text-to-speech (TTS) systems could generate word- and syllable-level prominence in the synthesized speech as natural as in native speech. To understand the effectiveness of prosody embeddings from TTS for prominence detection under nonnative context, a comparative analysis is conducted on the embeddings extracted from native and non-native speech considering the prominence-related embeddings: duration, energy, and pitch from a SOTA TTS named FastSpeech2. These embeddings are extracted under two conditions considering: 1) only text, 2) both speech and text. For the first condition, the embeddings are extracted directly from the TTS inference mode, whereas for the second condition, we propose to extract from the TTS under training mode. Experiments are conducted on native speech corpus: Tatoeba, and non-native speech corpus: ISLE. For experimentation, word-level prominence locations are manually annotated for both corpora. The highest relative improvement on word \& syllable-level prominence detection accuracies with the TTS embeddings are found to be 13.7% & 5.9% and 16.2% & 6.9% compared to those with the heuristic-based features and self-supervised Wav2Vec-2.0 representations, respectively.
Authors: Minh Le, Tien Ngoc Luu, An Nguyen The, Thanh-Thien Le, Trang Nguyen, Thanh Tung Nguyen, Linh Ngo Van, Thien Huu Nguyen
Abstract: To address catastrophic forgetting in Continual Relation Extraction (CRE), many current approaches rely on memory buffers to rehearse previously learned knowledge while acquiring new tasks. Recently, prompt-based methods have emerged as potent alternatives to rehearsal-based strategies, demonstrating strong empirical performance. However, upon analyzing existing prompt-based approaches for CRE, we identified several critical limitations, such as inaccurate prompt selection, inadequate mechanisms for mitigating forgetting in shared parameters, and suboptimal handling of cross-task and within-task variances. To overcome these challenges, we draw inspiration from the relationship between prefix-tuning and mixture of experts, proposing a novel approach that employs a prompt pool for each task, capturing variations within each task while enhancing cross-task variances. Furthermore, we incorporate a generative model to consolidate prior knowledge within shared parameters, eliminating the need for explicit data storage. Extensive experiments validate the efficacy of our approach, demonstrating superior performance over state-of-the-art prompt-based and rehearsal-free methods in continual relation extraction.
Authors: Nishat Raihan, Christian Newman, Marcos Zampieri
Abstract: Large language models (LLMs) have demonstrated remarkable capabilities across various NLP tasks and have recently expanded their impact to coding tasks, bridging the gap between natural languages (NL) and programming languages (PL). This taxonomy-based survey provides a comprehensive analysis of LLMs in the NL-PL domain, investigating how these models are utilized in coding tasks and examining their methodologies, architectures, and training processes. We propose a taxonomy-based framework that categorizes relevant concepts, providing a unified classification system to facilitate a deeper understanding of this rapidly evolving field. This survey offers insights into the current state and future directions of LLMs in coding tasks, including their applications and limitations.
Authors: Haotong Zhang
Abstract: We carry out a series of experiments to test large language models' multi-hop reasoning ability from three aspects: selecting and combining external knowledge, dealing with non-sequential reasoning tasks and generalising to data samples with larger numbers of hops. We test the GPT-3.5 model on four reasoning benchmarks with Chain-of-Thought prompting (and its variations). Our results reveal that despite the amazing performance achieved by large language models on various reasoning tasks, models still suffer from severe drawbacks which shows a large gap with humans.
Authors: Nikolay Banar, Ehsan Lotfi, Walter Daelemans
Abstract: Zero-shot evaluation of information retrieval (IR) models is often performed using BEIR; a large and heterogeneous benchmark composed of multiple datasets, covering different retrieval tasks across various domains. Although BEIR has become a standard benchmark for the zero-shot setup, its exclusively English content reduces its utility for underrepresented languages in IR, including Dutch. To address this limitation and encourage the development of Dutch IR models, we introduce BEIR-NL by automatically translating the publicly accessible BEIR datasets into Dutch. Using BEIR-NL, we evaluated a wide range of multilingual dense ranking and reranking models, as well as the lexical BM25 method. Our experiments show that BM25 remains a competitive baseline, and is only outperformed by the larger dense models trained for retrieval. When combined with reranking models, BM25 achieves performance on par with the best dense ranking models. In addition, we explored the impact of translation on the data by back-translating a selection of datasets to English, and observed a performance drop for both dense and lexical methods, indicating the limitations of translation for creating benchmarks. BEIR-NL is publicly available on the Hugging Face hub.
Authors: Sultan Alrashed
Abstract: We present SmolTulu-1.7b-Instruct, referenced in this report as SmolTulu-DPO-1130, an instruction-tuned language model that adapts AllenAI's Tulu 3 post-training pipeline to enhance Huggingface's SmolLM2-1.7B base model. Through comprehensive empirical analysis using a 135M parameter model, we demonstrate that the relationship between learning rate and batch size significantly impacts model performance in a task-dependent manner. Our findings reveal a clear split: reasoning tasks like ARC and GSM8K benefit from higher learning rate to batch size ratios, while pattern recognition tasks such as HellaSwag and IFEval show optimal performance with lower ratios. These insights informed the development of SmolTulu, which achieves state-of-the-art performance among sub-2B parameter models on instruction following, scoring 67.7% on IFEval ($\Delta$11%), and mathematical reasoning with 51.6% on GSM8K ($\Delta$3.4%), with an alternate version achieving scoring 57.1% on ARC ($\Delta5.4%$). We release our model, training recipes, and ablation studies to facilitate further research in efficient model alignment, demonstrating that careful adaptation of optimization dynamics can help bridge the capability gap between small and large language models.
Authors: Shubham Kumar Nigam, Balaramamahanthi Deepak Patnaik, Shivam Mishra, Noel Shallum, Kripabandhu Ghosh, Arnab Bhattacharya
Abstract: The integration of artificial intelligence (AI) in legal judgment prediction (LJP) has the potential to transform the legal landscape, particularly in jurisdictions like India, where a significant backlog of cases burdens the legal system. This paper introduces NyayaAnumana, the largest and most diverse corpus of Indian legal cases compiled for LJP, encompassing a total of 7,02,945 preprocessed cases. NyayaAnumana, which combines the words "Nyay" (judgment) and "Anuman" (prediction or inference) respectively for most major Indian languages, includes a wide range of cases from the Supreme Court, High Courts, Tribunal Courts, District Courts, and Daily Orders and, thus, provides unparalleled diversity and coverage. Our dataset surpasses existing datasets like PredEx and ILDC, offering a comprehensive foundation for advanced AI research in the legal domain. In addition to the dataset, we present INLegalLlama, a domain-specific generative large language model (LLM) tailored to the intricacies of the Indian legal system. It is developed through a two-phase training approach over a base LLaMa model. First, Indian legal documents are injected using continual pretraining. Second, task-specific supervised finetuning is done. This method allows the model to achieve a deeper understanding of legal contexts. Our experiments demonstrate that incorporating diverse court data significantly boosts model accuracy, achieving approximately 90% F1-score in prediction tasks. INLegalLlama not only improves prediction accuracy but also offers comprehensible explanations, addressing the need for explainability in AI-assisted legal decisions.
Authors: Jing Ye, Lu Xiang, Yaping Zhang, Chengqing Zong
Abstract: Large Language Models (LLMs) have demonstrated promising potential in providing empathetic support during interactions. However, their responses often become verbose or overly formulaic, failing to adequately address the diverse emotional support needs of real-world scenarios. To tackle this challenge, we propose an innovative strategy-enhanced role-playing framework, designed to simulate authentic emotional support conversations. Specifically, our approach unfolds in two steps: (1) Strategy-Enhanced Role-Playing Interactions, which involve three pivotal roles -- Seeker, Strategy Counselor, and Supporter -- engaging in diverse scenarios to emulate real-world interactions and promote a broader range of dialogues; and (2) Emotional Support Agent Training, achieved through fine-tuning LLMs using our specially constructed dataset. Within this framework, we develop the \textbf{ServeForEmo} dataset, comprising an extensive collection of 3.7K+ multi-turn dialogues and 62.8K+ utterances. We further present \textbf{SweetieChat}, an emotional support agent capable of handling diverse open-domain scenarios. Extensive experiments and human evaluations confirm the framework's effectiveness in enhancing emotional support, highlighting its unique ability to provide more nuanced and tailored assistance.
Authors: Wessel Poelman, Miryam de Lhoneux
Abstract: Multilingual natural language processing is getting increased attention, with numerous models, benchmarks, and methods being released for many languages. English is often used in multilingual evaluation to prompt language models (LMs), mainly to overcome the lack of instruction tuning data in other languages. In this position paper, we lay out two roles of English in multilingual LM evaluations: as an interface and as a natural language. We argue that these roles have different goals: task performance versus language understanding. This discrepancy is highlighted with examples from datasets and evaluation setups. Numerous works explicitly use English as an interface to boost task performance. We recommend to move away from this imprecise method and instead focus on furthering language understanding.
Authors: Kaiyuan Chen, Jin Wang, Xuejie Zhang
Abstract: Small language models (SLMs) are more efficient, cost-effective, and customizable than large language models (LLMs), though they often underperform in specific areas like reasoning. Past methods for enhancing SLMs' reasoning, such as supervised fine-tuning and distillation, often depend on costly external signals, resulting in SLMs being overly confident with limited supervision signals, thus limiting their abilities. Therefore, this study enables SLMs to learn to reason from self-iterative feedback. By combining odds ratio preference optimization (ORPO), we fine-tune and align SLMs using positive and negative signals generated by themselves. Additionally, we introduce process supervision for rewards in preference alignment by sampling-based inference simulation and process reward models. Compared to Supervised Fine-Tuning (SFT), our method improves the performance of Gemma-2B by 12.43 (Acc) on GSM8K and 3.95 (Pass@1) on MBPP. Furthermore, the proposed method also demonstrated superior out-of-domain generalization capabilities on MMLU_Math and HumanEval.
Authors: Jiayuan Ma, Hongbin Na, Zimu Wang, Yining Hua, Yue Liu, Wei Wang, Ling Chen
Abstract: Mental manipulation severely undermines mental wellness by covertly and negatively distorting decision-making. While there is an increasing interest in mental health care within the natural language processing community, progress in tackling manipulation remains limited due to the complexity of detecting subtle, covert tactics in conversations. In this paper, we propose Intent-Aware Prompting (IAP), a novel approach for detecting mental manipulations using large language models (LLMs), providing a deeper understanding of manipulative tactics by capturing the underlying intents of participants. Experimental results on the MentalManip dataset demonstrate superior effectiveness of IAP against other advanced prompting strategies. Notably, our approach substantially reduces false negatives, helping detect more instances of mental manipulation with minimal misjudgment of positive cases. The code of this paper is available at https://github.com/Anton-Jiayuan-MA/Manip-IAP.
Authors: Xiao Luo, Sean O'Connell, Shamima Mithun
Abstract: This paper provides an in-depth evaluation of three state-of-the-art Large Language Models (LLMs) for personalized career mentoring in the computing field, using three distinct student profiles that consider gender, race, and professional levels. We evaluated the performance of GPT-4, LLaMA 3, and Palm 2 using a zero-shot learning approach without human intervention. A quantitative evaluation was conducted through a custom natural language processing analytics pipeline to highlight the uniqueness of the responses and to identify words reflecting each student's profile, including race, gender, or professional level. The analysis of frequently used words in the responses indicates that GPT-4 offers more personalized mentoring compared to the other two LLMs. Additionally, a qualitative evaluation was performed to see if human experts reached similar conclusions. The analysis of survey responses shows that GPT-4 outperformed the other two LLMs in delivering more accurate and useful mentoring while addressing specific challenges with encouragement languages. Our work establishes a foundation for developing personalized mentoring tools based on LLMs, incorporating human mentors in the process to deliver a more impactful and tailored mentoring experience.
