Authors: Sebastian Jaskowski, Sahasra Chava, Agam Shah
Abstract: The BERTScore metric is commonly used to evaluate automatic text simplification systems. However, current implementations of the metric fail to provide complete visibility into all information the metric can produce. Notably, the specific token matchings can be incredibly useful in generating clause-level insight into the quality of simplified text. We address this by introducing BERTScoreVisualizer, a web application that goes beyond reporting precision, recall, and F1 score and provides a visualization of the matching between tokens. We believe that our software can help improve the analysis of text simplification systems by specifically showing where generated, simplified text deviates from reference text. We host our code and demo on GitHub.
Authors: Honggen Zhang, Igor Molybog, June Zhang, Xufeng Zhao
Abstract: Aligning large language models (LLMs) to human preferences is a crucial step in building helpful and safe AI tools, which usually involve training on supervised datasets. Popular algorithms such as Direct Preference Optimization rely on pairs of AI-generated responses ranked according to human feedback. The labeling process is the most labor-intensive and costly part of the alignment pipeline, and improving its efficiency would have a meaningful impact on AI development. We propose a strategy for sampling a high-quality training dataset that focuses on acquiring the most informative response pairs for labeling out of a set of AI-generated responses. Experimental results on synthetic HH-RLHF benchmarks indicate that choosing dissimilar response pairs enhances the direct alignment of LLMs while reducing inherited labeling errors. We also applied our method to the real-world dataset SHP2, selecting optimal pairs from multiple responses. The model aligned on dissimilar response pairs obtained the best win rate on the dialogue task. Our findings suggest that focusing on less similar pairs can improve the efficiency of LLM alignment, saving up to 65% of annotators' work.
Authors: Ankit Maloo Abhinav Garg
Abstract: Generating domain-specific content using small language models poses challenges, especially when dealing with multiple distinct datasets with minimal overlap. In this study, we explore methods to enable a small language model to produce coherent and relevant outputs for two different domains: stories (Dataset A) and recipes (Dataset B). Our initial experiments show that training individual models on each dataset yields satisfactory results, with each model generating appropriate content within its domain. We find that utilizing custom tokenizers tailored to each dataset significantly enhances generation quality compared to using a generic tokenizer. Attempts to adapt a single model to both domains using Low-Rank Adaptation (LoRA) or standard fine-tuning do not yield substantial results, often failing to produce meaningful outputs. Moreover, full fine-tuning without freezing the model's existing weights leads to catastrophic forgetting, where the model loses previously learned information and only retains knowledge from the new data. To overcome these challenges, we employ a knowledge expansion strategy: training only with additional parameters. This approach enables the model to generate both stories and recipes upon request, effectively handling multiple domains without suffering from catastrophic forgetting. Our findings demonstrate that knowledge expansion with frozen layers is an effective method for small language models to generate domain-specific content across distinct datasets. This work contributes to the development of efficient multi-domain language models and provides insights into managing catastrophic forgetting in small-scale architectures.
Authors: Shashidhar Reddy Javaji, Zining Zhu
Abstract: Large language models (LLMs) can store a massive amount of knowledge, yet their potential to acquire new knowledge remains unknown. We propose a novel evaluation framework that evaluates this capability. This framework prompts LLMs to generate questions about a statement introducing scientific knowledge, simulating a curious person when facing the statement for the first time. We score the qualities of the generated questions, thereby evaluating the knowledge acquisition potential of the LLM. We apply controlled ablation studies to validate our scoring procedures. Additionally, we created a synthetic dataset consisting of 1101 statements in physics, chemistry, and maths with distinct levels of difficulties, 300 general knowledge statements, and 567 incorrect statements. Human evaluations were conducted to validate our model assessments, achieving an approximate weighted Cohen's kappa of 0.7 on all three metrics considered. We find that while large models like GPT-4 and Mistral 8x7b are adept at generating coherent and relevant questions, the smaller Phi-2 model is equally or more effective. This indicates that size does not solely determine a model's knowledge acquisition potential. The proposed framework quantifies a critical model capability that was commonly overlooked and opens up research opportunities for developing more knowledgeable AI systems
Authors: Satoshi Munakata, Taku Fukui, Takao Mohri
Abstract: Large language models (LLMs) often fabricate a hallucinatory text. Several methods have been developed to detect such text by semantically comparing it with the multiple versions probabilistically regenerated. However, a significant issue is that if the storyline of each regenerated text changes, the generated texts become incomparable, which worsen detection accuracy. In this paper, we propose a hallucination detection method that incorporates a multiple-fill-in-the-blank exam approach to address this storyline-changing issue. First, our method creates a multiple-fill-in-the-blank exam by masking multiple objects from the original text. Second, prompts an LLM to repeatedly answer this exam. This approach ensures that the storylines of the exam answers align with the original ones. Finally, quantifies the degree of hallucination for each original sentence by scoring the exam answers, considering the potential for \emph{hallucination snowballing} within the original text itself. Experimental results show that our method alone not only outperforms existing methods, but also achieves clearer state-of-the-art performance in the ensembles with existing methods.
Authors: Kangsheng Wang, Xiao Zhang, Zizheng Guo, Tianyu Hu, Huimin Ma
Abstract: Chain-based reasoning methods like chain of thought (CoT) play a rising role in solving reasoning tasks for large language models (LLMs). However, the causal illusions between \textit{a step of reasoning} and \textit{corresponding state transitions} are becoming a significant obstacle to advancing LLMs' reasoning capabilities, especially in long-range reasoning tasks. This paper proposes a non-chain-based reasoning framework for simultaneous consideration of causal significance and consistency, i.e., the Causal Significance and Consistency Enhancer (CSCE). We customize LLM's loss function utilizing treatment effect assessments to enhance its reasoning ability from two aspects: causal significance and consistency. This ensures that the model captures essential causal relationships and maintains robust and consistent performance across various scenarios. Additionally, we transform the reasoning process from the cascading multiple one-step reasoning commonly used in Chain-Based methods, like CoT, to a causal-enhanced method that outputs the entire reasoning process in one go, further improving the model's reasoning efficiency. Extensive experiments show that our method improves both the reasoning success rate and speed. These improvements further demonstrate that non-chain-based methods can also aid LLMs in completing reasoning tasks.
Authors: Diego Marcos, Robert van de Vlasakker, Ioannis N. Athanasiadis, Pierre Bonnet, Herv\'e Goeau, Alexis Joly, W. Daniel Kissling, C\'esar Leblanc, Andr\'e S. J. van Proosdij, Konstantinos P. Panousis
Abstract: Plant morphological traits, their observable characteristics, are fundamental to understand the role played by each species within their ecosystem. However, compiling trait information for even a moderate number of species is a demanding task that may take experts years to accomplish. At the same time, massive amounts of information about species descriptions is available online in the form of text, although the lack of structure makes this source of data impossible to use at scale. To overcome this, we propose to leverage recent advances in large language models (LLMs) and devise a mechanism for gathering and processing information on plant traits in the form of unstructured textual descriptions, without manual curation. We evaluate our approach by automatically replicating three manually created species-trait matrices. Our method managed to find values for over half of all species-trait pairs, with an F1-score of over 75%. Our results suggest that large-scale creation of structured trait databases from unstructured online text is currently feasible thanks to the information extraction capabilities of LLMs, being limited by the availability of textual descriptions covering all the traits of interest.
Authors: Guanyi Mou, Kyumin Lee
Abstract: With the widespread online social networks, hate speeches are spreading faster and causing more damage than ever before. Existing hate speech detection methods have limitations in several aspects, such as handling data insufficiency, estimating model uncertainty, improving robustness against malicious attacks, and handling unintended bias (i.e., fairness). There is an urgent need for accurate, robust, and fair hate speech classification in online social networks. To bridge the gap, we design a data-augmented, fairness addressed, and uncertainty estimated novel framework. As parts of the framework, we propose Bidirectional Quaternion-Quasi-LSTM layers to balance effectiveness and efficiency. To build a generalized model, we combine five datasets collected from three platforms. Experiment results show that our model outperforms eight state-of-the-art methods under both no attack scenario and various attack scenarios, indicating the effectiveness and robustness of our model. We share our code along with combined dataset for better future research
Authors: Joshua Ashkinaze, Emily Fry, Narendra Edara, Eric Gilbert, Ceren Budak
Abstract: Recent debates raised concerns that language models may favor certain viewpoints. But what if the solution is not to aim for a 'view from nowhere' but rather to leverage different viewpoints? We introduce Plurals, a system and Python library for pluralistic AI deliberation. Plurals consists of Agents (LLMs, optionally with personas) which deliberate within customizable Structures, with Moderators overseeing deliberation. Plurals is a generator of simulated social ensembles. Plurals integrates with government datasets to create nationally representative personas, includes deliberation templates inspired by democratic deliberation theory, and allows users to customize both information-sharing structures and deliberation behavior within Structures. Six case studies demonstrate fidelity to theoretical constructs and efficacy. Three randomized experiments show simulated focus groups produced output resonant with an online sample of the relevant audiences (chosen over zero-shot generation in 75% of trials). Plurals is both a paradigm and a concrete system for pluralistic AI. The Plurals library is available at https://github.com/josh-ashkinaze/plurals and will be continually updated.
Authors: Jean-Loup Tastet, Inar Timiryasov
Abstract: We present BabyLlama-2, a 345 million parameter model distillation-pretrained from two teachers on a 10 million word corpus for the BabyLM competition. On BLiMP and SuperGLUE benchmarks, BabyLlama-2 outperforms baselines trained on both 10 and 100 million word datasets with the same data mix, as well as its teacher models. Through an extensive hyperparameter sweep, we demonstrate that the advantages of distillation cannot be attributed to suboptimal hyperparameter selection of the teachers. Our findings underscore the need for further investigation into distillation techniques, particularly in data-limited settings.
Authors: Yihong Liu, Mingyang Wang, Amir Hossein Kargaran, Ayyoob Imani, Orgest Xhelili, Haotian Ye, Chunlan Ma, Fran\c{c}ois Yvon, Hinrich Sch\"utze
Abstract: Recent studies have shown that post-aligning multilingual pretrained language models (mPLMs) using alignment objectives on both original and transliterated data can improve crosslingual alignment. This improvement further leads to better crosslingual transfer performance. However, it remains unclear how and why a better crosslingual alignment is achieved, as this technique only involves transliterations, and does not use any parallel data. This paper attempts to explicitly evaluate the crosslingual alignment and identify the key elements in transliteration-based approaches that contribute to better performance. For this, we train multiple models under varying setups for two pairs of related languages: (1) Polish and Ukrainian and (2) Hindi and Urdu. To assess alignment, we define four types of similarities based on sentence representations. Our experiments show that adding transliterations alone improves the overall similarities, even for random sentence pairs. With the help of auxiliary alignment objectives, especially the contrastive objective, the model learns to distinguish matched from random pairs, leading to better alignments. However, we also show that better alignment does not always yield better downstream performance, suggesting that further research is needed to clarify the connection between alignment and performance.
Authors: Robin Shing-Hei Yuen, Timothy Tin-Long Tse, Jian Zhu
Abstract: Current speech-based LLMs are predominantly trained on extensive ASR and TTS datasets, excelling in tasks related to these domains. However, their ability to handle direct speech-to-speech conversations remains notably constrained. These models often rely on an ASR-to-TTS chain-of-thought pipeline, converting speech into text for processing before generating audio responses, which introduces latency and loses audio features. We propose a method that implicitly internalizes ASR chain of thought into a speech LLM, enhancing its native speech understanding capabilities. Our approach reduces latency and improves the model's native understanding of speech, paving the way for more efficient and natural real-time audio interactions. We also release a large-scale synthetic conversational dataset to facilitate further research.
Authors: Hamidreza Amirzadeh, Sadegh Jafari, Anika Harju, Rob van der Goot
Abstract: Language typology databases enhance multi-lingual Natural Language Processing (NLP) by improving model adaptability to diverse linguistic structures. The widely-used lang2vec toolkit integrates several such databases, but its coverage remains limited at 28.9\%. Previous work on automatically increasing coverage predicts missing values based on features from other languages or focuses on single features, we propose to use textual data for better-informed feature prediction. To this end, we introduce a multi-lingual Part-of-Speech (POS) tagger, achieving over 70\% accuracy across 1,749 languages, and experiment with external statistical features and a variety of machine learning algorithms. We also introduce a more realistic evaluation setup, focusing on likely to be missing typology features, and show that our approach outperforms previous work in both setups.
Authors: Zhejian Zhou, Jiayu Wang, Dahua Lin, Kai Chen
Abstract: Though Large Language Models (LLMs) have shown remarkable abilities in mathematics reasoning, they are still struggling with performing numeric operations accurately, such as addition and multiplication. Numbers can be tokenized into tokens in various ways by different LLMs and affect the numeric operations performance. Currently, there are two representatives: 1) Tokenize into $1$-digit, and 2) Tokenize into $1\sim 3$ digit. The difference is roughly equivalent to using different numeral systems (namely base $10$ or base $10^{3}$). In light of this, we study the scaling behavior of different numeral systems in the context of transformer-based large language models. We empirically show that a base $10$ system is consistently more data-efficient than a base $10^{2}$ or $10^{3}$ system across training data scale, model sizes under from-scratch training settings, while different number systems have very similar fine-tuning performances. We attribute this to higher token frequencies of a base $10$ system. Additionally, we reveal extrapolation behavior patterns on addition and multiplication. We identify that base $100$ and base $1000$ systems struggle on token-level discernment and token-level operations. We also sheds light on the mechanism learnt by the models.
Authors: Konstantinos Skianis, John Pavlopoulos, A. Seza Do\u{g}ru\"oz
Abstract: Large Language Models (LLMs) are increasingly integrated into various medical fields, including mental health support systems. However, there is a gap in research regarding the effectiveness of LLMs in non-English mental health support applications. To address this problem, we present a novel multilingual adaptation of widely-used mental health datasets, translated from English into six languages (Greek, Turkish, French, Portuguese, German, and Finnish). This dataset enables a comprehensive evaluation of LLM performance in detecting mental health conditions and assessing their severity across multiple languages. By experimenting with GPT and Llama, we observe considerable variability in performance across languages, despite being evaluated on the same translated dataset. This inconsistency underscores the complexities inherent in multilingual mental health support, where language-specific nuances and mental health data coverage can affect the accuracy of the models. Through comprehensive error analysis, we emphasize the risks of relying exclusively on large language models (LLMs) in medical settings (e.g., their potential to contribute to misdiagnoses). Moreover, our proposed approach offers significant cost savings for multilingual tasks, presenting a major advantage for broad-scale implementation.