Authors: Guochao Jiang, Ziqin Luo, Chengwei Hu, Zepeng Ding, Deqing Yang
Abstract: Many previous models of named entity recognition (NER) suffer from the problem of Out-of-Entity (OOE), i.e., the tokens in the entity mentions of the test samples have not appeared in the training samples, which hinders the achievement of satisfactory performance. To improve OOE-NER performance, in this paper, we propose a new framework, namely S+NER, which fully leverages sentence-level information. Our S+NER achieves better OOE-NER performance mainly due to the following two particular designs. 1) It first exploits the pre-trained language model's capability of understanding the target entity's sentence-level context with a template set. 2) Then, it refines the sentence-level representation based on the positive and negative templates, through a contrastive learning strategy and template pooling method, to obtain better NER results. Our extensive experiments on five benchmark datasets have demonstrated that, our S+NER outperforms some state-of-the-art OOE-NER models.
Authors: Huiyuan Lai, Esther Ploeger, Rik van Noord, Antonio Toral
Abstract: Neural machine translation (NMT) systems amplify lexical biases present in their training data, leading to artificially impoverished language in output translations. These language-level characteristics render automatic translations different from text originally written in a language and human translations, which hinders their usefulness in for example creating evaluation datasets. Attempts to increase naturalness in NMT can fall short in terms of content preservation, where increased lexical diversity comes at the cost of translation accuracy. Inspired by the reinforcement learning from human feedback framework, we introduce a novel method that rewards both naturalness and content preservation. We experiment with multiple perspectives to produce more natural translations, aiming at reducing machine and human translationese. We evaluate our method on English-to-Dutch literary translation, and find that our best model produces translations that are lexically richer and exhibit more properties of human-written language, without loss in translation accuracy.
Authors: Hai-Yen Thi Nguyen, Cam-Van Thi Nguyen
Abstract: Comparative reviews are pivotal in understanding consumer preferences and influencing purchasing decisions. Comparative Quintuple Extraction (COQE) aims to identify five key components in text: the target entity, compared entities, compared aspects, opinions on these aspects, and polarity. Extracting precise comparative information from product reviews is challenging due to nuanced language and sequential task errors in traditional methods. To mitigate these problems, we propose MTP-COQE, an end-to-end model designed for COQE. Leveraging multi-perspective prompt-based learning, MTP-COQE effectively guides the generative model in comparative opinion mining tasks. Evaluation on the Camera-COQE (English) and VCOM (Vietnamese) datasets demonstrates MTP-COQE's efficacy in automating COQE, achieving superior performance with a 1.41% higher F1 score than the previous baseline models on the English dataset. Additionally, we designed a strategy to limit the generative model's creativity to ensure the output meets expectations. We also performed data augmentation to address data imbalance and to prevent the model from becoming biased towards dominant samples.
Authors: Pengyue Jia, Derong Xu, Xiaopeng Li, Zhaocheng Du, Xiangyang Li, Xiangyu Zhao, Yichao Wang, Yuhao Wang, Huifeng Guo, Ruiming Tang
Abstract: The reranker and generator are two critical components in the Retrieval-Augmented Generation (i.e., RAG) pipeline, responsible for ranking relevant documents and generating responses. However, due to differences in pre-training data and objectives, there is an inevitable gap between the documents ranked as relevant by the reranker and those required by the generator to support answering the query. To address this gap, we propose RADIO, a novel and practical preference alignment framework with RAtionale DIstillatiOn. Specifically, We first propose a rationale extraction method that leverages the reasoning capabilities of Large Language Models (LLMs) to extract the rationales necessary for answering the query. Subsequently, a rationale-based alignment process is designed to rerank the documents based on the extracted rationales, and fine-tune the reranker to align the preferences. We conduct extensive experiments on two tasks across three datasets to demonstrate the effectiveness of our approach compared to baseline methods. Our code is released online to ease reproduction.
Authors: Lefteris Loukas, Nikolaos Smyrnioudis, Chrysa Dikonomaki, Spyros Barbakos, Anastasios Toumazatos, John Koutsikakis, Manolis Kyriakakis, Mary Georgiou, Stavros Vassos, John Pavlopoulos, Ion Androutsopoulos
Abstract: We present GR-NLP-TOOLKIT, an open-source natural language processing (NLP) toolkit developed specifically for modern Greek. The toolkit provides state-of-the-art performance in five core NLP tasks, namely part-of-speech tagging, morphological tagging, dependency parsing, named entity recognition, and Greeklishto-Greek transliteration. The toolkit is based on pre-trained Transformers, it is freely available, and can be easily installed in Python (pip install gr-nlp-toolkit). It is also accessible through a demonstration platform on HuggingFace, along with a publicly available API for non-commercial use. We discuss the functionality provided for each task, the underlying methods, experiments against comparable open-source toolkits, and future possible enhancements. The toolkit is available at: https://github.com/nlpaueb/gr-nlp-toolkit
Authors: Yingxin Li, Ye Li, Yuan Meng, Xinzhu Ma, Zihan Geng, Shutao Xia, Zhi Wang
Abstract: As large language models (LLMs) continue to advance, the demand for higher quality and faster processing of long contexts across various applications is growing. KV cache is widely adopted as it stores previously generated key and value tokens, effectively reducing redundant computations during inference. However, as memory overhead becomes a significant concern, efficient compression of KV cache has gained increasing attention. Most existing methods perform compression from two perspectives: identifying important tokens and designing compression strategies. However, these approaches often produce biased distributions of important tokens due to the influence of accumulated attention scores or positional encoding. Furthermore, they overlook the sparsity and redundancy across different heads, which leads to difficulties in preserving the most effective information at the head level. To this end, we propose EMS to overcome these limitations, while achieving better KV cache compression under extreme compression ratios. Specifically, we introduce a Global-Local score that combines accumulated attention scores from both global and local KV tokens to better identify the token importance. For the compression strategy, we design an adaptive and unified Evict-then-Merge framework that accounts for the sparsity and redundancy of KV tokens across different heads. Additionally, we implement the head-wise parallel compression through a zero-class mechanism to enhance efficiency. Extensive experiments demonstrate our SOTA performance even under extreme compression ratios. EMS consistently achieves the lowest perplexity, improves scores by over 1.28 points across four LLMs on LongBench under a 256 cache budget, and preserves 95% retrieval accuracy with a cache budget less than 2% of the context length in the Needle-in-a-Haystack task.
Authors: Andor Diera, Lukas Galke, Fabian Karl, Ansgar Scherp
Abstract: Continual learning remains challenging across various natural language understanding tasks. When models are updated with new training data, they risk catastrophic forgetting of prior knowledge. In the present work, we introduce a discrete key-value bottleneck for encoder-only language models, allowing for efficient continual learning by requiring only localized updates. Inspired by the success of a discrete key-value bottleneck in vision, we address new and NLP-specific challenges. We experiment with different bottleneck architectures to find the most suitable variants regarding language, and present a generic discrete key initialization technique for NLP that is task independent. We evaluate the discrete key-value bottleneck in four continual learning NLP scenarios and demonstrate that it alleviates catastrophic forgetting. We showcase that it offers competitive performance to other popular continual learning methods, with lower computational costs.
Authors: Quynh-Mai Thi Nguyen, Lan-Nhi Thi Nguyen, Cam-Van Thi Nguyen
Abstract: The objective of multimodal intent recognition (MIR) is to leverage various modalities-such as text, video, and audio-to detect user intentions, which is crucial for understanding human language and context in dialogue systems. Despite advances in this field, two main challenges persist: (1) effectively extracting and utilizing semantic information from robust textual features; (2) aligning and fusing non-verbal modalities with verbal ones effectively. This paper proposes a Text Enhancement with CommOnsense Knowledge Extractor (TECO) to address these challenges. We begin by extracting relations from both generated and retrieved knowledge to enrich the contextual information in the text modality. Subsequently, we align and integrate visual and acoustic representations with these enhanced text features to form a cohesive multimodal representation. Our experimental results show substantial improvements over existing baseline methods.
Authors: Xiaodong Cui, A F M Saif, Songtao Lu, Lisha Chen, Tianyi Chen, Brian Kingsbury, George Saon
Abstract: In this paper, we propose a bilevel joint unsupervised and supervised training (BL-JUST) framework for automatic speech recognition. Compared to the conventional pre-training and fine-tuning strategy which is a disconnected two-stage process, BL-JUST tries to optimize an acoustic model such that it simultaneously minimizes both the unsupervised and supervised loss functions. Because BL-JUST seeks matched local optima of both loss functions, acoustic representations learned by the acoustic model strike a good balance between being generic and task-specific. We solve the BL-JUST problem using penalty-based bilevel gradient descent and evaluate the trained deep neural network acoustic models on various datasets with a variety of architectures and loss functions. We show that BL-JUST can outperform the widely-used pre-training and fine-tuning strategy and some other popular semi-supervised techniques.
Authors: Iman Munire Bilal, Zheng Fang, Miguel Arana-Catania, Felix-Anselm van Lier, Juliana Outes Velarde, Harry Bregazzi, Eleanor Carter, Mara Airoldi, Rob Procter
Abstract: As academic literature proliferates, traditional review methods are increasingly challenged by the sheer volume and diversity of available research. This article presents a study that aims to address these challenges by enhancing the efficiency and scope of systematic reviews in the social sciences through advanced machine learning (ML) and natural language processing (NLP) tools. In particular, we focus on automating stages within the systematic reviewing process that are time-intensive and repetitive for human annotators and which lend themselves to immediate scalability through tools such as information retrieval and summarisation guided by expert advice. The article concludes with a summary of lessons learnt regarding the integrated approach towards systematic reviews and future directions for improvement, including explainability.
Authors: Hang Zhao, Qile P. Chen, Yijing Barry Zhang, Gang Yang
Abstract: Both encoder-only models (e.g., BERT, RoBERTa) and large language models (LLMs, e.g., Llama3) have been widely used for text classification tasks. However, there is a lack of systematic studies comparing the performance of encoder-based models and LLMs in text classification, particularly when fine-tuning is involved. This study employed a diverse range of models and methods, varying in size and architecture, and including both fine-tuned and pre-trained approaches. We first assessed the performances of these LLMs on the 20 Newsgroups (20NG) and MASSIVE datasets, comparing them to encoder-only RoBERTa models. Additionally, we explored the multi-task capabilities of both model types by combining multiple classification tasks, including intent detection and slot-filling, into a single model using data from both datasets. Our results indicate that fully fine-tuned Llama3-70B models outperform RoBERTa-large and other decoder LLMs across various classification tasks and datasets. Moreover, the consolidated multi-task fine-tuned LLMs matched the performance of dual-model setups in both tasks across both datasets. Overall, our study provides a comprehensive benchmark of encoder-only and LLM models on text classification tasks and demonstrates a method to combine two or more fully fine-tuned decoder LLMs for reduced latency and equivalent performance.
Authors: Ivan Arias-Arias, Maria Jose Dominguez Vazquez, Carlos Valcarcel Riveiro
Abstract: By means of pilot experiments for the language pair German and Galician, this paper examines the concept of efficiency and intelligence in lexicography and artificial intelligence, AI. The aim of the experiments is to gain empirically and statistically based insights into the lexicographical text type,dictionary article, in the responses of ChatGPT 3.5, as well as into the lexicographical data on which this chatbot was trained. Both quantitative and qualitative methods are used for this purpose. The analysis is based on the evaluation of the outputs of several sessions with the same prompt in ChatGPT 3.5. On the one hand, the algorithmic performance of intelligent systems is evaluated in comparison with data from lexicographical works. On the other hand, the ChatGPT data supplied is analysed using specific text passages of the aforementioned lexicographical text type. The results of this study not only help to evaluate the efficiency of this chatbot regarding the creation of dictionary articles, but also to delve deeper into the concept of intelligence, the thought processes and the actions to be carried out in both disciplines.
Authors: Jiahui Li, Yongchang Hao, Haoyu Xu, Xing Wang, Yu Hong
Abstract: Despite the advancements in training Large Language Models (LLMs) with alignment techniques to enhance the safety of generated content, these models remain susceptible to jailbreak, an adversarial attack method that exposes security vulnerabilities in LLMs. Notably, the Greedy Coordinate Gradient (GCG) method has demonstrated the ability to automatically generate adversarial suffixes that jailbreak state-of-the-art LLMs. However, the optimization process involved in GCG is highly time-consuming, rendering the jailbreaking pipeline inefficient. In this paper, we investigate the process of GCG and identify an issue of Indirect Effect, the key bottleneck of the GCG optimization. To this end, we propose the Model Attack Gradient Index GCG (MAGIC), that addresses the Indirect Effect by exploiting the gradient information of the suffix tokens, thereby accelerating the procedure by having less computation and fewer iterations. Our experiments on AdvBench show that MAGIC achieves up to a 1.5x speedup, while maintaining Attack Success Rates (ASR) on par or even higher than other baselines. Our MAGIC achieved an ASR of 74% on the Llama-2 and an ASR of 54% when conducting transfer attacks on GPT-3.5. Code is available at https://github.com/jiah-li/magic.
Authors: Yutao Sun, Hangbo Bao, Wenhui Wang, Zhiliang Peng, Li Dong, Shaohan Huang, Jianyong Wang, Furu Wei
Abstract: Multimodal generative models require a unified approach to handle both discrete data (e.g., text and code) and continuous data (e.g., image, audio, video). In this work, we propose Latent Language Modeling (LatentLM), which seamlessly integrates continuous and discrete data using causal Transformers. Specifically, we employ a variational autoencoder (VAE) to represent continuous data as latent vectors and introduce next-token diffusion for autoregressive generation of these vectors. Additionally, we develop $\sigma$-VAE to address the challenges of variance collapse, which is crucial for autoregressive modeling. Extensive experiments demonstrate the effectiveness of LatentLM across various modalities. In image generation, LatentLM surpasses Diffusion Transformers in both performance and scalability. When integrated into multimodal large language models, LatentLM provides a general-purpose interface that unifies multimodal generation and understanding. Experimental results show that LatentLM achieves favorable performance compared to Transfusion and vector quantized models in the setting of scaling up training tokens. In text-to-speech synthesis, LatentLM outperforms the state-of-the-art VALL-E 2 model in speaker similarity and robustness, while requiring 10x fewer decoding steps. The results establish LatentLM as a highly effective and scalable approach to advance large multimodal models.
Authors: Khalil Mrini, Hanlin Lu, Linjie Yang, Weilin Huang, Heng Wang
Abstract: Text-to-image generation has advanced rapidly, yet aligning complex textual prompts with generated visuals remains challenging, especially with intricate object relationships and fine-grained details. This paper introduces Fast Prompt Alignment (FPA), a prompt optimization framework that leverages a one-pass approach, enhancing text-to-image alignment efficiency without the iterative overhead typical of current methods like OPT2I. FPA uses large language models (LLMs) for single-iteration prompt paraphrasing, followed by fine-tuning or in-context learning with optimized prompts to enable real-time inference, reducing computational demands while preserving alignment fidelity. Extensive evaluations on the COCO Captions and PartiPrompts datasets demonstrate that FPA achieves competitive text-image alignment scores at a fraction of the processing time, as validated through both automated metrics (TIFA, VQA) and human evaluation. A human study with expert annotators further reveals a strong correlation between human alignment judgments and automated scores, underscoring the robustness of FPA's improvements. The proposed method showcases a scalable, efficient alternative to iterative prompt optimization, enabling broader applicability in real-time, high-demand settings. The codebase is provided to facilitate further research: https://github.com/tiktok/fast_prompt_alignment
Authors: Jiali Chen, Xusen Hei, Yuqi Xue, Yuancheng Wei, Jiayuan Xie, Yi Cai, Qing Li
Abstract: Large multimodal models (LMMs) have shown remarkable performance in the visual commonsense reasoning (VCR) task, which aims to answer a multiple-choice question based on visual commonsense within an image. However, the ability of LMMs to correct potential visual commonsense errors in the distractor upon their occurrence is yet under-explored. Drawing inspiration from how a human teacher crafts challenging distractors to test students' comprehension of the concepts or skills and assists them in identifying and correcting errors toward the answer, we are the pioneering research for LMMs to simulate this error correction process. To this end, we employ GPT-4 as a ``teacher'' to collect the explainable feedback dataset VCR-DF for error correction, which serves as a benchmark to evaluate the ability of LMMs to identify misconceptions and clarify reasons behind the error in VCR distractors toward final answers. In addition, we propose an LMM-based Pedagogical Expert Instructed Feedback Generation (PEIFG) model to incorporate the learnable expert prompts and multimodal instruction as guidance for feedback generation. Experimental results show that our PEIFG significantly outperforms existing LMMs. We believe that our benchmark provides a new direction for evaluating the capabilities of LMMs.
Authors: Diederik Aerts, Jonito Aerts Argu\"elles, Lester Beltran, Massimiliano Sassoli de Bianchi, Sandro Sozzo
Abstract: We present a theoretical and empirical investigation of the statistical behaviour of the words in a text produced by human language. To this aim, we analyse the word distribution of various texts of Italian language selected from a specific literary corpus. We firstly generalise a theoretical framework elaborated by ourselves to identify 'quantum mechanical statistics' in large-size texts. Then, we show that, in all analysed texts, words distribute according to 'Bose--Einstein statistics' and show significant deviations from 'Maxwell--Boltzmann statistics'. Next, we introduce an effect of 'word randomization' which instead indicates that the difference between the two statistical models is not as pronounced as in the original cases. These results confirm the empirical patterns obtained in texts of English language and strongly indicate that identical words tend to 'clump together' as a consequence of their meaning, which can be explained as an effect of 'quantum entanglement' produced through a phenomenon of 'contextual updating'. More, word randomization can be seen as the linguistic-conceptual equivalent of an increase of temperature which destroys 'coherence' and makes classical statistics prevail over quantum statistics. Some insights into the origin of quantum statistics in physics are finally provided.
Authors: Anya Belz, Craig Thomson
Abstract: This paper presents version 3.0 of the Human Evaluation Datasheet (HEDS). This update is the result of our experience using HEDS in the context of numerous recent human evaluation experiments, including reproduction studies, and of feedback received. Our main overall goal was to improve clarity, and to enable users to complete the datasheet more consistently and comparably. The HEDS 3.0 package consists of the digital data sheet, documentation, and code for exporting completed data sheets as latex files, all available from the HEDS GitHub.
Authors: Davide Nunes, Luis Antunes
Abstract: One goal of Artificial Intelligence is to learn meaningful representations for natural language expressions, but what this entails is not always clear. A variety of new linguistic behaviours present themselves embodied as computers, enhanced humans, and collectives with various kinds of integration and communication. But to measure and understand the behaviours generated by such systems, we must clarify the language we use to talk about them. Computational models are often confused with the phenomena they try to model and shallow metaphors are used as justifications for (or to hype) the success of computational techniques on many tasks related to natural language; thus implying their progress toward human-level machine intelligence without ever clarifying what that means. This paper discusses the challenges in the specification of "machines of meaning", machines capable of acquiring meaningful semantics from natural language in order to achieve their goals. We characterize "meaning" in a computational setting, while highlighting the need for detachment from anthropocentrism in the study of the behaviour of machines of meaning. The pressing need to analyse AI risks and ethics requires a proper measurement of its capabilities which cannot be productively studied and explained while using ambiguous language. We propose a view of "meaning" to facilitate the discourse around approaches such as neural language models and help broaden the research perspectives for technology that facilitates dialogues between humans and machines.
Authors: Hang Gao, Chenhao Zhang, Fengge Wu, Junsuo Zhao, Changwen Zheng, Huaping Liu
Abstract: Graph representation learning methods are highly effective in handling complex non-Euclidean data by capturing intricate relationships and features within graph structures. However, traditional methods face challenges when dealing with heterogeneous graphs that contain various types of nodes and edges due to the diverse sources and complex nature of the data. Existing Heterogeneous Graph Neural Networks (HGNNs) have shown promising results but require prior knowledge of node and edge types and unified node feature formats, which limits their applicability. Recent advancements in graph representation learning using Large Language Models (LLMs) offer new solutions by integrating LLMs' data processing capabilities, enabling the alignment of various graph representations. Nevertheless, these methods often overlook heterogeneous graph data and require extensive preprocessing. To address these limitations, we propose a novel method that leverages the strengths of both LLM and GNN, allowing for the processing of graph data with any format and type of nodes and edges without the need for type information or special preprocessing. Our method employs LLM to automatically summarize and classify different data formats and types, aligns node features, and uses a specialized GNN for targeted learning, thus obtaining effective graph representations for downstream tasks. Theoretical analysis and experimental validation have demonstrated the effectiveness of our method.
Authors: Panlong Wu, Kangshuo Li, Junbao Nan, Fangxin Wang
Abstract: Large Language Models (LLMs) have revolutionized intelligent services by enabling logical reasoning, tool use, and interaction with external systems as agents. The advancement of LLMs is frequently hindered by the scarcity of high-quality data, much of which is inherently sensitive. Federated learning (FL) offers a potential solution by facilitating the collaborative training of distributed LLMs while safeguarding private data. However, FL frameworks face significant bandwidth and computational demands, along with challenges from heterogeneous data distributions. The emerging in-context learning capability of LLMs offers a promising approach by aggregating natural language rather than bulky model parameters. Yet, this method risks privacy leakage, as it necessitates the collection and presentation of data samples from various clients during aggregation. In this paper, we propose a novel privacy-preserving Federated In-Context LLM Agent Learning (FICAL) algorithm, which to our best knowledge for the first work unleashes the power of in-context learning to train diverse LLM agents through FL. In our design, knowledge compendiums generated by a novel LLM-enhanced Knowledge Compendiums Generation (KCG) module are transmitted between clients and the server instead of model parameters in previous FL methods. Apart from that, an incredible Retrieval Augmented Generation (RAG) based Tool Learning and Utilizing (TLU) module is designed and we incorporate the aggregated global knowledge compendium as a teacher to teach LLM agents the usage of tools. We conducted extensive experiments and the results show that FICAL has competitive performance compared to other SOTA baselines with a significant communication cost decrease of $\mathbf{3.33\times10^5}$ times.
Authors: Fuqiang Liu, Sicong Jiang, Luis Miranda-Moreno, Seongjin Choi, Lijun Sun
Abstract: Large Language Models (LLMs) have recently demonstrated significant potential in the field of time series forecasting, offering impressive capabilities in handling complex temporal data. However, their robustness and reliability in real-world applications remain under-explored, particularly concerning their susceptibility to adversarial attacks. In this paper, we introduce a targeted adversarial attack framework for LLM-based time series forecasting. By employing both gradient-free and black-box optimization methods, we generate minimal yet highly effective perturbations that significantly degrade the forecasting accuracy across multiple datasets and LLM architectures. Our experiments, which include models like TimeGPT and LLM-Time with GPT-3.5, GPT-4, LLaMa, and Mistral, show that adversarial attacks lead to much more severe performance degradation than random noise, and demonstrate the broad effectiveness of our attacks across different LLMs. The results underscore the critical vulnerabilities of LLMs in time series forecasting, highlighting the need for robust defense mechanisms to ensure their reliable deployment in practical applications.
Authors: Hee-Seon Kim, Minbeom Kim, Changick Kim
Abstract: Large Vision-Language Models (VLMs) have demonstrated remarkable performance across multimodal tasks by integrating vision encoders with large language models (LLMs). However, these models remain vulnerable to adversarial attacks. Among such attacks, Universal Adversarial Perturbations (UAPs) are especially powerful, as a single optimized perturbation can mislead the model across various input images. In this work, we introduce a novel UAP specifically designed for VLMs: the Doubly-Universal Adversarial Perturbation (Doubly-UAP), capable of universally deceiving VLMs across both image and text inputs. To successfully disrupt the vision encoder's fundamental process, we analyze the core components of the attention mechanism. After identifying value vectors in the middle-to-late layers as the most vulnerable, we optimize Doubly-UAP in a label-free manner with a frozen model. Despite being developed as a black-box to the LLM, Doubly-UAP achieves high attack success rates on VLMs, consistently outperforming baseline methods across vision-language tasks. Extensive ablation studies and analyses further demonstrate the robustness of Doubly-UAP and provide insights into how it influences internal attention mechanisms.