Authors: Dorsaf Sallami, Yuan-Chen Chang, Esma A\"imeur
Abstract: Fake news poses a significant threat to the integrity of information ecosystems and public trust. The advent of Large Language Models (LLMs) holds considerable promise for transforming the battle against fake news. Generally, LLMs represent a double-edged sword in this struggle. One major concern is that LLMs can be readily used to craft and disseminate misleading information on a large scale. This raises the pressing questions: Can LLMs easily generate biased fake news? Do all LLMs have this capability? Conversely, LLMs offer valuable prospects for countering fake news, thanks to their extensive knowledge of the world and robust reasoning capabilities. This leads to other critical inquiries: Can we use LLMs to detect fake news, and do they outperform typical detection models? In this paper, we aim to address these pivotal questions by exploring the performance of various LLMs. Our objective is to explore the capability of various LLMs in effectively combating fake news, marking this as the first investigation to analyze seven such models. Our results reveal that while some models adhere strictly to safety protocols, refusing to generate biased or misleading content, other models can readily produce fake news across a spectrum of biases. Additionally, our results show that larger models generally exhibit superior detection abilities and that LLM-generated fake news are less likely to be detected than human-written ones. Finally, our findings demonstrate that users can benefit from LLM-generated explanations in identifying fake news.
Authors: Chung-Chi Chen, Hen-Hsen Huang, Hsin-Hsi Chen, Hiroya Takamura, Ichiro Kobayashi, Yusuke Miyao
Abstract: In the era of rapid Internet and social media platform development, individuals readily share their viewpoints online. The overwhelming quantity of these posts renders comprehensive analysis impractical. This necessitates an efficient recommendation system to filter and present significant, relevant opinions. Our research introduces a dual-pronged argument mining technique to improve recommendation system effectiveness, considering both professional and amateur investor perspectives. Our first strategy involves using the discrepancy between target and closing prices as an opinion indicator. The second strategy applies argument mining principles to score investors' opinions, subsequently ranking them by these scores. Experimental results confirm the effectiveness of our approach, demonstrating its ability to identify opinions with higher profit potential. Beyond profitability, our research extends to risk analysis, examining the relationship between recommended opinions and investor behaviors. This offers a holistic view of potential outcomes following the adoption of these recommended opinions.
Authors: Chr-Jr Chiu, Chung-Chi Chen, Hen-Hsen Huang, Hsin-Hsi Chen
Abstract: Understanding the duration of news events' impact on the stock market is crucial for effective time-series forecasting, yet this facet is largely overlooked in current research. This paper addresses this research gap by introducing a novel dataset, the Impact Duration Estimation Dataset (IDED), specifically designed to estimate impact duration based on investor opinions. Our research establishes that pre-finetuning language models with IDED can enhance performance in text-based stock movement predictions. In addition, we juxtapose our proposed pre-finetuning task with sentiment analysis pre-finetuning, further affirming the significance of learning impact duration. Our findings highlight the promise of this novel research direction in stock movement prediction, offering a new avenue for financial forecasting. We also provide the IDED and pre-finetuned language models under the CC BY-NC-SA 4.0 license for academic use, fostering further exploration in this field.
Authors: Zhenmei Shi, Yifei Ming, Xuan-Phi Nguyen, Yingyu Liang, Shafiq Joty
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in handling long context inputs, but this comes at the cost of increased computational resources and latency. Our research introduces a novel approach for the long context bottleneck to accelerate LLM inference and reduce GPU memory consumption. Our research demonstrates that LLMs can identify relevant tokens in the early layers before generating answers to a query. Leveraging this insight, we propose an algorithm that uses early layers of an LLM as filters to select and compress input tokens, significantly reducing the context length for subsequent processing. Our method, GemFilter, demonstrates substantial improvements in both speed and memory efficiency compared to existing techniques, such as standard attention and SnapKV/H2O. Notably, it achieves a 2.4$\times$ speedup and 30\% reduction in GPU memory usage compared to SOTA methods. Evaluation on the Needle in a Haystack task shows that GemFilter significantly outperforms standard attention, SnapKV and demonstrates comparable performance on the LongBench challenge. GemFilter is simple, training-free, and broadly applicable across different LLMs. Crucially, it provides interpretability by allowing humans to inspect the selected input sequence. These findings not only offer practical benefits for LLM deployment, but also enhance our understanding of LLM internal mechanisms, paving the way for further optimizations in LLM design and inference. Our code is available at \url{https://github.com/SalesforceAIResearch/GemFilter}.
Authors: Jinghong Chen, Guangyu Yang, Weizhe Lin, Jingbiao Mei, Bill Byrne
Abstract: We derive and investigate two DPO variants that explicitly model the possibility of declaring a tie in pair-wise comparisons. We replace the Bradley-Terry model in DPO with two well-known modeling extensions, by Rao and Kupper and by Davidson, that assign probability to ties as alternatives to clear preferences. Our experiments in neural machine translation and summarization show that explicitly labeled ties can be added to the datasets for these DPO variants without the degradation in task performance that is observed when the same tied pairs are presented to DPO. We find empirically that the inclusion of ties leads to stronger regularization with respect to the reference policy as measured by KL divergence, and we see this even for DPO in its original form. These findings motivate and enable the inclusion of tied pairs in preference optimization as opposed to simply discarding them.
Authors: Wenlin Yao, Haitao Mi, Dong Yu
Abstract: Despite recent advancements in large language models (LLMs), their performance on complex reasoning problems requiring multi-step thinking and combining various skills is still limited. To address this, we propose a novel framework HDFlow for complex reasoning with LLMs that combines fast and slow thinking modes in an adaptive manner. Our approach consists of two key components: 1) a new approach for slow, deliberate reasoning called Dynamic Workflow, which automatically decomposes complex problems into more manageable sub-tasks and dynamically designs a workflow to assemble specialized LLM or symbolic reasoning tools to solve sub-tasks; 2) Hybrid Thinking, a general framework that dynamically combines fast and slow thinking based on problem complexity. Finally, we propose an easy-to-scale method for automatically synthesizing a large-scale dataset of 27K challenging reasoning problems for complex reasoning and a hybrid thinking tuning method that trains smaller LLMs on this dataset to internalize the fast/slow hybrid reasoning strategies. Experiments on four reasoning benchmark datasets demonstrate that our slow thinking with dynamic workflows significantly outperforms Chain-of-Thought, and hybrid thinking achieves the highest accuracy while providing an effective balance between computational efficiency and performance. Fine-tuning using our hybrid thinking approach also significantly boosts the complex reasoning capabilities of open-source language models. The results showcase the promise of slow thinking, dynamic workflows, and hybrid thinking in expanding the frontier of complex problem-solving with LLMs\footnote{Code and data will be released at \url{https://github.com/wenlinyao/HDFlow}.}.
Authors: Chung-Chi Chen, Hiroya Takamura, Ichiro Kobayashi, Yusuke Miyao
Abstract: This paper investigates the role of expert-designed hint in enhancing sentiment analysis on financial social media posts. We explore the capability of large language models (LLMs) to empathize with writer perspectives and analyze sentiments. Our findings reveal that expert-designed hint, i.e., pointing out the importance of numbers, significantly improve performances across various LLMs, particularly in cases requiring perspective-taking skills. Further analysis on tweets containing different types of numerical data demonstrates that the inclusion of expert-designed hint leads to notable improvements in sentiment analysis performance, especially for tweets with monetary-related numbers. Our findings contribute to the ongoing discussion on the applicability of Theory of Mind in NLP and open new avenues for improving sentiment analysis in financial domains through the strategic use of expert knowledge.
Authors: Yuqing Zhou, Ruixiang Tang, Ziyu Yao, Ziwei Zhu
Abstract: Language models (LMs), despite their advances, often depend on spurious correlations, undermining their accuracy and generalizability. This study addresses the overlooked impact of subtler, more complex shortcuts that compromise model reliability beyond oversimplified shortcuts. We introduce a comprehensive benchmark that categorizes shortcuts into occurrence, style, and concept, aiming to explore the nuanced ways in which these shortcuts influence the performance of LMs. Through extensive experiments across traditional LMs, large language models, and state-of-the-art robust models, our research systematically investigates models' resilience and susceptibilities to sophisticated shortcuts. Our benchmark and code can be found at: https://github.com/yuqing-zhou/shortcut-learning-in-text-classification.
URLs: https://github.com/yuqing-zhou/shortcut-learning-in-text-classification.
Authors: Sidney Gig-Jan Wong
Abstract: While NLP research into hate speech detection has grown exponentially in the last three decades, there has been minimal uptake or engagement from policy makers and non-profit organisations. We argue the absence of ethical frameworks have contributed to this rift between current practice and best practice. By adopting appropriate ethical frameworks, NLP researchers may enable the social impact potential of hate speech research. This position paper is informed by reviewing forty-eight hate speech detection systems associated with thirty-seven publications from different venues.
Authors: Heejin Do, Sangwon Ryu, Gary Geunbae Lee
Abstract: Recent advances in automated essay scoring (AES) have shifted towards evaluating multiple traits to provide enriched feedback. Like typical AES systems, multi-trait AES employs the quadratic weighted kappa (QWK) to measure agreement with human raters, aligning closely with the rating schema; however, its non-differentiable nature prevents its direct use in neural network training. In this paper, we propose Scoring-aware Multi-reward Reinforcement Learning (SaMRL), which integrates actual evaluation schemes into the training process by designing QWK-based rewards with a mean-squared error penalty for multi-trait AES. Existing reinforcement learning (RL) applications in AES are limited to classification models despite associated performance degradation, as RL requires probability distributions; instead, we adopt an autoregressive score generation framework to leverage token generation probabilities for robust multi-trait score predictions. Empirical analyses demonstrate that SaMRL facilitates model training, notably enhancing scoring of previously inferior prompts.
Authors: Guanyi Mou, Yichuan Li, Kyumin Lee
Abstract: Data augmentation has shown its effectiveness in resolving the data-hungry problem and improving model's generalization ability. However, the quality of augmented data can be varied, especially compared with the raw/original data. To boost deep learning models' performance given augmented data/samples in text classification tasks, we propose a novel framework, which leverages both meta learning and contrastive learning techniques as parts of our design for reweighting the augmented samples and refining their feature representations based on their quality. As part of the framework, we propose novel weight-dependent enqueue and dequeue algorithms to utilize augmented samples' weight/quality information effectively. Through experiments, we show that our framework can reasonably cooperate with existing deep learning models (e.g., RoBERTa-base and Text-CNN) and augmentation techniques (e.g., Wordnet and Easydata) for specific supervised learning tasks. Experiment results show that our framework achieves an average of 1.6%, up to 4.3% absolute improvement on Text-CNN encoders and an average of 1.4%, up to 4.4% absolute improvement on RoBERTa-base encoders on seven GLUE benchmark datasets compared with the best baseline. We present an indepth analysis of our framework design, revealing the non-trivial contributions of our network components. Our code is publicly available for better reproducibility.
Authors: Hao Liang, Keshi Zhao, Yajie Yang, Bin Cui, Guosheng Dong, Zenan Zhou, Wentao Zhang
Abstract: Large language models (LLMs) have demonstrated exceptional performance across a wide range of tasks and domains, with data preparation playing a critical role in achieving these results. Pre-training data typically combines information from multiple domains. To maximize performance when integrating data from various domains, determining the optimal data proportion is essential. However, state-of-the-art (SOTA) LLMs rarely disclose details about their pre-training data, making it difficult for researchers to identify ideal data proportions. In this paper, we introduce a new topic, \textit{data proportion detection}, which enables the automatic estimation of pre-training data proportions by analyzing the generated outputs of LLMs. We provide rigorous theoretical proofs, practical algorithms, and preliminary experimental results for data proportion detection. Based on these findings, we offer valuable insights into the challenges and future directions for effective data proportion detection and data management.
Authors: Pengjie Liu
Abstract: Knowledge Graph Completion (KGC) aims to predict the missing [relation] part of (head entity)--[relation]->(tail entity) triplet. Most existing KGC methods focus on single features (e.g., relation types) or sub-graph aggregation. However, they do not fully explore the Knowledge Graph (KG) features and neglect the guidance of external semantic knowledge. To address these shortcomings, we propose a knowledge-aware reasoning model (MUSE), which designs a novel multi-knowledge representation learning mechanism for missing relation prediction. Our model develops a tailored embedding space through three parallel components: 1) Prior Knowledge Learning for enhancing the triplets' semantic representation by fine-tuning BERT; 2) Context Message Passing for enhancing the context messages of KG; 3) Relational Path Aggregation for enhancing the path representation from the head entity to the tail entity. The experimental results show that MUSE significantly outperforms other baselines on four public datasets, achieving over 5.50% H@1 improvement and 4.20% MRR improvement on the NELL995 dataset. The code and datasets will be released via https://github.com/SUSTech-TP/ADMA2024-MUSE.git.
Authors: Tongxuan Liu, Wenjiang Xu, Weizhe Huang, Xingyu Wang, Jiaxing Wang, Hailong Yang, Jing Li
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks but their performance in complex logical reasoning tasks remains unsatisfactory. Although some prompting methods, such as Chain-of-Thought, can improve the reasoning ability of LLMs to some extent, they suffer from an unfaithful issue where derived conclusions may not align with the generated reasoning chain. To address this issue, some studies employ the approach of propositional logic to further enhance logical reasoning abilities of LLMs. However, the potential omissions in the extraction of logical expressions in these methods can cause information loss in the logical reasoning process, thereby generating incorrect results. To this end, we propose Logic-of-Thought (LoT) prompting which employs propositional logic to generate expanded logical information from input context, and utilizes the generated logical information as an additional augmentation to the input prompts, thereby enhancing the capability of logical reasoning. The LoT is orthogonal to existing prompting methods and can be seamlessly integrated with them. Extensive experiments demonstrate that LoT boosts the performance of various prompting methods with a striking margin across five logical reasoning tasks. In particular, the LoT enhances Chain-of-Thought's performance on the ReClor dataset by +4.35%; moreover, it improves Chain-of-Thought with Self-Consistency's performance on LogiQA by +5%; additionally, it boosts performance of Tree-of-Thoughts on ProofWriter dataset by +8%.
Authors: Cheolhun Jang
Abstract: Preference optimization methods typically begin training with a well-trained SFT model as a reference model. In RLHF and DPO, a regularization term is used during the preference optimization process to prevent the policy model from deviating too far from the reference model's distribution, thereby avoiding the generation of anomalous responses. When the reference model is already well-aligned with the given data or only requires slight adjustments, this approach can produce a well-aligned model. However, if the reference model is not aligned with the given data and requires significant deviation from its current state, a regularization term may actually hinder the model alignment. In this study, we propose \textbf{Modulated Intervention Preference Optimization (MIPO)} to address this issue. MIPO modulates the degree of intervention from the reference model based on how well the given data is aligned with it. If the data is well-aligned, the intervention is increased to prevent the policy model from diverging significantly from reference model. Conversely, if the alignment is poor, the interference is reduced to facilitate more extensive training. We compare the performance of MIPO and DPO using Mistral-7B and Llama3-8B in Alpaca Eval 2.0 and MT-Bench. The experimental results demonstrate that MIPO consistently outperforms DPO across various evaluation scenarios.
Authors: Jin Xu, Mari\"et Theune, Daniel Braun
Abstract: It is common practice in text classification to only use one majority label for model training even if a dataset has been annotated by multiple annotators. Doing so can remove valuable nuances and diverse perspectives inherent in the annotators' assessments. This paper proposes and compares three different strategies to leverage annotator disagreement for text classification: a probability-based multi-label method, an ensemble system, and instruction tuning. All three approaches are evaluated on the tasks of hate speech and abusive conversation detection, which inherently entail a high degree of subjectivity. Moreover, to evaluate the effectiveness of embracing annotation disagreements for model training, we conduct an online survey that compares the performance of the multi-label model against a baseline model, which is trained with the majority label. The results show that in hate speech detection, the multi-label method outperforms the other two approaches, while in abusive conversation detection, instruction tuning achieves the best performance. The results of the survey also show that the outputs from the multi-label models are considered a better representation of the texts than the single-label model.
Authors: Fuqiang Niu, Minghuan Tan, Bowen Zhang, Min Yang, Ruifeng Xu
Abstract: Idioms represent a ubiquitous vehicle for conveying sentiments in the realm of everyday discourse, rendering the nuanced analysis of idiom sentiment crucial for a comprehensive understanding of emotional expression within real-world texts. Nevertheless, the existing corpora dedicated to idiom sentiment analysis considerably limit research in text sentiment analysis. In this paper, we propose an innovative approach to automatically expand the sentiment lexicon for idioms, leveraging the capabilities of large language models through the application of Chain-of-Thought prompting. To demonstrate the effectiveness of this approach, we integrate multiple existing resources and construct an emotional idiom lexicon expansion dataset (called EmoIdiomE), which encompasses a comprehensive repository of Chinese and English idioms. Then we designed the Dual Chain-of-Thoughts (DualCoTs) method, which combines insights from linguistics and psycholinguistics, to demonstrate the effectiveness of using large models to automatically expand the sentiment lexicon for idioms. Experiments show that DualCoTs is effective in idioms sentiment lexicon expansion in both Chinese and English. For reproducibility, we will release the data and code upon acceptance.
Authors: Shifu Xiong, Mengzhi Wang, Genshun Wan, Hang Chen, Jianqing Gao, Lirong Dai
Abstract: Contextual-LAS (CLAS) has been shown effective in improving Automatic Speech Recognition (ASR) of rare words. It relies on phrase-level contextual modeling and attention-based relevance scoring without explicit contextual constraint which lead to insufficient use of contextual information. In this work, we propose deep CLAS to use contextual information better. We introduce bias loss forcing model to focus on contextual information. The query of bias attention is also enriched to improve the accuracy of the bias attention score. To get fine-grained contextual information, we replace phrase-level encoding with character-level encoding and encode contextual information with conformer rather than LSTM. Moreover, we directly use the bias attention score to correct the output probability distribution of the model. Experiments using the public AISHELL-1 and AISHELL-NER. On AISHELL-1, compared to CLAS baselines, deep CLAS obtains a 65.78% relative recall and a 53.49% relative F1-score increase in the named entity recognition scene.
Authors: Zhangpu Li, Changhong Zou, Suxue Ma, Zhicheng Yang, Chen Du, Youbao Tang, Zhenjie Cao, Ning Zhang, Jui-Hsin Lai, Ruei-Sung Lin, Yuan Ni, Xingzhi Sun, Jing Xiao, Kai Zhang, Mei Han
Abstract: The rocketing prosperity of large language models (LLMs) in recent years has boosted the prevalence of vision-language models (VLMs) in the medical sector. In our online medical consultation scenario, a doctor responds to the texts and images provided by a patient in multiple rounds to diagnose her/his health condition, forming a multi-turn multimodal medical dialogue format. Unlike high-quality images captured by professional equipment in traditional medical visual question answering (Med-VQA), the images in our case are taken by patients' mobile phones. These images have poor quality control, with issues such as excessive background elements and the lesion area being significantly off-center, leading to degradation of vision-language alignment in the model training phase. In this paper, we propose ZALM3, a Zero-shot strategy to improve vision-language ALignment in Multi-turn Multimodal Medical dialogue. Since we observe that the preceding text conversations before an image can infer the regions of interest (RoIs) in the image, ZALM3 employs an LLM to summarize the keywords from the preceding context and a visual grounding model to extract the RoIs. The updated images eliminate unnecessary background noise and provide more effective vision-language alignment. To better evaluate our proposed method, we design a new subjective assessment metric for multi-turn unimodal/multimodal medical dialogue to provide a fine-grained performance comparison. Our experiments across three different clinical departments remarkably demonstrate the efficacy of ZALM3 with statistical significance.
Authors: Xindi Tong, Yujin Zhu, Shijian Fan, Liang Xu
Abstract: Long text summarization, gradually being essential for efficiently processing large volumes of information, stays challenging for Large Language Models (LLMs) such as GPT and LLaMA families because of the insufficient open-sourced training datasets and the high requirement of contextual details dealing. To address the issue, we design a novel zero-shot transfer learning framework, abbreviated as T3, to iteratively training a baseline LLM on an assistant task for the target task, where the former should own richer data resources and share structural or semantic similarity with the latter. In practice, T3 is approached to deal with the long text summarization task by utilizing question answering as the assistant task, and further validated its effectiveness on the BBC summary, NarraSum, FairytaleQA, and NLQuAD datasets, with up to nearly 14% improvement in ROUGE, 35% improvement in BLEU, and 16% improvement in Factscore compared to three baseline LLMs, demonstrating its potential for more assistant-target task combinations.
Authors: Isaac Chung, Phat Vo, Arman Kizilkale, Aaron Reite
Abstract: Retrieval Augmented Generation (RAG) is a common method for integrating external knowledge into pretrained Large Language Models (LLMs) to enhance accuracy and relevancy in question answering (QA) tasks. However, prompt engineering and resource efficiency remain significant bottlenecks in developing optimal and robust RAG solutions for real-world QA applications. Recent studies have shown success in using fine tuning to address these problems; in particular, Retrieval Augmented Fine Tuning (RAFT) applied to smaller 7B models has demonstrated superior performance compared to RAG setups with much larger models such as GPT-3.5. The combination of RAFT with parameter-efficient fine tuning (PEFT) techniques, such as Low-Rank Adaptation (LoRA), promises an even more efficient solution, yet remains an unexplored area. In this work, we combine RAFT with LoRA to reduce fine tuning and storage requirements and gain faster inference times while maintaining comparable RAG performance. This results in a more compute-efficient RAFT, or CRAFT, which is particularly useful for knowledge-intensive QA tasks in resource-constrained environments where internet access may be restricted and hardware resources limited.
Authors: Kaden Uhlig, Joern Wuebker, Raphael Reinauer, John DeNero
Abstract: Reinforcement Learning from Human Feedback (RLHF) and derivative techniques like Direct Preference Optimization (DPO) are task-alignment algorithms used to repurpose general, foundational models for specific tasks. We show that applying task-alignment to neural machine translation (NMT) addresses an existing task--data mismatch in NMT, leading to improvements across all languages of a multilingual model, even when task-alignment is only applied to a subset of those languages. We do so by introducing Direct Quality Optimization (DQO), a variant of DPO leveraging a pre-trained translation quality estimation model as a proxy for human preferences, and verify the improvements with both automatic metrics and human evaluation.
Authors: Natthanaphop Isaradech, Andrea Riedel, Wachiranun Sirikul, Markus Kreuzthaler, Stefan Schulz
Abstract: Introduction: Medication prescriptions are often in free text and include a mix of two languages, local brand names, and a wide range of idiosyncratic formats and abbreviations. Large language models (LLMs) have shown promising ability to generate text in response to input prompts. We use ChatGPT 3.5 to automatically structure and expand medication statements in discharge summaries and thus make them easier to interpret for people and machines. Methods: Named-entity Recognition (NER) and Text Expansion (EX) are used in a zero- and few-shot setting with different prompt strategies. 100 medication statements were manually annotated and curated. NER performance was measured by using strict and partial matching. For the task EX, two experts interpreted the results by assessing semantic equivalence between original and expanded statements. The model performance was measured by precision, recall, and F1 score. Results: For NER, the best-performing prompt reached an average F1 score of 0.94 in the test set. For EX, the few-shot prompt showed superior performance among other prompts, with an average F1 score of 0.87. Conclusion: Our study demonstrates good performance for NER and EX tasks in free-text medication statements using ChatGPT. Compared to a zero-shot baseline, a few-shot approach prevented the system from hallucinating, which would be unacceptable when processing safety-relevant medication data.
Authors: Zekun Wang, King Zhu, Chunpu Xu, Wangchunshu Zhou, Jiaheng Liu, Yibo Zhang, Jiashuo Wang, Ning Shi, Siyu Li, Yizhi Li, Haoran Que, Zhaoxiang Zhang, Yuanxing Zhang, Ge Zhang, Ke Xu, Jie Fu, Wenhao Huang
Abstract: In this paper, we introduce MIO, a novel foundation model built on multimodal tokens, capable of understanding and generating speech, text, images, and videos in an end-to-end, autoregressive manner. While the emergence of large language models (LLMs) and multimodal large language models (MM-LLMs) propels advancements in artificial general intelligence through their versatile capabilities, they still lack true any-to-any understanding and generation. Recently, the release of GPT-4o has showcased the remarkable potential of any-to-any LLMs for complex real-world tasks, enabling omnidirectional input and output across images, speech, and text. However, it is closed-source and does not support the generation of multimodal interleaved sequences. To address this gap, we present MIO, which is trained on a mixture of discrete tokens across four modalities using causal multimodal modeling. MIO undergoes a four-stage training process: (1) alignment pre-training, (2) interleaved pre-training, (3) speech-enhanced pre-training, and (4) comprehensive supervised fine-tuning on diverse textual, visual, and speech tasks. Our experimental results indicate that MIO exhibits competitive, and in some cases superior, performance compared to previous dual-modal baselines, any-to-any model baselines, and even modality-specific baselines. Moreover, MIO demonstrates advanced capabilities inherent to its any-to-any feature, such as interleaved video-text generation, chain-of-visual-thought reasoning, visual guideline generation, instructional image editing, etc.
Authors: Qin Wang, Jianzhou Feng, Yiming Xu
Abstract: Manifestly and logically displaying the line of reasoning from evidence to answer is significant to explainable question answering (QA). The entailment tree exhibits the lines structurally, which is different from the self-explanation principle in large-scale language models. Existing methods rarely consider the semantic association of sentences between and within hierarchies within the tree structure, which is prone to apparent mistakes in combinations. In this work, we propose an architecture of integrating the Hierarchical Semantics of sentences under the framework of Controller-Generator (HiSCG) to explain answers. The HiSCG designs a hierarchical mapping between hypotheses and facts, discriminates the facts involved in tree constructions, and optimizes single-step entailments. To the best of our knowledge, We are the first to notice hierarchical semantics of sentences between the same layer and adjacent layers to yield improvements. The proposed method achieves comparable performance on all three settings of the EntailmentBank dataset. The generalization results on two out-of-domain datasets also demonstrate the effectiveness of our method.
Authors: Supriya Manna, Niladri Sett
Abstract: Faithfulness is arguably the most critical metric to assess the reliability of explainable AI. In NLP, current methods for faithfulness evaluation are fraught with discrepancies and biases, often failing to capture the true reasoning of models. We introduce Adversarial Sensitivity as a novel approach to faithfulness evaluation, focusing on the explainer's response when the model is under adversarial attack. Our method accounts for the faithfulness of explainers by capturing sensitivity to adversarial input changes. This work addresses significant limitations in existing evaluation techniques, and furthermore, quantifies faithfulness from a crucial yet underexplored paradigm.
Authors: Jian Li, Haojing Huang, Yujia Zhang, Pengfei Xu, Xi Chen, Rui Song, Lida Shi, Jingwen Wang, Hao Xu
Abstract: Recently, there has been significant interest in replacing the reward model in Reinforcement Learning with Human Feedback (RLHF) methods for Large Language Models (LLMs), such as Direct Preference Optimization (DPO) and its variants. These approaches commonly use a binary cross-entropy mechanism on pairwise samples, i.e., minimizing and maximizing the loss based on preferred or dis-preferred responses, respectively. However, while this training strategy omits the reward model, it also overlooks the varying preference degrees within different responses. We hypothesize that this is a key factor hindering LLMs from sufficiently understanding human preferences. To address this problem, we propose a novel Self-supervised Preference Optimization (SPO) framework, which constructs a self-supervised preference degree loss combined with the alignment loss, thereby helping LLMs improve their ability to understand the degree of preference. Extensive experiments are conducted on two widely used datasets of different tasks. The results demonstrate that SPO can be seamlessly integrated with existing preference optimization methods and significantly boost their performance to achieve state-of-the-art performance. We also conduct detailed analyses to offer comprehensive insights into SPO, which verifies its effectiveness. The code is available at https://github.com/lijian16/SPO.
Authors: Zhixuan Liu, Zhanhui Zhou, Yuanfu Wang, Chao Yang, Yu Qiao
Abstract: Large language models are typically fine-tuned to align with human preferences, but tuning large models is computationally intensive and complex. In this work, we introduce $\textit{Integrated Value Guidance}$ (IVG), a method that uses implicit and explicit value functions to guide language model decoding at token and chunk-level respectively, efficiently aligning large language models purely at inference time. This approach circumvents the complexities of direct fine-tuning and outperforms traditional methods. Empirically, we demonstrate the versatility of IVG across various tasks. In controlled sentiment generation and summarization tasks, our method significantly improves the alignment of large models using inference-time guidance from $\texttt{gpt2}$-based value functions. Moreover, in a more challenging instruction-following benchmark AlpacaEval 2.0, we show that both specifically tuned and off-the-shelf value functions greatly improve the length-controlled win rates of large models against $\texttt{gpt-4-turbo}$ (e.g., $19.51\% \rightarrow 26.51\%$ for $\texttt{Mistral-7B-Instruct-v0.2}$ and $25.58\% \rightarrow 33.75\%$ for $\texttt{Mixtral-8x7B-Instruct-v0.1}$ with Tulu guidance).