Authors: Jiayun Luo, Mir Rayat Imtiaz Hossain, Boyang Li, Leonid Sigal
Abstract: Vision-Language Models (VLMs) achieved strong performance on a variety of tasks (e.g., image-text retrieval, visual question answering). However, most VLMs rely on coarse-grained image-caption pairs for alignment, relying on data volume to resolve ambiguities and ground linguistic concepts in images. The richer semantic and syntactic structure within text is largely overlooked. To address this, we propose HIerarchically STructured Learning (HIST) that enhances VLM training without any additional supervision, by hierarchically decomposing captions into the constituent Subject, Noun Phrases, and Composite Phrases. Entailment between these constituent components allows us to formulate additional regularization constraints on the VLM attention maps. Specifically, we introduce two novel loss functions: (1) Subject Loss, which aligns image content with the subject of corresponding phrase, acting as an entailment of standard contrastive/matching losses at the Phrase level; (2) Addition Loss, to balance attention across multiple objects. HIST is general, and can be applied to any VLM for which attention between vision and language can be computed; we illustrate its efficacy on BLIP and ALBEF. HIST outperforms baseline VLMs, achieving up to +9.8% improvement in visual grounding, +6.3% in multi-object referring segmentation, +1.1% in image-text retrieval, and +0.2% in visual question answering, underscoring the value of structuring learning in VLMs.
Authors: Sri Harsha Dumpala, David Arps, Sageev Oore, Laura Kallmeyer, Hassan Sajjad
Abstract: Vision-language models (VLMs), serve as foundation models for multi-modal applications such as image captioning and text-to-image generation. Recent studies have highlighted limitations in VLM text encoders, particularly in areas like compositionality and semantic understanding, though the underlying reasons for these limitations remain unclear. In this work, we aim to address this gap by analyzing the syntactic information, one of the fundamental linguistic properties, encoded by the text encoders of VLMs. We perform a thorough analysis comparing VLMs with different objective functions, parameter size and training data size, and with uni-modal language models (ULMs) in their ability to encode syntactic knowledge. Our findings suggest that ULM text encoders acquire syntactic information more effectively than those in VLMs. The syntactic information learned by VLM text encoders is shaped primarily by the pre-training objective, which plays a more crucial role than other factors such as model architecture, model size, or the volume of pre-training data. Models exhibit different layer-wise trends where CLIP performance dropped across layers while for other models, middle layers are rich in encoding syntactic knowledge.
Authors: Haowei Lou, Helen Paik, Wen Hu, Lina Yao
Abstract: Recent advancements in text-to-speech (TTS) systems, such as FastSpeech and StyleSpeech, have significantly improved speech generation quality. However, these models often rely on duration generated by external tools like the Montreal Forced Aligner, which can be time-consuming and lack flexibility. The importance of accurate duration is often underestimated, despite their crucial role in achieving natural prosody and intelligibility. To address these limitations, we propose a novel Aligner-Guided Training Paradigm that prioritizes accurate duration labelling by training an aligner before the TTS model. This approach reduces dependence on external tools and enhances alignment accuracy. We further explore the impact of different acoustic features, including Mel-Spectrograms, MFCCs, and latent features, on TTS model performance. Our experimental results show that aligner-guided duration labelling can achieve up to a 16\% improvement in word error rate and significantly enhance phoneme and tone alignment. These findings highlight the effectiveness of our approach in optimizing TTS systems for more natural and intelligible speech generation.
Authors: Haowei Lou, Helen Paik, Pari Delir Haghighi, Wen Hu, Lina Yao
Abstract: Diffusion-based Generative AI gains significant attention for its superior performance over other generative techniques like Generative Adversarial Networks and Variational Autoencoders. While it has achieved notable advancements in fields such as computer vision and natural language processing, their application in speech generation remains under-explored. Mainstream Text-to-Speech systems primarily map outputs to Mel-Spectrograms in the spectral space, leading to high computational loads due to the sparsity of MelSpecs. To address these limitations, we propose LatentSpeech, a novel TTS generation approach utilizing latent diffusion models. By using latent embeddings as the intermediate representation, LatentSpeech reduces the target dimension to 5% of what is required for MelSpecs, simplifying the processing for the TTS encoder and vocoder and enabling efficient high-quality speech generation. This study marks the first integration of latent diffusion models in TTS, enhancing the accuracy and naturalness of generated speech. Experimental results on benchmark datasets demonstrate that LatentSpeech achieves a 25% improvement in Word Error Rate and a 24% improvement in Mel Cepstral Distortion compared to existing models, with further improvements rising to 49.5% and 26%, respectively, with additional training data. These findings highlight the potential of LatentSpeech to advance the state-of-the-art in TTS technology
Authors: Quang-Hung Le, Long Hoang Dang, Ngan Le, Truyen Tran, Thao Minh Le
Abstract: Existing Large Vision-Language Models (LVLMs) excel at matching concepts across multi-modal inputs but struggle with compositional concepts and high-level relationships between entities. This paper introduces Progressive multi-granular Vision-Language alignments (PromViL), a novel framework to enhance LVLMs' ability in performing grounded compositional visual reasoning tasks. Our approach constructs a hierarchical structure of multi-modal alignments, ranging from simple to complex concepts. By progressively aligning textual descriptions with corresponding visual regions, our model learns to leverage contextual information from lower levels to inform higher-level reasoning. To facilitate this learning process, we introduce a data generation process that creates a novel dataset derived from Visual Genome, providing a wide range of nested compositional vision-language pairs. Experimental results demonstrate that our PromViL framework significantly outperforms baselines on various visual grounding and compositional question answering tasks.
Authors: Yayun Qi, Hongxi Li, Yiqi Song, Xinxiao Wu, Jiebo Luo
Abstract: The exploration of various vision-language tasks, such as visual captioning, visual question answering, and visual commonsense reasoning, is an important area in artificial intelligence and continuously attracts the research community's attention. Despite the improvements in overall performance, classic challenges still exist in vision-language tasks and hinder the development of this area. In recent years, the rise of pre-trained models is driving the research on vision-language tasks. Thanks to the massive scale of training data and model parameters, pre-trained models have exhibited excellent performance in numerous downstream tasks. Inspired by the powerful capabilities of pre-trained models, new paradigms have emerged to solve the classic challenges. Such methods have become mainstream in current research with increasing attention and rapid advances. In this paper, we present a comprehensive overview of how vision-language tasks benefit from pre-trained models. First, we review several main challenges in vision-language tasks and discuss the limitations of previous solutions before the era of pre-training. Next, we summarize the recent advances in incorporating pre-trained models to address the challenges in vision-language tasks. Finally, we analyze the potential risks associated with the inherent limitations of pre-trained models and discuss possible solutions, attempting to provide future research directions.
Authors: Mohammadmostafa Rostamkhani, Baktash Ansari, Hoorieh Sabzevari, Farzan Rahmani, Sauleh Eetemadi
Abstract: In recent years, Visual Question Answering (VQA) has made significant strides, particularly with the advent of multimodal models that integrate vision and language understanding. However, existing VQA datasets often overlook the complexities introduced by image illusions, which pose unique challenges for both human perception and model interpretation. In this study, we introduce a novel task called Illusory VQA, along with four specialized datasets: IllusionMNIST, IllusionFashionMNIST, IllusionAnimals, and IllusionChar. These datasets are designed to evaluate the performance of state-of-the-art multimodal models in recognizing and interpreting visual illusions. We assess the zero-shot performance of various models, fine-tune selected models on our datasets, and propose a simple yet effective solution for illusion detection using Gaussian and blur low-pass filters. We show that this method increases the performance of models significantly and in the case of BLIP-2 on IllusionAnimals without any fine-tuning, it outperforms humans. Our findings highlight the disparity between human and model perception of illusions and demonstrate that fine-tuning and specific preprocessing techniques can significantly enhance model robustness. This work contributes to the development of more human-like visual understanding in multimodal models and suggests future directions for adapting filters using learnable parameters.
Authors: Xingchen Song, Mengtao Xing, Changwei Ma, Shengqiang Li, Di Wu, Binbin Zhang, Fuping Pan, Dinghao Zhou, Yuekai Zhang, Shun Lei, Zhendong Peng, Zhiyong Wu
Abstract: It is well known that LLM-based systems are data-hungry. Recent LLM-based TTS works typically employ complex data processing pipelines to obtain high-quality training data. These sophisticated pipelines require excellent models at each stage (e.g., speech denoising, speech enhancement, speaker diarization, and punctuation models), which themselves demand high-quality training data and are rarely open-sourced. Even with state-of-the-art models, issues persist, such as incomplete background noise removal and misalignment between punctuation and actual speech pauses. Moreover, the stringent filtering strategies often retain only 10-30\% of the original data, significantly impeding data scaling efforts. In this work, we leverage a noise-robust audio tokenizer (S3Tokenizer) to design a simplified yet effective TTS data processing pipeline that maintains data quality while substantially reducing data acquisition costs, achieving a data retention rate of over 50\%. Beyond data scaling challenges, LLM-based TTS systems also incur higher deployment costs compared to conventional approaches. Current systems typically use LLMs solely for text-to-token generation, while requiring separate models (e.g., flow matching models) for token-to-waveform generation, which cannot be directly executed by LLM inference engines, further complicating deployment. To address these challenges, we eliminate redundant modules in both LLM and flow components, replacing the flow model backbone with an LLM architecture. Building upon this simplified flow backbone, we propose a unified architecture for both streaming and non-streaming inference, significantly reducing deployment costs. Finally, we explore the feasibility of unifying TTS and ASR tasks using the same data for training, thanks to the simplified pipeline and the S3Tokenizer that reduces the quality requirements for TTS training data.
Authors: Xingyu Peng, Junran Wu, Ruomei Liu, Ke Xu
Abstract: Traditional methods for detecting rumors on social media primarily focus on analyzing textual content, often struggling to capture the complexity of online interactions. Recent research has shifted towards leveraging graph neural networks to model the hierarchical conversation structure that emerges during rumor propagation. However, these methods tend to overlook the temporal aspect of rumor propagation and may disregard potential noise within the propagation structure. In this paper, we propose a novel approach that incorporates temporal information by constructing a weighted propagation tree, where the weight of each edge represents the time interval between connected posts. Drawing upon the theory of structural entropy, we transform this tree into a coding tree. This transformation aims to preserve the essential structure of rumor propagation while reducing noise. Finally, we introduce a recursive neural network to learn from the coding tree for rumor veracity prediction. Experimental results on two common datasets demonstrate the superiority of our approach.
Authors: Zun Wang, Jialu Li, Yicong Hong, Songze Li, Kunchang Li, Shoubin Yu, Yi Wang, Yu Qiao, Yali Wang, Mohit Bansal, Limin Wang
Abstract: Creating high-quality data for training robust language-instructed agents is a long-lasting challenge in embodied AI. In this paper, we introduce a Self-Refining Data Flywheel (SRDF) that generates high-quality and large-scale navigational instruction-trajectory pairs by iteratively refining the data pool through the collaboration between two models, the instruction generator and the navigator, without any human-in-the-loop annotation. Specifically, SRDF starts with using a base generator to create an initial data pool for training a base navigator, followed by applying the trained navigator to filter the data pool. This leads to higher-fidelity data to train a better generator, which can, in turn, produce higher-quality data for training the next-round navigator. Such a flywheel establishes a data self-refining process, yielding a continuously improved and highly effective dataset for large-scale language-guided navigation learning. Our experiments demonstrate that after several flywheel rounds, the navigator elevates the performance boundary from 70% to 78% SPL on the classic R2R test set, surpassing human performance (76%) for the first time. Meanwhile, this process results in a superior generator, evidenced by a SPICE increase from 23.5 to 26.2, better than all previous VLN instruction generation methods. Finally, we demonstrate the scalability of our method through increasing environment and instruction diversity, and the generalization ability of our pre-trained navigator across various downstream navigation tasks, surpassing state-of-the-art methods by a large margin in all cases.
Authors: Martin Klissarov, Mikael Henaff, Roberta Raileanu, Shagun Sodhani, Pascal Vincent, Amy Zhang, Pierre-Luc Bacon, Doina Precup, Marlos C. Machado, Pierluca D'Oro
Abstract: Describing skills in natural language has the potential to provide an accessible way to inject human knowledge about decision-making into an AI system. We present MaestroMotif, a method for AI-assisted skill design, which yields high-performing and adaptable agents. MaestroMotif leverages the capabilities of Large Language Models (LLMs) to effectively create and reuse skills. It first uses an LLM's feedback to automatically design rewards corresponding to each skill, starting from their natural language description. Then, it employs an LLM's code generation abilities, together with reinforcement learning, for training the skills and combining them to implement complex behaviors specified in language. We evaluate MaestroMotif using a suite of complex tasks in the NetHack Learning Environment (NLE), demonstrating that it surpasses existing approaches in both performance and usability.