Authors: Siyan Wang, Bradford Levy
Abstract: Many of the recent breakthroughs in language modeling have resulted from scaling effectively the same model architecture to larger datasets. In this vein, recent work has highlighted performance gains from increasing training dataset size and quality, suggesting a need for novel sources of large-scale datasets. In this work, we introduce BeanCounter, a public dataset consisting of more than 159B tokens extracted from businesses' disclosures. We show that this data is indeed novel: less than 0.1% of BeanCounter appears in Common Crawl-based datasets and it is an order of magnitude larger than datasets relying on similar sources. Given the data's provenance, we hypothesize that BeanCounter is comparatively more factual and less toxic than web-based datasets. Exploring this hypothesis, we find that many demographic identities occur with similar prevalence in BeanCounter but with significantly less toxic context relative to other datasets. To demonstrate the utility of BeanCounter, we evaluate and compare two LLMs continually pre-trained on BeanCounter with their base models. We find an 18-33% reduction in toxic generation and improved performance within the finance domain for the continually pretrained models. Collectively, our work suggests that BeanCounter is a novel source of low-toxicity and high-quality domain-specific data with sufficient scale to train multi-billion parameter LLMs.
Authors: Tianfang Xie, Tianjing Li, Wei Zhu, Wei Han, Yi Zhao
Abstract: Due to their substantial sizes, large language models (LLMs) are typically deployed within a single-backbone multi-tenant framework. In this setup, a single instance of an LLM backbone must cater to multiple users or tasks through the application of various parameter-efficient fine-tuning (PEFT) models. Despite the availability of numerous effective PEFT techniques such as LoRA, there remains a need for a PEFT approach that achieves both high efficiency during inference and competitive performance on downstream tasks. In this research, we introduce a new and straightforward PEFT methodology named \underline{P}rompt D\underline{E}pen\underline{D}ent \underline{R}epresentation M\underline{O}dification (PEDRO). The proposed method involves integrating a lightweight vector generator into each Transformer layer, which generates vectors contingent upon the input prompts. These vectors then modify the hidden representations created by the LLM through a dot product operation, thereby influencing the semantic output and generated content of the model. Extensive experimentation across a variety of tasks indicates that: (a) PEDRO surpasses recent PEFT benchmarks when using a similar number of tunable parameters. (b) Under the single-backbone multi-tenant deployment model, PEDRO exhibits superior efficiency compared to LoRA, indicating significant industrial potential.
Authors: Shaoxiong Ji, Zihao Li, Indraneil Paul, Jaakko Paavola, Peiqin Lin, Pinzhen Chen, Dayy\'an O'Brien, Hengyu Luo, Hinrich Sch\"utze, J\"org Tiedemann, Barry Haddow
Abstract: In this work, we introduce EMMA-500, a large-scale multilingual language model continue-trained on texts across 546 languages designed for enhanced multilingual performance, focusing on improving language coverage for low-resource languages. To facilitate continual pre-training, we compile the MaLA corpus, a comprehensive multilingual dataset enriched with curated datasets across diverse domains. Leveraging this corpus, we conduct extensive continual pre-training of the Llama 2 7B model, resulting in EMMA-500, which demonstrates robust performance across a wide collection of benchmarks, including a comprehensive set of multilingual tasks and PolyWrite, an open-ended generation benchmark developed in this study. Our results highlight the effectiveness of continual pre-training in expanding large language models' language capacity, particularly for underrepresented languages, demonstrating significant gains in cross-lingual transfer, task generalization, and language adaptability.
Authors: Guokan Shang, Hadi Abdine, Yousef Khoubrane, Amr Mohamed, Yassine Abbahaddou, Sofiane Ennadir, Imane Momayiz, Xuguang Ren, Eric Moulines, Preslav Nakov, Michalis Vazirgiannis, Eric Xing
Abstract: We introduce Atlas-Chat, the first-ever collection of large language models specifically developed for dialectal Arabic. Focusing on Moroccan Arabic, also known as Darija, we construct our instruction dataset by consolidating existing Darija language resources, creating novel datasets both manually and synthetically, and translating English instructions with stringent quality control. Atlas-Chat-9B and 2B models, fine-tuned on the dataset, exhibit superior ability in following Darija instructions and performing standard NLP tasks. Notably, our models outperform both state-of-the-art and Arabic-specialized LLMs like LLaMa, Jais, and AceGPT, e.g., achieving a 13% performance boost over a larger 13B model on DarijaMMLU, in our newly introduced evaluation suite for Darija covering both discriminative and generative tasks. Furthermore, we perform an experimental analysis of various fine-tuning strategies and base model choices to determine optimal configurations. All our resources are publicly accessible, and we believe our work offers comprehensive design methodologies of instruction-tuning for low-resource language variants, which are often neglected in favor of data-rich languages by contemporary LLMs.
Authors: Hengrui Gu, Kaixiong Zhou, Yili Wang, Ruobing Wang, Xin Wang
Abstract: During pre-training, the Text-to-Image (T2I) diffusion models encode factual knowledge into their parameters. These parameterized facts enable realistic image generation, but they may become obsolete over time, thereby misrepresenting the current state of the world. Knowledge editing techniques aim to update model knowledge in a targeted way. However, facing the dual challenges posed by inadequate editing datasets and unreliable evaluation criterion, the development of T2I knowledge editing encounter difficulties in effectively generalizing injected knowledge. In this work, we design a T2I knowledge editing framework by comprehensively spanning on three phases: First, we curate a dataset \textbf{CAKE}, comprising paraphrase and multi-object test, to enable more fine-grained assessment on knowledge generalization. Second, we propose a novel criterion, \textbf{adaptive CLIP threshold}, to effectively filter out false successful images under the current criterion and achieve reliable editing evaluation. Finally, we introduce \textbf{MPE}, a simple but effective approach for T2I knowledge editing. Instead of tuning parameters, MPE precisely recognizes and edits the outdated part of the conditioning text-prompt to accommodate the up-to-date knowledge. A straightforward implementation of MPE (Based on in-context learning) exhibits better overall performance than previous model editors. We hope these efforts can further promote faithful evaluation of T2I knowledge editing methods.
Authors: Andreas Waldis, Joel Birrer, Anne Lauscher, Iryna Gurevych
Abstract: Gender-fair language, an evolving German linguistic variation, fosters inclusion by addressing all genders or using neutral forms. Nevertheless, there is a significant lack of resources to assess the impact of this linguistic shift on classification using language models (LMs), which are probably not trained on such variations. To address this gap, we present Lou, the first dataset featuring high-quality reformulations for German text classification covering seven tasks, like stance detection and toxicity classification. Evaluating 16 mono- and multi-lingual LMs on Lou shows that gender-fair language substantially impacts predictions by flipping labels, reducing certainty, and altering attention patterns. However, existing evaluations remain valid, as LM rankings of original and reformulated instances do not significantly differ. While we offer initial insights on the effect on German text classification, the findings likely apply to other languages, as consistent patterns were observed in multi-lingual and English LMs.
Authors: Richard Yue, John E. Ortega
Abstract: Translation memories (TMs) are the backbone for professional translation tools called computer-aided translation (CAT) tools. In order to perform a translation using a CAT tool, a translator uses the TM to gather translations similar to the desired segment to translate (s'). Many CAT tools offer a fuzzy-match algorithm to locate segments (s) in the TM that are close in distance to s'. After locating two similar segments, the CAT tool will present parallel segments (s, t) that contain one segment in the source language along with its translation in the target language. Additionally, CAT tools contain fuzzy-match repair (FMR) techniques that will automatically use the parallel segments from the TM to create new TM entries containing a modified version of the original with the idea in mind that it will be the translation of s'. Most FMR techniques use machine translation as a way of "repairing" those words that have to be modified. In this article, we show that for a large part of those words which are anchored, we can use other techniques that are based on machine learning approaches such as Word2Vec. BERT, and even ChatGPT. Specifically, we show that for anchored words that follow the continuous bag-of-words (CBOW) paradigm, Word2Vec, BERT, and GPT-4 can be used to achieve similar and, for some cases, better results than neural machine translation for translating anchored words from French to English.
Authors: Richard Yue, John E. Ortega, Kenneth Ward Church
Abstract: The typical workflow for a professional translator to translate a document from its source language (SL) to a target language (TL) is not always focused on what many language models in natural language processing (NLP) do - predict the next word in a series of words. While high-resource languages like English and French are reported to achieve near human parity using common metrics for measurement such as BLEU and COMET, we find that an important step is being missed: the translation of technical terms, specifically acronyms. Some state-of-the art machine translation systems like Google Translate which are publicly available can be erroneous when dealing with acronyms - as much as 50% in our findings. This article addresses acronym disambiguation for MT systems by proposing an additional step to the SL-TL (FR-EN) translation workflow where we first offer a new acronym corpus for public consumption and then experiment with a search-based thresholding algorithm that achieves nearly 10% increase when compared to Google Translate and OpusMT.
Authors: Amita Kamath, Cheng-Yu Hsieh, Kai-Wei Chang, Ranjay Krishna
Abstract: Several benchmarks have concluded that our best vision-language models (e.g., CLIP) are lacking in compositionality. Given an image, these benchmarks probe a model's ability to identify its associated caption amongst a set of compositional distractors. In response, a surge of recent proposals show improvements by finetuning CLIP with distractors as hard negatives. Our investigations reveal that these improvements have, in fact, been significantly overstated -- because existing benchmarks do not probe whether finetuned vision-language models remain invariant to hard positives. By curating an evaluation dataset with 112,382 hard negatives and hard positives, we uncover that including hard positives decreases CLIP's performance by 12.9%, while humans perform effortlessly at 99%. CLIP finetuned with hard negatives results in an even larger decrease, up to 38.7%. With this finding, we then produce a 1,775,259 image-text training set with both hard negative and hard positive captions. By training with both, we see improvements on existing benchmarks while simultaneously improving performance on hard positives, indicating a more robust improvement in compositionality. Our work suggests the need for future research to rigorously test and improve CLIP's understanding of semantic relationships between related "positive" concepts.
Authors: Linzhuang Sun, Hao Liang, Wentao Zhang
Abstract: Large Language Models (LLMs) have exhibited exceptional performance across a broad range of tasks and domains. However, they still encounter difficulties in solving mathematical problems due to the rigorous and logical nature of mathematics. Previous studies have employed techniques such as supervised fine-tuning (SFT), prompt engineering, and search-based methods to improve the mathematical problem-solving abilities of LLMs. Despite these efforts, their performance remains suboptimal and demands substantial computational resources. To address this issue, we propose a novel approach, BEATS, to enhance mathematical problem-solving abilities. Our method leverages newly designed prompts that guide the model to iteratively rewrite, advance by one step, and generate answers based on previous steps. Additionally, we introduce a new back-verification technique that uses LLMs to validate the correctness of the generated answers. Furthermore, we employ a pruning tree search to optimize search time while achieving strong performance. Notably, our method improves Qwen2-7b-Instruct's score from 36.94 to 61.52, outperforming GPT4's 42.5 on the MATH benchmark.
Authors: Ameeta Agrawal, Andy Dang, Sina Bagheri Nezhad, Rhitabrat Pokharel, Russell Scheinberg
Abstract: Recent large language models (LLMs) demonstrate impressive capabilities in handling long contexts, some exhibiting near-perfect recall on synthetic retrieval tasks. However, these evaluations have mainly focused on English text and involved a single target sentence within lengthy contexts. Our work investigates how LLM performance generalizes to multilingual settings with multiple hidden target sentences. We comprehensively evaluate several long-context LLMs on retrieval and reasoning tasks across five languages: English, Vietnamese, Indonesian, Swahili, and Somali. These languages share the Latin script but belong to distinct language families and resource levels. Our analysis reveals a significant performance gap between languages. The best-performing models such as Gemini-1.5 and GPT-4o, achieve around 96% accuracy in English to around 36% in Somali with a single target sentence. However, this accuracy drops to 40% in English and 0% in Somali when dealing with three target sentences. Our findings highlight the challenges long-context LLMs face when processing longer contexts, an increase in the number of target sentences, or languages of lower resource levels.
Authors: Hannah Sterz, Jonas Pfeiffer, Ivan Vuli\'c
Abstract: Vision Language Models (VLMs) extend remarkable capabilities of text-only large language models and vision-only models, and are able to learn from and process multi-modal vision-text input. While modern VLMs perform well on a number of standard image classification and image-text matching tasks, they still struggle with a number of crucial vision-language (VL) reasoning abilities such as counting and spatial reasoning. Moreover, while they might be very brittle to small variations in instructions and/or evaluation protocols, existing benchmarks fail to evaluate their robustness (or rather the lack of it). In order to couple challenging VL scenarios with comprehensive robustness evaluation, we introduce DARE, Diverse Visual Question Answering with Robustness Evaluation, a carefully created and curated multiple-choice VQA benchmark. DARE evaluates VLM performance on five diverse categories and includes four robustness-oriented evaluations based on the variations of: prompts, the subsets of answer options, the output format and the number of correct answers. Among a spectrum of other findings, we report that state-of-the-art VLMs still struggle with questions in most categories and are unable to consistently deliver their peak performance across the tested robustness evaluations. The worst case performance across the subsets of options is up to 34% below the performance in the standard case. The robustness of the open-source VLMs such as LLaVA 1.6 and Idefics2 cannot match the closed-source models such as GPT-4 and Gemini, but even the latter remain very brittle to different variations.
Authors: Sahil Garje
Abstract: Power words are terms that evoke strong emotional responses and significantly influence readers' behavior, playing a crucial role in fields like marketing, politics, and motivational writing. This study proposes a methodology for the automated detection and analysis of power words in persuasive text using a custom lexicon and the TextBlob library in Python. By identifying the presence and frequency of power words within a given text, we aim to classify and analyze their impact on sentiment and reader engagement. This research examines diverse datasets across various domains to provide insights into the effectiveness of power words, offering practical applications for content creators, advertisers, and policymakers.