Authors: Michal Shlapentokh-Rothman, Yu-Xiong Wang, Derek Hoiem
Abstract: Visual programming prompts LLMs (large language mod-els) to generate executable code for visual tasks like visual question answering (VQA). Prompt-based methods are difficult to improve while also being unreliable and costly in both time and money. Our goal is to develop an efficient visual programming system without 1) using prompt-based LLMs at inference time and 2) a large set of program and answer annotations. We develop a synthetic data augmentation approach and alternative program generation method based on decoupling programs into higher-level skills called templates and the corresponding arguments. Our results show that with data augmentation, prompt-free smaller LLMs ($\approx$ 1B parameters) are competitive with state-of-the art models with the added benefit of much faster inference
Authors: Thong Nguyen, Cong-Duy Nguyen, Xiaobao Wu, See-Kiong Ng, Anh Tuan Luu
Abstract: With the burgeoning amount of data of image-text pairs and diversity of Vision-and-Language (V\&L) tasks, scholars have introduced an abundance of deep learning models in this research domain. Furthermore, in recent years, transfer learning has also shown tremendous success in Computer Vision for tasks such as Image Classification, Object Detection, etc., and in Natural Language Processing for Question Answering, Machine Translation, etc. Inheriting the spirit of Transfer Learning, research works in V\&L have devised multiple pretraining techniques on large-scale datasets in order to enhance the performance of downstream tasks. The aim of this article is to provide a comprehensive revision of contemporary V\&L pretraining models. In particular, we categorize and delineate pretraining approaches, along with the summary of state-of-the-art vision-and-language pretrained models. Moreover, a list of training datasets and downstream tasks is supplied to further polish the perspective into V\&L pretraining. Lastly, we decided to take a further step to discuss numerous directions for future research.
Authors: Fei Zhao, Taotian Pang, Zhen Wu, Zheng Ma, Shujian Huang, Xinyu Dai
Abstract: In-context learning (ICL) is a new paradigm for natural language processing (NLP), where a large language model (LLM) observes a small number of demonstrations and a test instance as its input, and directly makes predictions without updating model parameters. Previous studies have revealed that ICL is sensitive to the selection and the ordering of demonstrations. However, there are few studies regarding the impact of the demonstration number on the ICL performance within a limited input length of LLM, because it is commonly believed that the number of demonstrations is positively correlated with model performance. In this paper, we found this conclusion does not always hold true. Through pilot experiments, we discover that increasing the number of demonstrations does not necessarily lead to improved performance. Building upon this insight, we propose a Dynamic Demonstrations Controller (D$^2$Controller), which can improve the ICL performance by adjusting the number of demonstrations dynamically. The experimental results show that D$^2$Controller yields a 4.6% relative improvement on ten different sizes of LLMs across ten datasets. Moreover, we also extend our method to previous ICL models and achieve competitive results.
Authors: Yupei Du, Albert Gatt, Dong Nguyen
Abstract: Despite the massive success of fine-tuning Pre-trained Language Models (PLMs), they remain susceptible to out-of-distribution input. Dataset cartography is a simple yet effective dual-model approach that improves the robustness of fine-tuned PLMs. It involves fine-tuning a model on the original training set (i.e. reference model), selecting a subset of important training instances based on the training dynamics, and fine-tuning again only on these selected examples (i.e. main model). However, this approach requires fine-tuning the same model twice, which is computationally expensive for large PLMs. In this paper, we show that (1) training dynamics are highly transferable across model sizes and pre-training methods, and that (2) fine-tuning main models using these selected training instances achieves higher training efficiency than empirical risk minimization (ERM). Building on these observations, we propose a novel fine-tuning approach: Fine-Tuning by transFerring Training dynamics (FTFT). Compared with dataset cartography, FTFT uses more efficient reference models and aggressive early stopping. FTFT achieves robustness improvements over ERM while lowering the training cost by up to $\sim 50\%$.
Authors: Nanna Inie, Jonathan Stray, Leon Derczynski
Abstract: Engaging in the deliberate generation of abnormal outputs from Large Language Models (LLMs) by attacking them is a novel human activity. This paper presents a thorough exposition of how and why people perform such attacks, defining LLM red-teaming based on extensive and diverse evidence. Using a formal qualitative methodology, we interviewed dozens of practitioners from a broad range of backgrounds, all contributors to this novel work of attempting to cause LLMs to fail. We focused on the research questions of defining LLM red teaming, uncovering the motivations and goals for performing the activity, and characterizing the strategies people use when attacking LLMs. Based on the data, LLM red teaming is defined as a limit-seeking, non-malicious, manual activity, which depends highly on a team-effort and an alchemist mindset. It is highly intrinsically motivated by curiosity, fun, and to some degrees by concerns for various harms of deploying LLMs. We identify a taxonomy of 12 strategies and 35 different techniques of attacking LLMs. These findings are presented as a comprehensive grounded theory of how and why people attack large language models: LLM red teaming.
Authors: Sara Pieri, Sahal Shaji Mullappilly, Fahad Shahbaz Khan, Rao Muhammad Anwer, Salman Khan, Timothy Baldwin, Hisham Cholakkal
Abstract: In this paper, we introduce BiMediX, the first bilingual medical mixture of experts LLM designed for seamless interaction in both English and Arabic. Our model facilitates a wide range of medical interactions in English and Arabic, including multi-turn chats to inquire about additional details such as patient symptoms and medical history, multiple-choice question answering, and open-ended question answering. We propose a semi-automated English-to-Arabic translation pipeline with human refinement to ensure high-quality translations. We also introduce a comprehensive evaluation benchmark for Arabic medical LLMs. Furthermore, we introduce BiMed1.3M, an extensive Arabic-English bilingual instruction set covering 1.3 Million diverse medical interactions, resulting in over 632 million healthcare specialized tokens for instruction tuning. Our BiMed1.3M dataset includes 250k synthesized multi-turn doctor-patient chats and maintains a 1:2 Arabic-to-English ratio. Our model outperforms state-of-the-art Med42 and Meditron by average absolute gains of 2.5% and 4.1%, respectively, computed across multiple medical evaluation benchmarks in English, while operating at 8-times faster inference. Moreover, our BiMediX outperforms the generic Arabic-English bilingual LLM, Jais-30B, by average absolute gains of 10% on our Arabic medical benchmark and 15% on bilingual evaluations across multiple datasets. Our project page with source code and trained model is available at https://github.com/mbzuai-oryx/BiMediX .
Authors: Mikayel Samvelyan, Sharath Chandra Raparthy, Andrei Lupu, Eric Hambro, Aram H. Markosyan, Manish Bhatt, Yuning Mao, Minqi Jiang, Jack Parker-Holder, Jakob Foerster, Tim Rockt\"aschel, Roberta Raileanu
Abstract: As large language models (LLMs) become increasingly prevalent across many real-world applications, understanding and enhancing their robustness to adversarial attacks is of paramount importance. Existing methods for identifying adversarial prompts tend to focus on specific domains, lack diversity, or require extensive human annotations. To address these limitations, we present Rainbow Teaming, a novel black-box approach for producing a diverse collection of adversarial prompts. Rainbow Teaming casts adversarial prompt generation as a quality-diversity problem and uses open-ended search to generate prompts that are both effective and diverse. Focusing on the safety domain, we use Rainbow Teaming to target various state-of-the-art LLMs, including the Llama 2 and Llama 3 models. Our approach reveals hundreds of effective adversarial prompts, with an attack success rate exceeding 90% across all tested models. Furthermore, we demonstrate that prompts generated by Rainbow Teaming are highly transferable and that fine-tuning models with synthetic data generated by our method significantly enhances their safety without sacrificing general performance or helpfulness. We additionally explore the versatility of Rainbow Teaming by applying it to question answering and cybersecurity, showcasing its potential to drive robust open-ended self-improvement in a wide range of applications.
Authors: Zihao Li, Yucheng Shi, Zirui Liu, Fan Yang, Ali Payani, Ninghao Liu, Mengnan Du
Abstract: The development of Large Language Models (LLMs) relies on extensive text corpora, which are often unevenly distributed across languages. This imbalance results in LLMs performing significantly better on high-resource languages like English, German, and French, while their capabilities in low-resource languages remain inadequate. Currently, there is a lack of quantitative methods to evaluate the performance of LLMs in these low-resource languages. To address this gap, we propose the Language Ranker, an intrinsic metric designed to benchmark and rank languages based on LLM performance using internal representations. By comparing the LLM's internal representation of various languages against a baseline derived from English, we can assess the model's multilingual capabilities in a robust and language-agnostic manner. Our analysis reveals that high-resource languages exhibit higher similarity scores with English, demonstrating superior performance, while low-resource languages show lower similarity scores, underscoring the effectiveness of our metric in assessing language-specific capabilities. Besides, the experiments show that there is a strong correlation between the LLM's performance in different languages and the proportion of those languages in its pre-training corpus. These insights underscore the efficacy of the Language Ranker as a tool for evaluating LLM performance across different languages, particularly those with limited resources.
Authors: Pengxiang Lan, Enneng Yang, Yuting Liu, Guibing Guo, Jianzhe Zhao, Xingwei Wang
Abstract: Prompt tuning is a promising method to fine-tune a pre-trained language model without retraining its large-scale parameters. Instead, it attaches a soft prompt to the input text, whereby downstream tasks can be well adapted by merely learning the embeddings of prompt tokens. Nevertheless, existing methods still suffer from two challenges: (i) they are hard to balance accuracy and efficiency. A longer (shorter) soft prompt generally leads to a better(worse) accuracy but at the cost of more (less) training time. (ii)The performance may not be consistent when adapting to different downstream tasks. We attribute it to the same embedding space but responsible for different requirements of downstream tasks. To address these issues, we propose an Efficient Prompt Tuning method (EPT) by multi-space projection and prompt fusion. Specifically, it decomposes a given soft prompt into a shorter prompt and two low-rank matrices, significantly reducing the training time. Accuracy is also enhanced by leveraging low-rank matrices and the short prompt as additional knowledge sources to enrich the semantics of the original short prompt. In addition, we project the soft prompt into multiple subspaces to improve the performance consistency, and then adaptively learn the combination weights of different spaces through a gating network. Experiments on 13 natural language processing downstream tasks show that our method significantly and consistently outperforms 11 comparison methods with the relative percentage of improvements up to 12.9%, and training time decreased by 14%.
Authors: Holli Sargeant, Ahmed Izzidien, Felix Steffek
Abstract: This paper addresses a critical gap in legal analytics by developing and applying a novel taxonomy for topic classification of summary judgment cases in the United Kingdom. Using a curated dataset of summary judgment cases, we use the Large Language Model Claude 3 Opus to explore functional topics and trends. We find that Claude 3 Opus correctly classified the topic with an accuracy of 87.13% and an F1 score of 0.87. The analysis reveals distinct patterns in the application of summary judgments across various legal domains. As case law in the United Kingdom is not originally labelled with keywords or a topic filtering option, the findings not only refine our understanding of the thematic underpinnings of summary judgments but also illustrate the potential of combining traditional and AI-driven approaches in legal classification. Therefore, this paper provides a new and general taxonomy for UK law. The implications of this work serve as a foundation for further research and policy discussions in the field of judicial administration and computational legal research methodologies.
Authors: Yutao Zhu, Zhaoheng Huang, Zhicheng Dou, Ji-Rong Wen
Abstract: Retrieval-augmented generation (RAG) is a promising way to improve large language models (LLMs) for generating more factual, accurate, and up-to-date content. Existing methods either optimize prompts to guide LLMs in leveraging retrieved information or directly fine-tune LLMs to adapt to RAG scenarios. Although fine-tuning can yield better performance, it often compromises the LLMs' general generation capabilities by modifying their parameters. This limitation poses challenges in practical applications, especially when LLMs are already deployed, as parameter adjustments may affect their original functionality. To address this, we propose a novel method that involves learning scalable and pluggable virtual tokens for RAG. By maintaining the LLMs' original parameters and fine-tuning only the embeddings of these pluggable tokens, our approach not only enhances LLMs' performance but also preserves their general generation capabilities. Furthermore, we design several training strategies to improve the scalability, flexibility, and generalizability of our method. Comprehensive experiments across 12 question-answering tasks demonstrate the superiority of our approach.