Authors: Belen Alastruey, Gerard I. G\'allego, Marta R. Costa-juss\`a
Abstract: Direct speech-to-text translation systems encounter an important drawback in data scarcity. A common solution consists on pretraining the encoder on automatic speech recognition, hence losing efficiency in the training process. In this study, we compare the training dynamics of a system using a pretrained encoder, the conventional approach, and one trained from scratch. We observe that, throughout the training, the randomly initialized model struggles to incorporate information from the speech inputs for its predictions. Hence, we hypothesize that this issue stems from the difficulty of effectively training an encoder for direct speech translation. While a model trained from scratch needs to learn acoustic and semantic modeling simultaneously, a pretrained one can just focus on the latter. Based on these findings, we propose a subtle change in the decoder cross-attention to integrate source information from earlier steps in training. We show that with this change, the model trained from scratch can achieve comparable performance to the pretrained one, while reducing the training time.
Authors: Hung-Ting Chen, Eunsol Choi
Abstract: We study retrieving a set of documents that covers various perspectives on a complex and contentious question (e.g., will ChatGPT do more harm than good?). We curate a Benchmark for Retrieval Diversity for Subjective questions (BERDS), where each example consists of a question and diverse perspectives associated with the question, sourced from survey questions and debate websites. On this data, retrievers paired with a corpus are evaluated to surface a document set that contains diverse perspectives. Our framing diverges from most retrieval tasks in that document relevancy cannot be decided by simple string matches to references. Instead, we build a language model based automatic evaluator that decides whether each retrieved document contains a perspective. This allows us to evaluate the performance of three different types of corpus (Wikipedia, web snapshot, and corpus constructed on the fly with retrieved pages from the search engine) paired with retrievers. Retrieving diverse documents remains challenging, with the outputs from existing retrievers covering all perspectives on only 33.74% of the examples. We further study the impact of query expansion and diversity-focused reranking approaches and analyze retriever sycophancy. Together, we lay the foundation for future studies in retrieval diversity handling complex queries.
Authors: Guobin Shen, Dongcheng Zhao, Aorigele Bao, Xiang He, Yiting Dong, Yi Zeng
Abstract: Human beings often experience stress, which can significantly influence their performance. This study explores whether Large Language Models (LLMs) exhibit stress responses similar to those of humans and whether their performance fluctuates under different stress-inducing prompts. To investigate this, we developed a novel set of prompts, termed StressPrompt, designed to induce varying levels of stress. These prompts were derived from established psychological frameworks and carefully calibrated based on ratings from human participants. We then applied these prompts to several LLMs to assess their responses across a range of tasks, including instruction-following, complex reasoning, and emotional intelligence. The findings suggest that LLMs, like humans, perform optimally under moderate stress, consistent with the Yerkes-Dodson law. Notably, their performance declines under both low and high-stress conditions. Our analysis further revealed that these StressPrompts significantly alter the internal states of LLMs, leading to changes in their neural representations that mirror human responses to stress. This research provides critical insights into the operational robustness and flexibility of LLMs, demonstrating the importance of designing AI systems capable of maintaining high performance in real-world scenarios where stress is prevalent, such as in customer service, healthcare, and emergency response contexts. Moreover, this study contributes to the broader AI research community by offering a new perspective on how LLMs handle different scenarios and their similarities to human cognition.
Authors: Debargha Ganguly, Srinivasan Iyengar, Vipin Chaudhary, Shivkumar Kalyanaraman
Abstract: Large Language Models (LLMs) have revolutionized natural language processing, yet they struggle with inconsistent reasoning, particularly in novel domains and complex logical sequences. This research introduces Proof of Thought, a framework that enhances the reliability and transparency of LLM outputs. Our approach bridges LLM-generated ideas with formal logic verification, employing a custom interpreter to convert LLM outputs into First Order Logic constructs for theorem prover scrutiny. Central to our method is an intermediary JSON-based Domain-Specific Language, which by design balances precise logical structures with intuitive human concepts. This hybrid representation enables both rigorous validation and accessible human comprehension of LLM reasoning processes. Key contributions include a robust type system with sort management for enhanced logical integrity, explicit representation of rules for clear distinction between factual and inferential knowledge, and a flexible architecture that allows for easy extension to various domain-specific applications. We demonstrate Proof of Thought's effectiveness through benchmarking on StrategyQA and a novel multimodal reasoning task, showing improved performance in open-ended scenarios. By providing verifiable and interpretable results, our technique addresses critical needs for AI system accountability and sets a foundation for human-in-the-loop oversight in high-stakes domains.
Authors: Xun Xian, Ganghua Wang, Xuan Bi, Jayanth Srinivasa, Ashish Kundu, Charles Fleming, Mingyi Hong, Jie Ding
Abstract: Retrieval-Augmented Generation (RAG) has been empirically shown to enhance the performance of large language models (LLMs) in knowledge-intensive domains such as healthcare, finance, and legal contexts. Given a query, RAG retrieves relevant documents from a corpus and integrates them into the LLMs' generation process. In this study, we investigate the adversarial robustness of RAG, focusing specifically on examining the retrieval system. First, across 225 different setup combinations of corpus, retriever, query, and targeted information, we show that retrieval systems are vulnerable to universal poisoning attacks in medical Q\&A. In such attacks, adversaries generate poisoned documents containing a broad spectrum of targeted information, such as personally identifiable information. When these poisoned documents are inserted into a corpus, they can be accurately retrieved by any users, as long as attacker-specified queries are used. To understand this vulnerability, we discovered that the deviation from the query's embedding to that of the poisoned document tends to follow a pattern in which the high similarity between the poisoned document and the query is retained, thereby enabling precise retrieval. Based on these findings, we develop a new detection-based defense to ensure the safe use of RAG. Through extensive experiments spanning various Q\&A domains, we observed that our proposed method consistently achieves excellent detection rates in nearly all cases.
Authors: Zehao Wang, Minye Wu, Yixin Cao, Yubo Ma, Meiqi Chen, Tinne Tuytelaars
Abstract: This study presents a novel evaluation framework for the Vision-Language Navigation (VLN) task. It aims to diagnose current models for various instruction categories at a finer-grained level. The framework is structured around the context-free grammar (CFG) of the task. The CFG serves as the basis for the problem decomposition and the core premise of the instruction categories design. We propose a semi-automatic method for CFG construction with the help of Large-Language Models (LLMs). Then, we induct and generate data spanning five principal instruction categories (i.e. direction change, landmark recognition, region recognition, vertical movement, and numerical comprehension). Our analysis of different models reveals notable performance discrepancies and recurrent issues. The stagnation of numerical comprehension, heavy selective biases over directional concepts, and other interesting findings contribute to the development of future language-guided navigation systems.
Authors: Zeyu Huang, Zihan Qiu, Zili Wang, Edoardo M. Ponti, Ivan Titov
Abstract: Reinforcement Learning from Human Feedback aligns the outputs of Large Language Models with human values and preferences. Central to this process is the reward model (RM), which translates human feedback into training signals for optimising LLM behaviour. However, RMs can develop biases by exploiting spurious correlations in their training data, such as favouring outputs based on length or style rather than true quality. These biases can lead to incorrect output rankings, sub-optimal model evaluations, and the amplification of undesirable behaviours in LLMs alignment. This paper addresses the challenge of correcting such biases without additional data and training, introducing the concept of Post-hoc Reward Calibration. We first propose an intuitive approach to estimate the bias term and, thus, remove it to approximate the underlying true reward. We then extend the approach to a more general and robust form with the Locally Weighted Regression. Focusing on the prevalent length bias, we validate our proposed approaches across three experimental settings, demonstrating consistent improvements: (1) a 3.11 average performance gain across 33 reward models on the RewardBench dataset; (2) enhanced alignment of RM rankings with GPT-4 evaluations and human preferences based on the AlpacaEval benchmark; and (3) improved Length-Controlled win rate of the RLHF process in multiple LLM--RM combinations. Our method is computationally efficient and generalisable to other types of bias and RMs, offering a scalable and robust solution for mitigating biases in LLM alignment. Our code and results are available at https://github.com/ZeroYuHuang/Reward-Calibration.
Authors: Ryuichi Yamamoto, Yuma Shirahata, Masaya Kawamura, Kentaro Tachibana
Abstract: We propose a novel description-based controllable text-to-speech (TTS) method with cross-lingual control capability. To address the lack of audio-description paired data in the target language, we combine a TTS model trained on the target language with a description control model trained on another language, which maps input text descriptions to the conditional features of the TTS model. These two models share disentangled timbre and style representations based on self-supervised learning (SSL), allowing for disentangled voice control, such as controlling speaking styles while retaining the original timbre. Furthermore, because the SSL-based timbre and style representations are language-agnostic, combining the TTS and description control models while sharing the same embedding space effectively enables cross-lingual control of voice characteristics. Experiments on English and Japanese TTS demonstrate that our method achieves high naturalness and controllability for both languages, even though no Japanese audio-description pairs are used.
Authors: Yifan Jiang, Kriti Aggarwal, Tanmay Laud, Kashif Munir, Jay Pujara, Subhabrata Mukherjee
Abstract: The rapid progress of Large Language Models (LLMs) has opened up new opportunities across various domains and applications; yet it also presents challenges related to potential misuse. To mitigate such risks, red teaming has been employed as a proactive security measure to probe language models for harmful outputs via jailbreak attacks. However, current jailbreak attack approaches are single-turn with explicit malicious queries that do not fully capture the complexity of real-world interactions. In reality, users can engage in multi-turn interactions with LLM-based chat assistants, allowing them to conceal their true intentions in a more covert manner. To bridge this gap, we, first, propose a new jailbreak approach, RED QUEEN ATTACK. This method constructs a multi-turn scenario, concealing the malicious intent under the guise of preventing harm. We craft 40 scenarios that vary in turns and select 14 harmful categories to generate 56k multi-turn attack data points. We conduct comprehensive experiments on the RED QUEEN ATTACK with four representative LLM families of different sizes. Our experiments reveal that all LLMs are vulnerable to RED QUEEN ATTACK, reaching 87.62% attack success rate on GPT-4o and 75.4% on Llama3-70B. Further analysis reveals that larger models are more susceptible to the RED QUEEN ATTACK, with multi-turn structures and concealment strategies contributing to its success. To prioritize safety, we introduce a straightforward mitigation strategy called RED QUEEN GUARD, which aligns LLMs to effectively counter adversarial attacks. This approach reduces the attack success rate to below 1% while maintaining the model's performance across standard benchmarks. Full implementation and dataset are publicly accessible at https://github.com/kriti-hippo/red_queen.
Authors: Gongfan Fang, Hongxu Yin, Saurav Muralidharan, Greg Heinrich, Jeff Pool, Jan Kautz, Pavlo Molchanov, Xinchao Wang
Abstract: Large Language Models (LLMs) are distinguished by their massive parameter counts, which typically result in significant redundancy. This work introduces MaskLLM, a learnable pruning method that establishes Semi-structured (or ``N:M'') Sparsity in LLMs, aimed at reducing computational overhead during inference. Instead of developing a new importance criterion, MaskLLM explicitly models N:M patterns as a learnable distribution through Gumbel Softmax sampling. This approach facilitates end-to-end training on large-scale datasets and offers two notable advantages: 1) High-quality Masks - our method effectively scales to large datasets and learns accurate masks; 2) Transferability - the probabilistic modeling of mask distribution enables the transfer learning of sparsity across domains or tasks. We assessed MaskLLM using 2:4 sparsity on various LLMs, including LLaMA-2, Nemotron-4, and GPT-3, with sizes ranging from 843M to 15B parameters, and our empirical results show substantial improvements over state-of-the-art methods. For instance, leading approaches achieve a perplexity (PPL) of 10 or greater on Wikitext compared to the dense model's 5.12 PPL, but MaskLLM achieves a significantly lower 6.72 PPL solely by learning the masks with frozen weights. Furthermore, MaskLLM's learnable nature allows customized masks for lossless application of 2:4 sparsity to downstream tasks or domains. Code is available at \url{https://github.com/NVlabs/MaskLLM}.
Authors: Xuefeng Du, Chaowei Xiao, Yixuan Li
Abstract: The surge in applications of large language models (LLMs) has prompted concerns about the generation of misleading or fabricated information, known as hallucinations. Therefore, detecting hallucinations has become critical to maintaining trust in LLM-generated content. A primary challenge in learning a truthfulness classifier is the lack of a large amount of labeled truthful and hallucinated data. To address the challenge, we introduce HaloScope, a novel learning framework that leverages the unlabeled LLM generations in the wild for hallucination detection. Such unlabeled data arises freely upon deploying LLMs in the open world, and consists of both truthful and hallucinated information. To harness the unlabeled data, we present an automated membership estimation score for distinguishing between truthful and untruthful generations within unlabeled mixture data, thereby enabling the training of a binary truthfulness classifier on top. Importantly, our framework does not require extra data collection and human annotations, offering strong flexibility and practicality for real-world applications. Extensive experiments show that HaloScope can achieve superior hallucination detection performance, outperforming the competitive rivals by a significant margin. Code is available at https://github.com/deeplearningwisc/haloscope.
Authors: Gary A. McCully, John D. Hastings, Shengjie Xu, Adam Fortier
Abstract: Ransomware and other forms of malware cause significant financial and operational damage to organizations by exploiting long-standing and often difficult-to-detect software vulnerabilities. To detect vulnerabilities such as buffer overflows in compiled code, this research investigates the application of unidirectional transformer-based embeddings, specifically GPT-2. Using a dataset of LLVM functions, we trained a GPT-2 model to generate embeddings, which were subsequently used to build LSTM neural networks to differentiate between vulnerable and non-vulnerable code. Our study reveals that embeddings from the GPT-2 model significantly outperform those from bidirectional models of BERT and RoBERTa, achieving an accuracy of 92.5% and an F1-score of 89.7%. LSTM neural networks were developed with both frozen and unfrozen embedding model layers. The model with the highest performance was achieved when the embedding layers were unfrozen. Further, the research finds that, in exploring the impact of different optimizers within this domain, the SGD optimizer demonstrates superior performance over Adam. Overall, these findings reveal important insights into the potential of unidirectional transformer-based approaches in enhancing cybersecurity defenses.