Authors: Chujie Gao, Siyuan Wu, Yue Huang, Dongping Chen, Qihui Zhang, Zhengyan Fu, Yao Wan, Lichao Sun, Xiangliang Zhang
Abstract: Large Language Models (LLMs) have achieved remarkable success across various industries due to their exceptional generative capabilities. However, for safe and effective real-world deployments, ensuring honesty and helpfulness is critical. This paper addresses the question: Can we prioritize the helpfulness of LLMs while preserving their honesty? To begin with, we establish exhaustive principles aimed at guaranteeing the honesty of LLM. Additionally, we introduce a novel dataset, referred to as HoneSet, comprising 930 queries spanning six categories meticulously crafted to assess an LLM's capacity for maintaining honesty. Subsequently, we present two approaches to augmenting honesty and helpfulness in LLMs: a training-free enhancement and a fine-tuning-based improvement. The training-free approach, which is based on curiosity-driven prompting, empowers LLMs to articulate internal confusion and uncertainty regarding queries, thereby optimizing their responses. Conversely, the fine-tuning-based method employs a two-stage process inspired by curriculum learning: initially instructing LLMs to discern between honest and dishonest responses, then refining their training to enhance helpfulness. Experiments conducted on nine prominent LLMs demonstrate a significant improvement in alignment with honesty across all models through the implementation of our proposed enhancements. Particularly noteworthy is the 65.3% enhancement observed in Llama3-8b and the remarkable 124.7% improvement in Mistral-7b, as measured by the H$^{2}$ (honest and helpful) assessment. We believe that our work can pave the way for developing more trustworthy LLMs for real-world applications.
Authors: Hakaze Cho, Yoshihiro Sakai, Kenshiro Tanaka, Mariko Kato, Naoya Inoue
Abstract: In this paper, we investigate the output token probability information in the output embedding of language models. We find an approximate common log-linear encoding of output token probabilities within the output embedding vectors and empirically demonstrate that it is accurate and sparse. As a causality examination, we steer the encoding in output embedding to modify the output probability distribution accurately. Moreover, the sparsity we find in output probability encoding suggests that a large number of dimensions in the output embedding do not contribute to causal language modeling. Therefore, we attempt to delete the output-unrelated dimensions and find more than 30% of the dimensions can be deleted without significant movement in output distribution and sequence generation. Additionally, in the pre-training dynamics of language models, we find that the output embeddings capture the corpus token frequency information in early steps, even before an obvious convergence of parameters starts.
Authors: Sherzod Hakimov, Yerkezhan Abdullayeva, Kushal Koshti, Antonia Schmidt, Yan Weiser, Anne Beyer, David Schlangen
Abstract: While the situation has improved for text-only models, it again seems to be the case currently that multimodal (text and image) models develop faster than ways to evaluate them. In this paper, we bring a recently developed evaluation paradigm from text models to multimodal models, namely evaluation through the goal-oriented game (self) play, complementing reference-based and preference-based evaluation. Specifically, we define games that challenge a model's capability to represent a situation from visual information and align such representations through dialogue. We find that the largest closed models perform rather well on the games that we define, while even the best open-weight models struggle with them. On further analysis, we find that the exceptional deep captioning capabilities of the largest models drive some of the performance. There is still room to grow for both kinds of models, ensuring the continued relevance of the benchmark.
Authors: Sheikh Asif Imran, Mohammad Nur Hossain Khan, Subrata Biswas, Bashima Islam
Abstract: Integrating inertial measurement units (IMUs) with large language models (LLMs) expands the potential of multimodal AI, enabling more nuanced human activity analysis. In this paper, we introduce LLaSA (Large Language and Sensor Assistant), a multimodal large language model built on LIMU-BERT and Llama, designed to interpret and answer queries related to human activities and motion analysis, leveraging sensor data and contextual reasoning. To develop LLaSA, we introduce two key datasets: SensorCaps, a comprehensive collection of 35,960 IMU-derived narratives with handcrafted features, and OpenSQA, an instruction-following dataset containing 179,727 question-answer pairs aware of the sensor and human activity context. These datasets provide diverse and rich inputs to train LLaSA for complex sensor-based queries. To optimize LLaSA's performance, we apply a unique hyperparameter tuning method, which significantly enhances its effectiveness in contextual question-answering tasks. Extensive evaluations, including a human-led assessment of the question-answering, demonstrate that LLaSA achieves superior data interpretation and context-aware responses compared to GPT-3.5-Turbo and Vicuna-1.5-13b-16K. These contributions advance the frontier of sensor-aware LLMs and create new opportunities for impactful multimodal research in healthcare, sports science, and human-computer interactions. Our code repository and datasets can be found at https://github.com/BASHLab/LLaSA.
Authors: Jasivan Alex Sivakumar, Nafise Sadat Moosavi
Abstract: Within numerical reasoning, understanding numbers themselves is still a challenge for existing language models. Simple generalisations, such as solving 100+200 instead of 1+2, can substantially affect model performance (Sivakumar and Moosavi, 2023). Among various techniques, character-level embeddings of numbers have emerged as a promising approach to improve number representation. However, this method has limitations as it leaves the task of aggregating digit representations to the model, which lacks direct supervision for this process. In this paper, we explore the use of mathematical priors to compute aggregated digit embeddings and explicitly incorporate these aggregates into transformer models. This can be achieved either by adding a special token to the input embeddings or by introducing an additional loss function to enhance correct predictions. We evaluate the effectiveness of incorporating this explicit aggregation, analysing its strengths and shortcomings, and discuss future directions to better benefit from this approach. Our methods, while simple, are compatible with any pretrained model, easy to implement, and have been made publicly available.
Authors: Sangjoon Park, Chan Woo Wee, Seo Hee Choi, Kyung Hwan Kim, Jee Suk Chang, Hong In Yoon, Ik Jae Lee, Yong Bae Kim, Jaeho Cho, Ki Chang Keum, Chang Geol Lee, Hwa Kyung Byun, Woong Sub Koom
Abstract: Accurate survival prediction in radiotherapy (RT) is critical for optimizing treatment decisions. This study developed and validated the RT-Surv framework, which integrates general-domain, open-source large language models (LLMs) to structure unstructured electronic health records alongside structured clinical data. Using data from 34,276 patients and an external cohort of 852, the framework successfully transformed unstructured clinical information into structured formats. Incorporating LLM-structured clinical features improved the concordance index from 0.779 to 0.842 during external validation, demonstrating a significant performance enhancement. Key LLM-structured features, such as disease extent, general condition, and RT purpose, showed high predictive importance and aligned closely with statistically significant predictors identified through conventional statistical analyses, thereby improving model interpretability. Furthermore, the framework enhanced risk stratification, enabling more distinct differentiation among low-, intermediate-, and high-risk groups (p < 0.001) using LLM-structured clinical features. These findings highlight the potential of LLMs to convert unstructured data into actionable insights, improving predictive modeling and patient outcomes in clinics.
Authors: Xufeng Duan, Xinyu Zhou, Bei Xiao, Zhenguang G. Cai
Abstract: As large language models (LLMs) advance in their linguistic capacity, understanding how they capture aspects of language competence remains a significant challenge. This study therefore employs psycholinguistic paradigms in English, which are well-suited for probing deeper cognitive aspects of language processing, to explore neuron-level representations in language model across three tasks: sound-shape association, sound-gender association, and implicit causality. Our findings indicate that while GPT-2-XL struggles with the sound-shape task, it demonstrates human-like abilities in both sound-gender association and implicit causality. Targeted neuron ablation and activation manipulation reveal a crucial relationship: When GPT-2-XL displays a linguistic ability, specific neurons correspond to that competence; conversely, the absence of such an ability indicates a lack of specialized neurons. This study is the first to utilize psycholinguistic experiments to investigate deep language competence at the neuron level, providing a new level of granularity in model interpretability and insights into the internal mechanisms driving language ability in the transformer-based LLM.
Authors: Alexander Scarlatos, Ryan S. Baker, Andrew Lan
Abstract: Recent advances in large language models (LLMs) have led to the development of artificial intelligence (AI)-powered tutoring chatbots, showing promise in providing broad access to high-quality personalized education. Existing works have studied how to make LLMs follow tutoring principles, but have not studied broader uses of LLMs for supporting tutoring. Up until now, tracing student knowledge and analyzing misconceptions has been difficult and time-consuming to implement for open-ended dialogue tutoring. In this work, we investigate whether LLMs can be supportive of this task: we first use LLM prompting methods to identify the knowledge components/skills involved in each dialogue turn, i.e., a tutor utterance posing a task or a student utterance that responds to it. We also evaluate whether the student responds correctly to the tutor and verify the LLM's accuracy using human expert annotations. We then apply a range of knowledge tracing (KT) methods on the resulting labeled data to track student knowledge levels over an entire dialogue. We conduct experiments on two tutoring dialogue datasets, and show that a novel yet simple LLM-based method, LLMKT, significantly outperforms existing KT methods in predicting student response correctness in dialogues. We perform extensive qualitative analyses to highlight the challenges in dialogueKT and outline multiple avenues for future work.
Authors: Daniil Gurgurov, Rishu Kumar, Simon Ostermann
Abstract: Contextualized embeddings based on large language models (LLMs) are available for various languages, but their coverage is often limited for lower resourced languages. Using LLMs for such languages is often difficult due to a high computational cost; not only during training, but also during inference. Static word embeddings are much more resource-efficient ("green"), and thus still provide value, particularly for very low-resource languages. There is, however, a notable lack of comprehensive repositories with such embeddings for diverse languages. To address this gap, we present GrEmLIn, a centralized repository of green, static baseline embeddings for 87 mid- and low-resource languages. We compute GrEmLIn embeddings with a novel method that enhances GloVe embeddings by integrating multilingual graph knowledge, which makes our static embeddings competitive with LLM representations, while being parameter-free at inference time. Our experiments demonstrate that GrEmLIn embeddings outperform state-of-the-art contextualized embeddings from E5 on the task of lexical similarity. They remain competitive in extrinsic evaluation tasks like sentiment analysis and natural language inference, with average performance gaps of just 5-10\% or less compared to state-of-the-art models, given a sufficient vocabulary overlap with the target task, and underperform only on topic classification. Our code and embeddings are publicly available at https://huggingface.co/DFKI.
Authors: Jun Shern Chan, Neil Chowdhury, Oliver Jaffe, James Aung, Dane Sherburn, Evan Mays, Giulio Starace, Kevin Liu, Leon Maksin, Tejal Patwardhan, Lilian Weng, Aleksander M\k{a}dry
Abstract: We introduce MLE-bench, a benchmark for measuring how well AI agents perform at machine learning engineering. To this end, we curate 75 ML engineering-related competitions from Kaggle, creating a diverse set of challenging tasks that test real-world ML engineering skills such as training models, preparing datasets, and running experiments. We establish human baselines for each competition using Kaggle's publicly available leaderboards. We use open-source agent scaffolds to evaluate several frontier language models on our benchmark, finding that the best-performing setup--OpenAI's o1-preview with AIDE scaffolding--achieves at least the level of a Kaggle bronze medal in 16.9% of competitions. In addition to our main results, we investigate various forms of resource scaling for AI agents and the impact of contamination from pre-training. We open-source our benchmark code (github.com/openai/mle-bench/) to facilitate future research in understanding the ML engineering capabilities of AI agents.