Authors: Hanlin Wu, Zhenguang G. Cai
Abstract: Spoken language is often, if not always, understood in a context that includes the identities of speakers. For instance, we can easily make sense of an utterance such as "I'm going to have a manicure this weekend" or "The first time I got pregnant I had a hard time" when the utterance is spoken by a woman, but it would be harder to understand when it is spoken by a man. Previous event-related potential (ERP) studies have shown mixed results regarding the neurophysiological responses to such speaker-mismatched utterances, with some reporting an N400 effect and others a P600 effect. In an experiment involving 64 participants, we showed that these different ERP effects reflect distinct cognitive processes employed to resolve the speaker-message mismatch. When possible, the message is integrated with the speaker context to arrive at an interpretation, as in the case of violations of social stereotypes (e.g., men getting a manicure), resulting in an N400 effect. However, when such integration is impossible due to violations of biological knowledge (e.g., men getting pregnant), listeners engage in an error correction process to revise either the perceived utterance or the speaker context, resulting in a P600 effect. Additionally, we found that the social N400 effect decreased as a function of the listener's personality trait of openness, while the biological P600 effect remained robust. Our findings help to reconcile the empirical inconsistencies in the literature and provide a rational account of speaker-contextualized language comprehension.
Authors: Saber Malekmohammadi, Golnoosh Farnadi
Abstract: A significant approach in natural language processing involves large-scale pre-training on general domain data followed by adaptation to specific tasks or domains. As models grow in size, full fine-tuning all parameters becomes increasingly impractical. To address this, some methods for low-rank task adaptation of language models have been proposed, e.g. LoRA and FLoRA. These methods keep the pre-trained model weights fixed and incorporate trainable low-rank decomposition matrices into some layers of the transformer architecture, called adapters. This approach significantly reduces the number of trainable parameters required for downstream tasks compared to full fine-tuning all parameters. In this work, we look at low-rank adaptation from the lens of data privacy. We show theoretically that the low-rank adaptation used in LoRA and FLoRA is equivalent to injecting some random noise into the batch gradients w.r.t the adapter parameters coming from their full fine-tuning, and we quantify the variance of the injected noise. By establishing a Berry-Esseen type bound on the total variation distance between the noise distribution and a Gaussian distribution with the same variance, we show that the dynamics of LoRA and FLoRA are very close to differentially private full fine-tuning the adapters, which suggests that low-rank adaptation implicitly provides privacy w.r.t the fine-tuning data. Finally, using Johnson-Lindenstrauss lemma, we show that when augmented with gradient clipping, low-rank adaptation is almost equivalent to differentially private full fine-tuning adapters with a fixed noise scale.
Authors: Himanshu Pandey, Akhil Amod, Shivang, Kshitij Jaggi, Ruchi Garg, Abheet Jain, Vinayak Tantia
Abstract: Artificial Intelligence (AI) and Large Language Models (LLMs) hold significant promise in revolutionizing healthcare, especially in clinical applications. Simultaneously, Digital Twin technology, which models and simulates complex systems, has gained traction in enhancing patient care. However, despite the advances in experimental clinical settings, the potential of AI and digital twins to streamline clinical operations remains largely untapped. This paper introduces a novel digital twin framework specifically designed to enhance oncology clinical operations. We propose the integration of multiple specialized digital twins, such as the Medical Necessity Twin, Care Navigator Twin, and Clinical History Twin, to enhance workflow efficiency and personalize care for each patient based on their unique data. Furthermore, by synthesizing multiple data sources and aligning them with the National Comprehensive Cancer Network (NCCN) guidelines, we create a dynamic Cancer Care Path, a continuously evolving knowledge base that enables these digital twins to provide precise, tailored clinical recommendations.
Authors: Nilanjan Sinhababu, Andrew Parry, Debasis Ganguly, Debasis Samanta, Pabitra Mitra
Abstract: A supervised ranking model, despite its advantage of being effective, usually involves complex processing - typically multiple stages of task-specific pre-training and fine-tuning. This has motivated researchers to explore simpler pipelines leveraging large language models (LLMs) that are capable of working in a zero-shot manner. However, since zero-shot inference does not make use of a training set of pairs of queries and their relevant documents, its performance is mostly worse than that of supervised models, which are trained on such example pairs. Motivated by the existing findings that training examples generally improve zero-shot performance, in our work, we explore if this also applies to ranking models. More specifically, given a query and a pair of documents, the preference prediction task is improved by augmenting examples of preferences for similar queries from a training set. Our proposed pairwise few-shot ranker demonstrates consistent improvements over the zero-shot baseline on both in-domain (TREC DL) and out-domain (BEIR subset) retrieval benchmarks. Our method also achieves a close performance to that of a supervised model without requiring any complex training pipeline.
Authors: Keyu An, Shiliang Zhang, Zhijie Yan
Abstract: In this study, we delve into the efficacy of transformers within pre-trained language models (PLMs) when repurposed as encoders for Automatic Speech Recognition (ASR). Our underlying hypothesis posits that, despite being initially trained on text-based corpora, these transformers possess a remarkable capacity to extract effective features from the input sequence. This inherent capability, we argue, is transferrable to speech data, thereby augmenting the acoustic modeling ability of ASR. Through rigorous empirical analysis, our findings reveal a notable improvement in Character Error Rate (CER) and Word Error Rate (WER) across diverse ASR tasks when transformers from pre-trained LMs are incorporated. Particularly, they serve as an advantageous starting point for initializing ASR encoders. Furthermore, we uncover that these transformers, when integrated into a well-established ASR encoder, can significantly boost performance, especially in scenarios where profound semantic comprehension is pivotal. This underscores the potential of leveraging the semantic prowess embedded within pre-trained transformers to advance ASR systems' capabilities.
Authors: Rimvydas Rubavicius, Peter David Fagan, Alex Lascarides, Subramanian Ramamoorthy
Abstract: This paper addresses a challenging interactive task learning scenario we call rearrangement under unawareness: to manipulate a rigid-body environment in a context where the robot is unaware of a concept that's key to solving the instructed task. We propose SECURE, an interactive task learning framework designed to solve such problems by fixing a deficient domain model using embodied conversation. Through dialogue, the robot discovers and then learns to exploit unforeseen possibilities. Using SECURE, the robot not only learns from the user's corrective feedback when it makes a mistake, but it also learns to make strategic dialogue decisions for revealing useful evidence about novel concepts for solving the instructed task. Together, these abilities allow the robot to generalise to subsequent tasks using newly acquired knowledge. We demonstrate that a robot that is semantics-aware -- that is, it exploits the logical consequences of both sentence and discourse semantics in the learning and inference process -- learns to solve rearrangement under unawareness more effectively than a robot that lacks such capabilities.
Authors: Taridzo Chomutare, Aleksandar Babic, Laura-Maria Peltonen, Silja Elunurm, Peter Lundberg, Arne J\"onsson, Emma Eneling, Ciprian-Virgil Gerstenberger, Troels Siggaard, Raivo Kolde, Oskar Jerdhaf, Martin Hansson, Alexandra Makhlysheva, Miroslav Muzny, Erik Ylip\"a\"a, S{\o}ren Brunak, Hercules Dalianis
Abstract: Background: Centralized collection and processing of healthcare data across national borders pose significant challenges, including privacy concerns, data heterogeneity and legal barriers. To address some of these challenges, we formed an interdisciplinary consortium to develop a feder-ated health data network, comprised of six institutions across five countries, to facilitate Nordic-Baltic cooperation on secondary use of health data. The objective of this report is to offer early insights into our experiences developing this network. Methods: We used a mixed-method ap-proach, combining both experimental design and implementation science to evaluate the factors affecting the implementation of our network. Results: Technically, our experiments indicate that the network functions without significant performance degradation compared to centralized simu-lation. Conclusion: While use of interdisciplinary approaches holds a potential to solve challeng-es associated with establishing such collaborative networks, our findings turn the spotlight on the uncertain regulatory landscape playing catch up and the significant operational costs.
Authors: Yujia Sun, Zeyu Zhao, Korin Richmond, Yuanchao Li
Abstract: Emotion recognition from speech and music shares similarities due to their acoustic overlap, which has led to interest in transferring knowledge between these domains. However, the shared acoustic cues between speech and music, particularly those encoded by Self-Supervised Learning (SSL) models, remain largely unexplored, given the fact that SSL models for speech and music have rarely been applied in cross-domain research. In this work, we revisit the acoustic similarity between emotion speech and music, starting with an analysis of the layerwise behavior of SSL models for Speech Emotion Recognition (SER) and Music Emotion Recognition (MER). Furthermore, we perform cross-domain adaptation by comparing several approaches in a two-stage fine-tuning process, examining effective ways to utilize music for SER and speech for MER. Lastly, we explore the acoustic similarities between emotional speech and music using Frechet audio distance for individual emotions, uncovering the issue of emotion bias in both speech and music SSL models. Our findings reveal that while speech and music SSL models do capture shared acoustic features, their behaviors can vary depending on different emotions due to their training strategies and domain-specificities. Additionally, parameter-efficient fine-tuning can enhance SER and MER performance by leveraging knowledge from each other. This study provides new insights into the acoustic similarity between emotional speech and music, and highlights the potential for cross-domain generalization to improve SER and MER systems.
Authors: Bingyao Liu, Iris Li, Jianhua Yao, Yuan Chen, Guanming Huang, Jiajing Wang
Abstract: This paper takes the graph neural network as the technical framework, integrates the intrinsic connections between enterprise financial indicators, and proposes a model for enterprise credit risk assessment. The main research work includes: Firstly, based on the experience of predecessors, we selected 29 enterprise financial data indicators, abstracted each indicator as a vertex, deeply analyzed the relationships between the indicators, constructed a similarity matrix of indicators, and used the maximum spanning tree algorithm to achieve the graph structure mapping of enterprises; secondly, in the representation learning phase of the mapped graph, a graph neural network model was built to obtain its embedded representation. The feature vector of each node was expanded to 32 dimensions, and three GraphSAGE operations were performed on the graph, with the results pooled using the Pool operation, and the final output of three feature vectors was averaged to obtain the graph's embedded representation; finally, a classifier was constructed using a two-layer fully connected network to complete the prediction task. Experimental results on real enterprise data show that the model proposed in this paper can well complete the multi-level credit level estimation of enterprises. Furthermore, the tree-structured graph mapping deeply portrays the intrinsic connections of various indicator data of the company, and according to the ROC and other evaluation criteria, the model's classification effect is significant and has good "robustness".
Authors: Shuai Zhao, Leilei Gan, Zhongliang Guo, Xiaobao Wu, Luwei Xiao, Xiaoyu Xu, Cong-Duy Nguyen, Luu Anh Tuan
Abstract: Despite being widely applied due to their exceptional capabilities, Large Language Models (LLMs) have been proven to be vulnerable to backdoor attacks. These attacks introduce targeted vulnerabilities into LLMs by poisoning training samples and full-parameter fine-tuning. However, this kind of backdoor attack is limited since they require significant computational resources, especially as the size of LLMs increases. Besides, parameter-efficient fine-tuning (PEFT) offers an alternative but the restricted parameter updating may impede the alignment of triggers with target labels. In this study, we first verify that backdoor attacks with PEFT may encounter challenges in achieving feasible performance. To address these issues and improve the effectiveness of backdoor attacks with PEFT, we propose a novel backdoor attack algorithm from weak to strong based on contrastive knowledge distillation (W2SAttack). Specifically, we poison small-scale language models through full-parameter fine-tuning to serve as the teacher model. The teacher model then covertly transfers the backdoor to the large-scale student model through contrastive knowledge distillation, which employs PEFT. Theoretical analysis reveals that W2SAttack has the potential to augment the effectiveness of backdoor attacks. We demonstrate the superior performance of W2SAttack on classification tasks across four language models, four backdoor attack algorithms, and two different architectures of teacher models. Experimental results indicate success rates close to 100% for backdoor attacks targeting PEFT.
Authors: Georg Ahnert, Max Pellert, David Garcia, Markus Strohmaier
Abstract: This paper proposes temporally aligned Large Language Models (LLMs) as a tool for longitudinal analysis of social media data. We fine-tune Temporal Adapters for Llama 3 8B on full timelines from a panel of British Twitter users, and extract longitudinal aggregates of emotions and attitudes with established questionnaires. We validate our estimates against representative British survey data and find strong positive, significant correlations for several collective emotions. The obtained estimates are robust across multiple training seeds and prompt formulations, and in line with collective emotions extracted using a traditional classification model trained on labeled data. To the best of our knowledge, this is the first work to extend the analysis of affect in LLMs to a longitudinal setting through Temporal Adapters. Our work enables new approaches towards the longitudinal analysis of social media data.
Authors: Jakub {\L}ucki, Boyi Wei, Yangsibo Huang, Peter Henderson, Florian Tram\`er, Javier Rando
Abstract: Large language models are finetuned to refuse questions about hazardous knowledge, but these protections can often be bypassed. Unlearning methods aim at completely removing hazardous capabilities from models and make them inaccessible to adversaries. This work challenges the fundamental differences between unlearning and traditional safety post-training from an adversarial perspective. We demonstrate that existing jailbreak methods, previously reported as ineffective against unlearning, can be successful when applied carefully. Furthermore, we develop a variety of adaptive methods that recover most supposedly unlearned capabilities. For instance, we show that finetuning on 10 unrelated examples or removing specific directions in the activation space can recover most hazardous capabilities for models edited with RMU, a state-of-the-art unlearning method. Our findings challenge the robustness of current unlearning approaches and question their advantages over safety training.
Authors: Yotam Wolf, Binyamin Rothberg, Dorin Shteyman, Amnon Shashua
Abstract: A common practice in large language model (LLM) usage for complex analytical tasks such as code generation, is to sample a solution for the entire task within the model's context window. Previous works have shown that subtask decomposition within the model's context (chain of thought), is beneficial for solving such tasks. In this work, we point a limitation of LLMs' ability to perform several sub-tasks within the same context window - an in-context hardness of composition, pointing to an advantage for distributing a decomposed problem in a multi-agent system of LLMs. The hardness of composition is quantified by a generation complexity metric, i.e., the number of LLM generations required to sample at least one correct solution. We find a gap between the generation complexity of solving a compositional problem within the same context relative to distributing it among multiple agents, that increases exponentially with the solution's length. We prove our results theoretically and demonstrate them empirically.