Authors: Yiming Chen, Xianghu Yue, Chen Zhang, Xiaoxue Gao, Robby T. Tan, Haizhou Li
Abstract: Building on the success of large language models (LLMs), recent advancements such as GPT-4o have enabled real-time speech interactions through LLM-based voice assistants, offering a significantly improved user experience compared to traditional text-based interactions. However, the absence of benchmarks designed to evaluate these speech interaction capabilities has hindered progress of LLM-based voice assistants development. Current evaluations focus primarily on automatic speech recognition (ASR) or general knowledge evaluation with clean speeches, neglecting the more intricate, real-world scenarios that involve diverse speaker characteristics, environmental and content factors. To address this, we introduce VoiceBench, the first benchmark designed to provide a multi-faceted evaluation of LLM-based voice assistants. VoiceBench also includes both real and synthetic spoken instructions that incorporate the above three key real-world variations. Extensive experiments reveal the limitations of current LLM-based voice assistant models and offer valuable insights for future research and development in this field.
Authors: Hengyuan Zhang, Chenming Shang, Sizhe Wang, Dongdong Zhang, Feng Yao, Renliang Sun, Yiyao Yu, Yujiu Yang, Furu Wei
Abstract: Although fine-tuning Large Language Models (LLMs) with multilingual data can rapidly enhance the multilingual capabilities of LLMs, they still exhibit a performance gap between the dominant language (e.g., English) and non-dominant ones due to the imbalance of training data across languages. To further enhance the performance of non-dominant languages, we propose ShifCon, a Shift-based Contrastive framework that aligns the internal forward process of other languages toward that of the dominant one. Specifically, it shifts the representations of non-dominant languages into the dominant language subspace, allowing them to access relatively rich information encoded in the model parameters. The enriched representations are then shifted back into their original language subspace before generation. Moreover, we introduce a subspace distance metric to pinpoint the optimal layer area for shifting representations and employ multilingual contrastive learning to further enhance the alignment of representations within this area. Experiments demonstrate that our ShifCon framework significantly enhances the performance of non-dominant languages, particularly for low-resource ones. Further analysis offers extra insights to verify the effectiveness of ShifCon and propel future research
Authors: Brian R. Y. Huang
Abstract: Large language models (LLMs) remain vulnerable to a slew of adversarial attacks and jailbreaking methods. One common approach employed by white-hat attackers, or red-teamers, is to process model inputs and outputs using string-level obfuscations, which can include leetspeak, rotary ciphers, Base64, ASCII, and more. Our work extends these encoding-based attacks by unifying them in a framework of invertible string transformations. With invertibility, we can devise arbitrary string compositions, defined as sequences of transformations, that we can encode and decode end-to-end programmatically. We devise a automated best-of-n attack that samples from a combinatorially large number of string compositions. Our jailbreaks obtain competitive attack success rates on several leading frontier models when evaluated on HarmBench, highlighting that encoding-based attacks remain a persistent vulnerability even in advanced LLMs.
Authors: Jinhao Jiang, Zhipeng Chen, Yingqian Min, Jie Chen, Xiaoxue Cheng, Jiapeng Wang, Yiru Tang, Haoxiang Sun, Jia Deng, Wayne Xin Zhao, Zheng Liu, Dong Yan, Jian Xie, Zhongyuan Wang, Ji-Rong Wen
Abstract: Recently, test-time scaling has garnered significant attention from the research community, largely due to the substantial advancements of the o1 model released by OpenAI. By allocating more computational resources during the inference phase, large language models~(LLMs) can extensively explore the solution space by generating more thought tokens or diverse solutions, thereby producing more accurate responses. However, developing an o1-like reasoning approach is challenging, and researchers have been making various attempts to advance this open area of research. In this paper, we present a preliminary exploration into enhancing the reasoning abilities of LLMs through reward-guided tree search algorithms. This framework is implemented by integrating the policy model, reward model, and search algorithm. It is primarily constructed around a tree search algorithm, where the policy model navigates a dynamically expanding tree guided by a specially trained reward model. We thoroughly explore various design considerations necessary for implementing this framework and provide a detailed report of the technical aspects. To assess the effectiveness of our approach, we focus on mathematical reasoning tasks and conduct extensive evaluations on four challenging datasets, significantly enhancing the reasoning abilities of LLMs.
Authors: Jinming Xing, Ruilin Xing, Yan Sun
Abstract: Large Language Models (LLMs) have revolutionized natural language processing (NLP) by delivering state-of-the-art performance across a variety of tasks. Among these, Transformer-based models like BERT and GPT rely on pooling layers to aggregate token-level embeddings into sentence-level representations. Common pooling mechanisms such as Mean, Max, and Weighted Sum play a pivotal role in this aggregation process. Despite their widespread use, the comparative performance of these strategies on different LLM architectures remains underexplored. To address this gap, this paper investigates the effects of these pooling mechanisms on two prominent LLM families -- BERT and GPT, in the context of sentence-level sentiment analysis. Comprehensive experiments reveal that each pooling mechanism exhibits unique strengths and weaknesses depending on the task's specific requirements. Our findings underline the importance of selecting pooling methods tailored to the demands of particular applications, prompting a re-evaluation of common assumptions regarding pooling operations. By offering actionable insights, this study contributes to the optimization of LLM-based models for downstream tasks.
Authors: Ivory Yang, Weicheng Ma, Soroush Vosoughi
Abstract: The preservation and revitalization of endangered and extinct languages is a meaningful endeavor, conserving cultural heritage while enriching fields like linguistics and anthropology. However, these languages are typically low-resource, making their reconstruction labor-intensive and costly. This challenge is exemplified by Nushu, a rare script historically used by Yao women in China for self-expression within a patriarchal society. To address this challenge, we introduce NushuRescue, an AI-driven framework designed to train large language models (LLMs) on endangered languages with minimal data. NushuRescue automates evaluation and expands target corpora to accelerate linguistic revitalization. As a foundational component, we developed NCGold, a 500-sentence Nushu-Chinese parallel corpus, the first publicly available dataset of its kind. Leveraging GPT-4-Turbo, with no prior exposure to Nushu and only 35 short examples from NCGold, NushuRescue achieved 48.69% translation accuracy on 50 withheld sentences and generated NCSilver, a set of 98 newly translated modern Chinese sentences of varying lengths. A sample of both NCGold and NCSilver is included in the Supplementary Materials. Additionally, we developed FastText-based and Seq2Seq models to further support research on Nushu. NushuRescue provides a versatile and scalable tool for the revitalization of endangered languages, minimizing the need for extensive human input.
Authors: Marco Molinari, Victor Shao, Vladimir Tregubiak, Abhimanyu Pandey, Mateusz Mikolajczak, Sebastian Kuznetsov Ryder Torres Pereira
Abstract: Determining company similarity is a vital task in finance, underpinning hedging, risk management, portfolio diversification, and more. Practitioners often rely on sector and industry classifications to gauge similarity, such as SIC-codes and GICS-codes - the former being used by the U.S. Securities and Exchange Commission (SEC), and the latter widely used by the investment community. Since these classifications can lack granularity and often need to be updated, using clusters of embeddings of company descriptions has been proposed as a potential alternative, but the lack of interpretability in token embeddings poses a significant barrier to adoption in high-stakes contexts. Sparse Autoencoders (SAEs) have shown promise in enhancing the interpretability of Large Language Models (LLMs) by decomposing LLM activations into interpretable features. We apply SAEs to company descriptions, obtaining meaningful clusters of equities in the process. We benchmark SAE features against SIC-codes, Major Group codes, and Embeddings. Our results demonstrate that SAE features not only replicate but often surpass sector classifications and embeddings in capturing fundamental company characteristics. This is evidenced by their superior performance in correlating monthly returns - a proxy for similarity - and generating higher Sharpe ratio co-integration strategies, which underscores deeper fundamental similarities among companies.
Authors: Lingxiao Wei, He Yan, Xiangju Lu, Junmin Zhu, Jun Wang, Wei Zhang
Abstract: Large Language Models (LLMs) have been well-researched in many long-context tasks. However, due to high annotation costs, high-quality long-context summary datasets for training or evaluation are scarce, limiting further research. In this work, we introduce CNNSum, a new multi-scale Chinese long-context novel summarization benchmark, including four subsets, length covering 16k to 128k, 695 samples in total, the annotations are human-driven. We evaluate commercial and open-source models on CNNSum and conduct a detailed analysis. Based on the observations, we further conduct fine-tuning exploration with short-context summary data. In our study: (1) GPT-4o underperformed, due to excessive subjective commentary. (2) Currently, long-context summarization mainly relies on memory ability, small LLMs with stable longer context lengths are the most cost-effective. Using long data concatenated from short-context summaries makes a significant improvement. (3) Prompt templates may cause a large performance gap but can be mitigated through fine-tuning. (4) Fine-tuned Chat or Instruction versions may harm the Base model and further fine-tuning cannot bridge performance gap. (5) while models with RoPE base scaling exhibit strong extrapolation potential, their performance may vary significantly when combined with other interpolation methods and need careful selection. (6) CNNSum provides more reliable and insightful evaluation results than other benchmarks. We release CNNSum to advance research in this field (https://github.com/CxsGhost/CNNSum).
Authors: Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, Yuandong Tian
Abstract: Large language models (LLMs) are restricted to reason in the "language space", where they typically express the reasoning process with a chain-of-thought (CoT) to solve a complex reasoning problem. However, we argue that language space may not always be optimal for reasoning. For example, most word tokens are primarily for textual coherence and not essential for reasoning, while some critical tokens require complex planning and pose huge challenges to LLMs. To explore the potential of LLM reasoning in an unrestricted latent space instead of using natural language, we introduce a new paradigm Coconut (Chain of Continuous Thought). We utilize the last hidden state of the LLM as a representation of the reasoning state (termed "continuous thought"). Rather than decoding this into a word token, we feed it back to the LLM as the subsequent input embedding directly in the continuous space. Experiments show that Coconut can effectively augment the LLM on several reasoning tasks. This novel latent reasoning paradigm leads to emergent advanced reasoning patterns: the continuous thought can encode multiple alternative next reasoning steps, allowing the model to perform a breadth-first search (BFS) to solve the problem, rather than prematurely committing to a single deterministic path like CoT. Coconut outperforms CoT in certain logical reasoning tasks that require substantial backtracking during planning, with fewer thinking tokens during inference. These findings demonstrate the promise of latent reasoning and offer valuable insights for future research.
Authors: Yuxiao Lu, Arunesh Sinha, Pradeep Varakantham
Abstract: Large Language Models (LLMs) generating unsafe responses to toxic prompts is a significant issue in their applications. While various efforts aim to address this safety concern, previous approaches often demand substantial human data collection or rely on the less dependable option of using another LLM to generate corrective data. In this paper, we aim to take this problem and overcome limitations of requiring significant high-quality human data. Our method requires only a small set of unsafe responses to toxic prompts, easily obtained from the unsafe LLM itself. By employing a semantic cost combined with a negative Earth Mover Distance (EMD) loss, we guide the LLM away from generating unsafe responses. Additionally, we propose a novel lower bound for EMD loss, enabling more efficient optimization. Our results demonstrate superior performance and data efficiency compared to baselines, and we further examine the nuanced effects of over-alignment and potential degradation of language capabilities when using contrastive data.
Authors: Yongqi Li, Xin Miao, Shen Zhou, Mayi Xu, Yuyang Ren, Tieyun Qian
Abstract: Despite the rapid progress that existing automated feedback methods have made in correcting the output of large language models (LLMs), these methods cannot be well applied to the relation extraction (RE) task due to their designated feedback objectives and correction manner. To address this problem, we propose a novel automated feedback framework for RE, which presents a rationale supervisor to verify the rationale and provides re-selected demonstrations as feedback to correct the initial prediction. Specifically, we first design a causal intervention and observation method to collect biased/unbiased rationales for contrastive training the rationale supervisor. Then, we present a verification-feedback-correction procedure to iteratively enhance LLMs' capability of handling the RE task. Extensive experiments prove that our proposed framework significantly outperforms existing methods.
Authors: Lance Calvin Lim Gamboa, Mark Lee
Abstract: Bias studies on multilingual models confirm the presence of gender-related stereotypes in masked models processing languages with high NLP resources. We expand on this line of research by introducing Filipino CrowS-Pairs and Filipino WinoQueer: benchmarks that assess both sexist and anti-queer biases in pretrained language models (PLMs) handling texts in Filipino, a low-resource language from the Philippines. The benchmarks consist of 7,074 new challenge pairs resulting from our cultural adaptation of English bias evaluation datasets, a process that we document in detail to guide similar forthcoming efforts. We apply the Filipino benchmarks on masked and causal multilingual models, including those pretrained on Southeast Asian data, and find that they contain considerable amounts of bias. We also find that for multilingual models, the extent of bias learned for a particular language is influenced by how much pretraining data in that language a model was exposed to. Our benchmarks and insights can serve as a foundation for future work analyzing and mitigating bias in multilingual models.
Authors: Tom Kouwenhoven, Max Peeperkorn, Tessa Verhoef
Abstract: Human languages have evolved to be structured through repeated language learning and use. These processes introduce biases that operate during language acquisition and shape linguistic systems toward communicative efficiency. In this paper, we investigate whether the same happens if artificial languages are optimised for implicit biases of Large Language Models (LLMs). To this end, we simulate a classical referential game in which LLMs learn and use artificial languages. Our results show that initially unstructured holistic languages are indeed shaped to have some structural properties that allow two LLM agents to communicate successfully. Similar to observations in human experiments, generational transmission increases the learnability of languages, but can at the same time result in non-humanlike degenerate vocabularies. Taken together, this work extends experimental findings, shows that LLMs can be used as tools in simulations of language evolution, and opens possibilities for future human-machine experiments in this field.
Authors: Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, Enhong Chen
Abstract: Hallucination is a big shadow hanging over the rapidly evolving Multimodal Large Language Models (MLLMs), referring to the phenomenon that the generated text is inconsistent with the image content. In order to mitigate hallucinations, existing studies mainly resort to an instruction-tuning manner that requires retraining the models with specific data. In this paper, we pave a different way, introducing a training-free method named Woodpecker. Like a woodpecker heals trees, it picks out and corrects hallucinations from the generated text. Concretely, Woodpecker consists of five stages: key concept extraction, question formulation, visual knowledge validation, visual claim generation, and hallucination correction. Implemented in a post-remedy manner, Woodpecker can easily serve different MLLMs, while being interpretable by accessing intermediate outputs of the five stages. We evaluate Woodpecker both quantitatively and qualitatively and show the huge potential of this new paradigm. On the POPE benchmark, our method obtains a 30.66%/24.33% improvement in accuracy over the baseline MiniGPT-4/mPLUG-Owl. The source code is released at https://github.com/BradyFU/Woodpecker.
Authors: Youpeng Zhao, Ming Lin, Huadong Tang, Qiang Wu, Jun Wang
Abstract: Generative Large Language Models (LLMs) stand as a revolutionary advancement in the modern era of artificial intelligence (AI). However, scaling down LLMs for resource-constrained hardware, such as Internet-of-Things (IoT) devices requires non-trivial efforts and domain knowledge. In this paper, we propose a novel information-entropy framework for designing mobile-friendly generative language models. The whole design procedure involves solving a mathematical programming (MP) problem, which can be done on the CPU within minutes, making it nearly zero-cost. We evaluate our designed models, termed MeRino, across fourteen NLP downstream tasks, showing their competitive performance against the state-of-the-art autoregressive transformer models under the mobile setting. Notably, MeRino achieves similar or better performance on both language modeling and zero-shot learning tasks, compared to the 350M parameter OPT while being 4.9x faster on NVIDIA Jetson Nano with 5.5x reduction in model size.
Authors: Filippos Gouidis, Katerina Papantoniou, Konstantinos Papoutsakis, Theodore Patkos, Antonis Argyros, Dimitris Plexousakis
Abstract: Domain-specific knowledge can significantly contribute to addressing a wide variety of vision tasks. However, the generation of such knowledge entails considerable human labor and time costs. This study investigates the potential of Large Language Models (LLMs) in generating and providing domain-specific information through semantic embeddings. To achieve this, an LLM is integrated into a pipeline that utilizes Knowledge Graphs and pre-trained semantic vectors in the context of the Vision-based Zero-shot Object State Classification task. We thoroughly examine the behavior of the LLM through an extensive ablation study. Our findings reveal that the integration of LLM-based embeddings, in combination with general-purpose pre-trained embeddings, leads to substantial performance improvements. Drawing insights from this ablation study, we conduct a comparative analysis against competing models, thereby highlighting the state-of-the-art performance achieved by the proposed approach.
Authors: Susan Athey, Herman Brunborg, Tianyu Du, Ayush Kanodia, Keyon Vafa
Abstract: Vafa et al. (2024) introduced a transformer-based econometric model, CAREER, that predicts a worker's next job as a function of career history (an "occupation model"). CAREER was initially estimated ("pre-trained") using a large, unrepresentative resume dataset, which served as a "foundation model," and parameter estimation was continued ("fine-tuned") using data from a representative survey. CAREER had better predictive performance than benchmarks. This paper considers an alternative where the resume-based foundation model is replaced by a large language model (LLM). We convert tabular data from the survey into text files that resemble resumes and fine-tune the LLMs using these text files with the objective to predict the next token (word). The resulting fine-tuned LLM is used as an input to an occupation model. Its predictive performance surpasses all prior models. We demonstrate the value of fine-tuning and further show that by adding more career data from a different population, fine-tuning smaller LLMs surpasses the performance of fine-tuning larger models.
Authors: Wei Wang, Qing Li
Abstract: Language models have emerged as a critical area of focus in artificial intelligence, particularly with the introduction of groundbreaking innovations like ChatGPT. Large-scale Transformer networks have quickly become the leading approach for advancing natural language processing algorithms. Built on the Transformer architecture, these models enable interactions that closely mimic human communication and, equipped with extensive knowledge, can even assist in guiding human tasks. Despite their impressive capabilities and growing complexity, a key question remains-the theoretical foundations of large language models (LLMs). What makes Transformer so effective for powering intelligent language applications, such as translation and coding? What underlies LLMs' ability for In-Context Learning (ICL)? How does the LoRA scheme enhance the fine-tuning of LLMs? And what supports the practicality of pruning LLMs? To address these critical questions and explore the technological strategies within LLMs, we leverage the Universal Approximation Theory (UAT) to offer a theoretical backdrop, shedding light on the mechanisms that underpin these advancements.
Authors: Alliot Nagle, Adway Girish, Marco Bondaschi, Michael Gastpar, Ashok Vardhan Makkuva, Hyeji Kim
Abstract: We formalize the problem of prompt compression for large language models (LLMs) and present a framework to unify token-level prompt compression methods which create hard prompts for black-box models. We derive the distortion-rate function for this setup as a linear program, and provide an efficient algorithm to compute this fundamental limit via the dual of the linear program. Using the distortion-rate function as the baseline, we study the performance of existing compression schemes on a synthetic dataset consisting of prompts generated from a Markov chain, natural language queries, and their respective answers. Our empirical analysis demonstrates the criticality of query-aware prompt compression, where the compressor has knowledge of the downstream task/query for the black-box LLM. We show that there is a large gap between the performance of current prompt compression methods and the optimal strategy, and propose Adaptive QuerySelect, a query-aware, variable-rate adaptation of a prior work to close the gap. We extend our experiments to a small natural language dataset to further confirm our findings on our synthetic dataset.
Authors: Ethan Ewer, Daewon Chae, Thomas Zeng, Jinkyu Kim, Kangwook Lee
Abstract: Next-token prediction is conventionally done using decoder-only Transformers with causal attention, as this approach allows for efficient reuse of keys and values. What if we were not compute-limited, should we still use decoder-only Transformers? In this work, we introduce Encoder-only Next Token Prediction (ENTP). We use small scale experiments to explore the differences between ENTP and decoders, highlighting potential advantages of ENTP in setting with unbounded compute. We introduce the Count3 task and show, both theoretically and experimentally, that while ENTP can perform this task easily, a decoder-only Transformer cannot. Finally, we empirically demonstrate ENTP's superior performance across various synthetic tasks, such as length generalization and in-context learning.
Authors: David R. Nelson, Ashish Kumar Jaiswal, Noha Ismail, Alexandra Mystikou, Kourosh Salehi-Ashtiani
Abstract: AI language models (LMs) show promise for biological sequence analysis. We re-engineered open-source LMs (GPT-2, BLOOM, DistilRoBERTa, ELECTRA, and Mamba, ranging from 70M to 12B parameters) for microbial sequence classification. The models achieved F1 scores up to 95 and operated 16,580x faster and at 2.9x the recall of BLASTP. They effectively classified the algal dark proteome - uncharacterized proteins comprising about 65% of total proteins - validated on new data including a new, complete Hi-C/Pacbio Chlamydomonas genome. Larger (>1B) LA4SR models reached high accuracy (F1 > 86) when trained on less than 2% of available data, rapidly achieving strong generalization capacity. High accuracy was achieved when training data had intact or scrambled terminal information, demonstrating robust generalization to incomplete sequences. Finally, we provide custom AI explainability software tools for attributing amino acid patterns to AI generative processes and interpret their outputs in evolutionary and biophysical contexts.
Authors: Vishnou Vinayagame, Gregory Senay, Luis Mart\'i
Abstract: Mathematical reasoning capabilities are increasing with tool-augmented language agents, but methods often rely either on closed-source or large models, external data, or extensive prompt engineering. This work introduces MATATA, a novel cost-effective method to train LLM agents for tabular data problems through reasoning, planning, and tool use. With a progressive self-improvement paradigm and an iterative weak supervision, it empowers 3.8B/8B Small Language Models (SLMs), particularly suited for local hosting and sensitive business contexts where data privacy is crucial. By employing a flexible and reusable tools across different datasets, it achieves robust performance with effective scalability across shared tasks. Experiments show that MATATA reaches state-of-the-art performances on FinQA and TAT-QA among reasoning frameworks based on open-source models. Moreover, MATATA models compete with GPT-4 based frameworks on TabMWP, while being SLMs.
Authors: Yuci Liang, Xinheng Lyu, Meidan Ding, Wenting Chen, Jipeng Zhang, Yuexiang Ren, Xiangjian He, Song Wu, Sen Yang, Xiyue Wang, Xiaohan Xing, Linlin Shen
Abstract: Recent advancements in computational pathology have produced patch-level Multi-modal Large Language Models (MLLMs), but these models are limited by their inability to analyze whole slide images (WSIs) comprehensively and their tendency to bypass crucial morphological features that pathologists rely on for diagnosis. To address these challenges, we first introduce WSI-Bench, a large-scale morphology-aware benchmark containing 180k VQA pairs from 9,850 WSIs across 30 cancer types, designed to evaluate MLLMs' understanding of morphological characteristics crucial for accurate diagnosis. Building upon this benchmark, we present WSI-LLaVA, a novel framework for gigapixel WSI understanding that employs a three-stage training approach: WSI-text alignment, feature space alignment, and task-specific instruction tuning. To better assess model performance in pathological contexts, we develop two specialized WSI metrics: WSI-Precision and WSI-Relevance. Experimental results demonstrate that WSI-LLaVA outperforms existing models across all capability dimensions, with a significant improvement in morphological analysis, establishing a clear correlation between morphological understanding and diagnostic accuracy.
Authors: Lanxiang Hu, Qiyu Li, Anze Xie, Nan Jiang, Ion Stoica, Haojian Jin, Hao Zhang
Abstract: Evaluating the reasoning abilities of large language models (LLMs) is challenging. Existing benchmarks often depend on static datasets, which are vulnerable to data contamination and may get saturated over time, or on binary live human feedback that conflates reasoning with other abilities. As the most prominent dynamic benchmark, Chatbot Arena evaluates open-ended questions in real-world settings, but lacks the granularity in assessing specific reasoning capabilities. We introduce GameArena, a dynamic benchmark designed to evaluate LLM reasoning capabilities through interactive gameplay with humans. GameArena consists of three games designed to test specific reasoning capabilities (e.g., deductive and inductive reasoning), while keeping participants entertained and engaged. We analyze the gaming data retrospectively to uncover the underlying reasoning processes of LLMs and measure their fine-grained reasoning capabilities. We collect over 2000 game sessions and provide detailed assessments of various reasoning capabilities for five state-of-the-art LLMs. Our user study with 100 participants suggests that GameArena improves user engagement compared to Chatbot Arena. For the first time, GameArena enables the collection of step-by-step LLM reasoning data in the wild.