Authors: Kai Chen, Yunhao Gou, Runhui Huang, Zhili Liu, Daxin Tan, Jing Xu, Chunwei Wang, Yi Zhu, Yihan Zeng, Kuo Yang, Dingdong Wang, Kun Xiang, Haoyuan Li, Haoli Bai, Jianhua Han, Xiaohui Li, Weike Jin, Nian Xie, Yu Zhang, James T. Kwok, Hengshuang Zhao, Xiaodan Liang, Dit-Yan Yeung, Xiao Chen, Zhenguo Li, Wei Zhang, Qun Liu, Lanqing Hong, Lu Hou, Hang Xu
Abstract: GPT-4o, an omni-modal model that enables vocal conversations with diverse emotions and tones, marks a milestone for omni-modal foundation models. However, empowering Large Language Models to perceive and generate images, texts, and speeches end-to-end with publicly available data remains challenging in the open-source community. Existing vision-language models rely on external tools for the speech processing, while speech-language models still suffer from limited or even without vision-understanding abilities. To address this gap, we propose EMOVA (EMotionally Omni-present Voice Assistant), to enable Large Language Models with end-to-end speech capabilities while maintaining the leading vision-language performance. With a semantic-acoustic disentangled speech tokenizer, we notice surprisingly that omni-modal alignment can further enhance vision-language and speech abilities compared with the corresponding bi-modal aligned counterparts. Moreover, a lightweight style module is proposed for flexible speech style controls (e.g., emotions and pitches). For the first time, EMOVA achieves state-of-the-art performance on both the vision-language and speech benchmarks, and meanwhile, supporting omni-modal spoken dialogue with vivid emotions.
Authors: Soeun Lee, Si-Woo Kim, Taewhan Kim, Dong-Jin Kim
Abstract: Recent advancements in image captioning have explored text-only training methods to overcome the limitations of paired image-text data. However, existing text-only training methods often overlook the modality gap between using text data during training and employing images during inference. To address this issue, we propose a novel approach called Image-like Retrieval, which aligns text features with visually relevant features to mitigate the modality gap. Our method further enhances the accuracy of generated captions by designing a Fusion Module that integrates retrieved captions with input features. Additionally, we introduce a Frequency-based Entity Filtering technique that significantly improves caption quality. We integrate these methods into a unified framework, which we refer to as IFCap ($\textbf{I}$mage-like Retrieval and $\textbf{F}$requency-based Entity Filtering for Zero-shot $\textbf{Cap}$tioning). Through extensive experimentation, our straightforward yet powerful approach has demonstrated its efficacy, outperforming the state-of-the-art methods by a significant margin in both image captioning and video captioning compared to zero-shot captioning based on text-only training.
Authors: Yanming Wan, Yue Wu, Yiping Wang, Jiayuan Mao, Natasha Jaques
Abstract: For AI agents to be helpful to humans, they should be able to follow natural language instructions to complete everyday cooperative tasks in human environments. However, real human instructions inherently possess ambiguity, because the human speakers assume sufficient prior knowledge about their hidden goals and intentions. Standard language grounding and planning methods fail to address such ambiguities because they do not model human internal goals as additional partially observable factors in the environment. We propose a new framework, Follow Instructions with Social and Embodied Reasoning (FISER), aiming for better natural language instruction following in collaborative embodied tasks. Our framework makes explicit inferences about human goals and intentions as intermediate reasoning steps. We implement a set of Transformer-based models and evaluate them over a challenging benchmark, HandMeThat. We empirically demonstrate that using social reasoning to explicitly infer human intentions before making action plans surpasses purely end-to-end approaches. We also compare our implementation with strong baselines, including Chain of Thought prompting on the largest available pre-trained language models, and find that FISER provides better performance on the embodied social reasoning tasks under investigation, reaching the state-of-the-art on HandMeThat.
Authors: Mohammad Khosravani, Amine Trabelsi
Abstract: Unsupervised summarization is a powerful technique that enables training summarizing models without requiring labeled datasets. This survey covers different recent techniques and models used for unsupervised summarization. We cover extractive, abstractive, and hybrid models and strategies used to achieve unsupervised summarization. While the main focus of this survey is on recent research, we also cover some of the important previous research. We additionally introduce a taxonomy, classifying different research based on their approach to unsupervised training. Finally, we discuss the current approaches and mention some datasets and evaluation methods.
Authors: Haofei Yu, Zhengyang Qi, Lawrence Jang, Ruslan Salakhutdinov, Louis-Philippe Morency, Paul Pu Liang
Abstract: Advances in multimodal models have greatly improved how interactions relevant to various tasks are modeled. Today's multimodal models mainly focus on the correspondence between images and text, using this for tasks like image-text matching. However, this covers only a subset of real-world interactions. Novel interactions, such as sarcasm expressed through opposing spoken words and gestures or humor expressed through utterances and tone of voice, remain challenging. In this paper, we introduce an approach to enhance multimodal models, which we call Multimodal Mixtures of Experts (MMoE). The key idea in MMoE is to train separate expert models for each type of multimodal interaction, such as redundancy present in both modalities, uniqueness in one modality, or synergy that emerges when both modalities are fused. On a sarcasm detection task (MUStARD) and a humor detection task (URFUNNY), we obtain new state-of-the-art results. MMoE is also able to be applied to various types of models to gain improvement.
Authors: Guiming Hardy Chen, Shunian Chen, Ziche Liu, Feng Jiang, Benyou Wang
Abstract: Adopting human and large language models (LLM) as judges (a.k.a human- and LLM-as-a-judge) for evaluating the performance of LLMs has recently gained attention. Nonetheless, this approach concurrently introduces potential biases from human and LLMs, questioning the reliability of the evaluation results. In this paper, we propose a novel framework that is free from referencing groundtruth annotations for investigating Misinformation Oversight Bias, Gender Bias, Authority Bias and Beauty Bias on LLM and human judges. We curate a dataset referring to the revised Bloom's Taxonomy and conduct thousands of evaluations. Results show that human and LLM judges are vulnerable to perturbations to various degrees, and that even the cutting-edge judges possess considerable biases. We further exploit these biases to conduct attacks on LLM judges. We hope that our work can notify the community of the bias and vulnerability of human- and LLM-as-a-judge, as well as the urgency of developing robust evaluation systems.
Authors: Atsuki Yamaguchi, Aline Villavicencio, Nikolaos Aletras
Abstract: The development of state-of-the-art generative large language models (LLMs) disproportionately relies on English-centric tokenizers, vocabulary and pre-training data. Despite the fact that some LLMs have multilingual capabilities, recent studies have shown that their inference efficiency deteriorates when generating text in languages other than English. This results in increased inference time and costs. Cross-lingual vocabulary adaptation (CVA) methods have been proposed for adapting models to a target language aiming to improve downstream performance. However, the effectiveness of these methods on increasing inference efficiency of generative LLMs has yet to be explored. In this paper, we perform an empirical study of five CVA methods on four generative LLMs (including monolingual and multilingual models) across four typologically-diverse languages and four natural language understanding tasks. We find that CVA substantially contributes to LLM inference speedups of up to 271.5\%. We also show that adapting LLMs that have been pre-trained on more balanced multilingual data results in downstream performance comparable to the original models.
Authors: Heydar Soudani, Evangelos Kanoulas, Faegheh Hasibi
Abstract: Language Models (LMs) memorize a vast amount of factual knowledge, exhibiting strong performance across diverse tasks and domains. However, it has been observed that the performance diminishes when dealing with less-popular or low-frequency concepts and entities, for example in domain specific applications. The two prominent approaches to enhance the performance of LMs on low-frequent topics are: Retrieval Augmented Generation (RAG) and fine-tuning (FT) over synthetic data. This paper explores and evaluates the impact of RAG and FT on customizing LMs in handling low-frequency entities on question answering tasks. We conduct extensive experiments on twelve LMs of varying size and type and different fine tuning, data augmentation, and retrieval models. Our findings indicate that while FT boosts the performance across entities of varying popularity, RAG surpasses FT by a large margin particularly for least popular factual knowledge. Additionally, the success of both RAG and FT approaches is amplified by improving retrieval and data augmentation techniques. Fine tuning, while beneficial for small LMs, requires extensive resources. To address this issue, we propose the new Stimulus RAG approach that surpasses the effectiveness of fine tuning based approaches, thereby eliminating the need for the costly data augmentation and fine tuning step for enriching LMs with less popular factual knowledge.
Authors: Xin Lu, Yanyan Zhao, Bing Qin, Liangyu Huo, Qing Yang, Dongliang Xu
Abstract: Pre-trained language models have been proven to possess strong base capabilities, which not only excel in in-distribution language modeling but also show powerful abilities in out-of-distribution language modeling, transfer learning and few-shot learning. Unlike existing work focusing on the influence of scale on base capabilities, our work examines the influence of architecture on those. Specifically, our concern is: How does architecture influence the base capabilities of pre-trained language models? In this work, we attempt to explain and reverse the decline in base capabilities caused by the architecture of FFN-Wider Transformers, seeking to provide some insights. Through analysis, we found the contribution ratio of Multi-Head Attention (a combination function) to pre-trained language modeling is a key factor affecting base capabilities. FFN-Wider Transformers reduce the contribution ratio of this combination function, leading to a decline in base capabilities. We confirmed this by experiments and proposed Combination Enhanced Architecture (CEA) to address the decline in base capabilities of such models. Significantly, we extended our explanation and CEA to Mixture of Experts (MoE) Transformers. We successfully achieved significant improvements in base capabilities on a 14B parameter MoE model, demonstrating the practical application value of our work. This also indicates that our analysis has a certain guiding significance for architecture analysis, architecture improvement and architecture design.
Authors: Eli Schwartz, Leshem Choshen, Joseph Shtok, Sivan Doveh, Leonid Karlinsky, Assaf Arbelle
Abstract: Language models struggle with handling numerical data and performing arithmetic operations. We hypothesize that this limitation can be partially attributed to non-intuitive textual numbers representation. When a digit is read or generated by a causal language model it does not know its place value (e.g. thousands vs. hundreds) until the entire number is processed. To address this issue, we propose a simple adjustment to how numbers are represented by including the count of digits before each number. For instance, instead of "42", we suggest using "{2:42}" as the new format. This approach, which we term NumeroLogic, offers an added advantage in number generation by serving as a Chain of Thought (CoT). By requiring the model to consider the number of digits first, it enhances the reasoning process before generating the actual number. We use arithmetic tasks to demonstrate the effectiveness of the NumeroLogic formatting. We further demonstrate NumeroLogic applicability to general natural language modeling, improving language understanding performance in the MMLU benchmark.
Authors: Kaixin Li, Yuchen Tian, Qisheng Hu, Ziyang Luo, Zhiyong Huang, Jing Ma
Abstract: Programming often involves converting detailed and complex specifications into code, a process during which developers typically utilize visual aids to more effectively convey concepts. While recent developments in Large Multimodal Models have demonstrated remarkable abilities in visual reasoning and mathematical tasks, there is little work on investigating whether these models can effectively interpret visual elements for code generation. To this end, we present MMCode, the first multi-modal coding dataset for evaluating algorithmic problem-solving skills in visually rich contexts. MMCode contains 3,548 questions and 6,620 images collected from real-world programming challenges harvested from 10 code competition websites, presenting significant challenges due to the extreme demand for reasoning abilities. Our experiment results show that current state-of-the-art models struggle to solve these problems. The results highlight the lack of powerful vision-code models, and we hope MMCode can serve as an inspiration for future works in this domain. The data and code are publicly available at https://github.com/likaixin2000/MMCode.
Authors: Jumbly Grindrod
Abstract: The transformer architecture, introduced by Vaswani et al. (2017), is at the heart of the remarkable recent progress in the development of language models, including widely-used chatbots such as Chat-GPT and Claude. In this paper, I argue that we can extract from the way the transformer architecture works a theory of the relationship between context and meaning. I call this the transformer theory, and I argue that it is novel with regard to two related philosophical debates: the contextualism debate regarding the extent of context-sensitivity across natural language, and the polysemy debate regarding how polysemy should be captured within an account of word meaning.
Authors: Wenhao Huang, Zhouhong Gu, Chenghao Peng, Zhixu Li, Jiaqing Liang, Yanghua Xiao, Liqian Wen, Zulong Chen
Abstract: Web scraping is a powerful technique that extracts data from websites, enabling automated data collection, enhancing data analysis capabilities, and minimizing manual data entry efforts. Existing methods, wrappers-based methods suffer from limited adaptability and scalability when faced with a new website, while language agents, empowered by large language models (LLMs), exhibit poor reusability in diverse web environments. In this work, we introduce the paradigm of generating web scrapers with LLMs and propose AutoScraper, a two-stage framework that can handle diverse and changing web environments more efficiently. AutoScraper leverages the hierarchical structure of HTML and similarity across different web pages for generating web scrapers. Besides, we propose a new executability metric for better measuring the performance of web scraper generation tasks. We conduct comprehensive experiments with multiple LLMs and demonstrate the effectiveness of our framework. Resources of this paper can be found at \url{https://github.com/EZ-hwh/AutoScraper}
Authors: Andreas Waldis, Yotam Perlitz, Leshem Choshen, Yufang Hou, Iryna Gurevych
Abstract: We introduce Holmes, a new benchmark designed to assess language models (LMs) linguistic competence - their unconscious understanding of linguistic phenomena. Specifically, we use classifier-based probing to examine LMs' internal representations regarding distinct linguistic phenomena (e.g., part-of-speech tagging). As a result, we meet recent calls to disentangle LMs' linguistic competence from other cognitive abilities, such as following instructions in prompting-based evaluations. Composing Holmes, we review over 270 probing studies and include more than 200 datasets to assess syntax, morphology, semantics, reasoning, and discourse phenomena. Analyzing over 50 LMs reveals that, aligned with known trends, their linguistic competence correlates with model size. However, surprisingly, model architecture and instruction tuning also significantly influence performance, particularly in morphology and syntax. Finally, we propose FlashHolmes, a streamlined version that reduces the computation load while maintaining high-ranking precision.
Authors: Raul Salles de Padua, Imran Qureshi
Abstract: Two fundamental problems in health-care stem from patient handoff and triage. Doctors are often required to perform complex findings summarization to facilitate efficient communication with specialists and decision making on the urgency of each case. To address these challenges, we present a state of the art radiology report summarization model utilizing adjusted bidirectional encoder representation from transformers BERTtoBERT encoder and decoder architecture. We also provide a data processing pipeline for future models developed on the the MIMIC CXR dataset. Our approach includes a novel method for augmenting medical data and a comprehensive performance analysis. Our best performing model achieved a recall oriented understudy for gisting evaluation L F1 score of 58.75/100, outperforming specialized checkpoints with more sophisticated attention mechanisms. We also provide a data processing pipeline for future models developed on the MIMIC chest X-ray dataset. The model introduced in this paper demonstrates significantly improved capacity in radiology report summarization, highlighting the potential for ensuring better clinical workflows and enhanced patient care.
Authors: Chuanyang Zheng, Yihang Gao, Han Shi, Minbin Huang, Jingyao Li, Jing Xiong, Xiaozhe Ren, Michael Ng, Xin Jiang, Zhenguo Li, Yu Li
Abstract: Positional encoding plays a crucial role in transformers, significantly impacting model performance and length generalization. Prior research has introduced absolute positional encoding (APE) and relative positional encoding (RPE) to distinguish token positions in given sequences. However, both APE and RPE remain fixed after model training regardless of input data, limiting their adaptability and flexibility. Hence, we expect that the desired positional encoding should be data-adaptive and can be dynamically adjusted with the given attention. In this paper, we propose a Data-Adaptive Positional Encoding (DAPE) method, which dynamically and semantically adjusts based on input context and learned fixed priors. Experimental validation on real-world datasets (Arxiv, Books3, and CHE) demonstrates that DAPE enhances model performances in terms of trained length and length generalization, where the improvements are statistically significant. The model visualization suggests that our model can keep both local and anti-local information. Finally, we successfully train the model on sequence length 128 and achieve better performance at evaluation sequence length 8192, compared with other static positional encoding methods, revealing the benefit of the adaptive positional encoding method.
Authors: Gal Yona, Roee Aharoni, Mor Geva
Abstract: We posit that large language models (LLMs) should be capable of expressing their intrinsic uncertainty in natural language. For example, if the LLM is equally likely to output two contradicting answers to the same question, then its generated response should reflect this uncertainty by hedging its answer (e.g., "I'm not sure, but I think..."). We formalize faithful response uncertainty based on the gap between the model's intrinsic confidence in the assertions it makes and the decisiveness by which they are conveyed. This example-level metric reliably indicates whether the model reflects its uncertainty, as it penalizes both excessive and insufficient hedging. We evaluate a variety of aligned LLMs at faithfully communicating uncertainty on several knowledge-intensive question answering tasks. Our results provide strong evidence that modern LLMs are poor at faithfully conveying their uncertainty, and that better alignment is necessary to improve their trustworthiness.
Authors: Krystian Zawistowski
Abstract: LLM text decoding is key component for perceived LLM quality. We demonstrate two experiments showing that decoding methods could be improved by manipulation of token probabilities. First, we test few LLM on SummEval summary scoring dataset, to measure reading comprehension. We compare scores from greedy decoding to expected values over the next token distribution. We scale logits by large temperature to increase the entropy of scores. This allows strong improvement of performance on SummEval (in terms of correlations to human judgement). We see improvement from 6-8% to 13-28% for 7B Mistral and from 20%-46% to 37%-56% for Mixtral, beating GPT 4 0314 result on two metrics. Part of the gain seems related to positional bias. Secondly, we use probability-based tree sampling algorithm, to examine all most probable generations for given prompt.
Authors: Ruixin Hong, Hongming Zhang, Xiaoman Pan, Dong Yu, Changshui Zhang
Abstract: Abstract reasoning, the ability to reason from the abstract essence of a problem, serves as a key to generalization in human reasoning. However, eliciting language models to perform reasoning with abstraction remains unexplored. This paper seeks to bridge this gap by introducing a novel structured reasoning format called Abstraction-of-Thought (AoT). The uniqueness of AoT lies in its explicit requirement for varying levels of abstraction within the reasoning process. This approach could elicit language models to first contemplate on the abstract level before incorporating concrete details, which is overlooked by the prevailing step-by-step Chain-of-Thought (CoT) method. To align models with the AoT format, we present AoT Collection, a generic finetuning dataset consisting of 348k high-quality samples with AoT reasoning processes, collected via an automated and scalable pipeline. We finetune a wide range of language models with AoT Collection and conduct extensive evaluations on 23 unseen tasks from the challenging benchmark Big-Bench Hard. Experimental results indicate that models aligned to AoT reasoning format substantially outperform those aligned to CoT in many reasoning tasks.
Authors: Pinzhen Chen, Simon Yu, Zhicheng Guo, Barry Haddow
Abstract: Multilingual large language models are designed, claimed, and expected to cater to speakers of varied languages. We hypothesise that the current practices of fine-tuning and evaluating these models may not perfectly align with this objective owing to a heavy reliance on translation, which cannot cover language-specific knowledge but can introduce translation defects. It remains unknown whether the nature of the instruction data has an impact on the model output; conversely, it is questionable whether translated test sets can capture such nuances. Due to the often coupled practices of using translated data in both stages, such imperfections could have been overlooked. This work investigates these issues using controlled native or translated data during the instruction tuning and evaluation stages. We show that native or generation benchmarks reveal a notable difference between native and translated instruction data especially when model performance is high, whereas other types of test sets cannot. The comparison between round-trip and single-pass translations reflects the importance of knowledge from language-native resources. Finally, we demonstrate that regularization is beneficial to bridging this gap on structured but not generative tasks.
Authors: Bo Wang, Heyan Huang, Yixin Cao, Jiahao Ying, Wei Tang, Chong Feng
Abstract: While large language models (LLMs) have made notable advancements in natural language processing, they continue to struggle with processing extensive text. Memory mechanism offers a flexible solution for managing long contexts, utilizing techniques such as compression, summarization, and structuring to facilitate nuanced and efficient handling of large volumes of text. However, existing techniques face challenges with static knowledge integration, leading to insufficient adaptation to task-specific needs and missing multi-segmentation relationships, which hinders the dynamic reorganization and logical combination of relevant segments during the response process. To address these issues, we introduce a novel strategy, Question then Reflection Memory Mechanism (QRMeM), incorporating a dual-structured memory pool. This pool synergizes static textual content with structured graph guidance, fostering a reflective trial-and-error approach for navigating and identifying relevant segments. Our evaluation across multiple-choice questions (MCQ) and multi-document question answering (Multi-doc QA) benchmarks showcases QRMeM enhanced performance compared to existing approaches.
Authors: Zhenlong Dai, Chang Yao, WenKang Han, Ying Yuan, Zhipeng Gao, Jingyuan Chen
Abstract: Large Language Models (LLMs) have demonstrated great potential for assisting developers in their daily development. However, most research focuses on generating correct code, how to use LLMs to generate personalized code has seldom been investigated. To bridge this gap, we proposed MPCoder (Multi-user Personalized Code Generator) to generate personalized code for multiple users. To better learn coding style features, we utilize explicit coding style residual learning to capture the syntax code style standards and implicit style learning to capture the semantic code style conventions. We train a multi-user style adapter to better differentiate the implicit feature representations of different users through contrastive learning, ultimately enabling personalized code generation for multiple users. We further propose a novel evaluation metric for estimating similarities between codes of different coding styles. The experimental results show the effectiveness of our approach for this novel task.
Authors: Tanush Chopra, Michael Li, Jacob Haimes
Abstract: When large language models (LLMs) are asked to perform certain tasks, how can we be sure that their learned representations align with reality? We propose a domain-agnostic framework for systematically evaluating distribution shifts in LLMs decision-making processes, where they are given control of mechanisms governed by pre-defined rules. While individual LLM actions may appear consistent with expected behavior, across a large number of trials, statistically significant distribution shifts can emerge. To test this, we construct a well-defined environment with known outcome logic: blackjack. In more than 1,000 trials, we uncover statistically significant evidence suggesting behavioral misalignment in the learned representations of LLM.
Authors: Pedro Ferreira, Ivan Titov, Wilker Aziz
Abstract: Explanation regularisation (ER) has been introduced as a way to guide text classifiers to form their predictions relying on input tokens that humans consider plausible. This is achieved by introducing an auxiliary explanation loss that measures how well the output of an input attribution technique for the model agrees with human-annotated rationales. The guidance appears to benefit performance in out-of-domain (OOD) settings, presumably due to an increased reliance on "plausible" tokens. However, previous work has under-explored the impact of guidance on that reliance, particularly when reliance is measured using attribution techniques different from those used to guide the model. In this work, we seek to close this gap, and also explore the relationship between reliance on plausible features and OOD performance. We find that the connection between ER and the ability of a classifier to rely on plausible features has been overstated and that a stronger reliance on plausible tokens does not seem to be the cause for OOD improvements.
Authors: Doan Nam Long Vu, Timour Igamberdiev, Ivan Habernal
Abstract: Applying differential privacy (DP) by means of the DP-SGD algorithm to protect individual data points during training is becoming increasingly popular in NLP. However, the choice of granularity at which DP is applied is often neglected. For example, neural machine translation (NMT) typically operates on the sentence-level granularity. From the perspective of DP, this setup assumes that each sentence belongs to a single person and any two sentences in the training dataset are independent. This assumption is however violated in many real-world NMT datasets, e.g., those including dialogues. For proper application of DP we thus must shift from sentences to entire documents. In this paper, we investigate NMT at both the sentence and document levels, analyzing the privacy/utility trade-off for both scenarios, and evaluating the risks of not using the appropriate privacy granularity in terms of leaking personally identifiable information (PII). Our findings indicate that the document-level NMT system is more resistant to membership inference attacks, emphasizing the significance of using the appropriate granularity when working with DP.
Authors: Ziyuan Zhuang, Zhiyang Zhang, Sitao Cheng, Fangkai Yang, Jia Liu, Shujian Huang, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang, Qi Zhang
Abstract: Retrieval-augmented generation (RAG) methods encounter difficulties when addressing complex questions like multi-hop queries. While iterative retrieval methods improve performance by gathering additional information, current approaches often rely on multiple calls of large language models (LLMs). In this paper, we introduce EfficientRAG, an efficient retriever for multi-hop question answering. EfficientRAG iteratively generates new queries without the need for LLM calls at each iteration and filters out irrelevant information. Experimental results demonstrate that EfficientRAG surpasses existing RAG methods on three open-domain multi-hop question-answering datasets.
Authors: Alex Gorbulev, Vasiliy Alekseev, Konstantin Vorontsov
Abstract: Topic modelling is fundamentally a soft clustering problem (of known objects -- documents, over unknown clusters -- topics). That is, the task is incorrectly posed. In particular, the topic models are unstable and incomplete. All this leads to the fact that the process of finding a good topic model (repeated hyperparameter selection, model training, and topic quality assessment) can be particularly long and labor-intensive. We aim to simplify the process, to make it more deterministic and provable. To this end, we present a method for iterative training of a topic model. The essence of the method is that a series of related topic models are trained so that each subsequent model is at least as good as the previous one, i.e., that it retains all the good topics found earlier. The connection between the models is achieved by additive regularization. The result of this iterative training is the last topic model in the series, which we call the iteratively updated additively regularized topic model (ITAR). Experiments conducted on several collections of natural language texts show that the proposed ITAR model performs better than other popular topic models (LDA, ARTM, BERTopic), its topics are diverse, and its perplexity (ability to "explain" the underlying data) is moderate.
Authors: John Mendon\c{c}a, Isabel Trancoso, Alon Lavie
Abstract: Although human evaluation remains the gold standard for open-domain dialogue evaluation, the growing popularity of automated evaluation using Large Language Models (LLMs) has also extended to dialogue. However, most frameworks leverage benchmarks that assess older chatbots on aspects such as fluency and relevance, which are not reflective of the challenges associated with contemporary models. In fact, a qualitative analysis on Soda, a GPT-3.5 generated dialogue dataset, suggests that current chatbots may exhibit several recurring issues related to coherence and commonsense knowledge, but generally produce highly fluent and relevant responses. Noting the aforementioned limitations, this paper introduces Soda-Eval, an annotated dataset based on Soda that covers over 120K turn-level assessments across 10K dialogues, where the annotations were generated by GPT-4. Using Soda-Eval as a benchmark, we then study the performance of several open-access instruction-tuned LLMs, finding that dialogue evaluation remains challenging. Fine-tuning these models improves performance over few-shot inferences, both in terms of correlation and explanation.
Authors: Lei Liang, Mengshu Sun, Zhengke Gui, Zhongshu Zhu, Zhouyu Jiang, Ling Zhong, Yuan Qu, Peilong Zhao, Zhongpu Bo, Jin Yang, Huaidong Xiong, Lin Yuan, Jun Xu, Zaoyang Wang, Zhiqiang Zhang, Wen Zhang, Huajun Chen, Wenguang Chen, Jun Zhou
Abstract: The recently developed retrieval-augmented generation (RAG) technology has enabled the efficient construction of domain-specific applications. However, it also has limitations, including the gap between vector similarity and the relevance of knowledge reasoning, as well as insensitivity to knowledge logic, such as numerical values, temporal relations, expert rules, and others, which hinder the effectiveness of professional knowledge services. In this work, we introduce a professional domain knowledge service framework called Knowledge Augmented Generation (KAG). KAG is designed to address the aforementioned challenges with the motivation of making full use of the advantages of knowledge graph(KG) and vector retrieval, and to improve generation and reasoning performance by bidirectionally enhancing large language models (LLMs) and KGs through five key aspects: (1) LLM-friendly knowledge representation, (2) mutual-indexing between knowledge graphs and original chunks, (3) logical-form-guided hybrid reasoning engine, (4) knowledge alignment with semantic reasoning, and (5) model capability enhancement for KAG. We compared KAG with existing RAG methods in multihop question answering and found that it significantly outperforms state-of-theart methods, achieving a relative improvement of 19.6% on 2wiki and 33.5% on hotpotQA in terms of F1 score. We have successfully applied KAG to two professional knowledge Q&A tasks of Ant Group, including E-Government Q&A and E-Health Q&A, achieving significant improvement in professionalism compared to RAG methods.
Authors: Michael D. Skarlinski, Sam Cox, Jon M. Laurent, James D. Braza, Michaela Hinks, Michael J. Hammerling, Manvitha Ponnapati, Samuel G. Rodriques, Andrew D. White
Abstract: Language models are known to hallucinate incorrect information, and it is unclear if they are sufficiently accurate and reliable for use in scientific research. We developed a rigorous human-AI comparison methodology to evaluate language model agents on real-world literature search tasks covering information retrieval, summarization, and contradiction detection tasks. We show that PaperQA2, a frontier language model agent optimized for improved factuality, matches or exceeds subject matter expert performance on three realistic literature research tasks without any restrictions on humans (i.e., full access to internet, search tools, and time). PaperQA2 writes cited, Wikipedia-style summaries of scientific topics that are significantly more accurate than existing, human-written Wikipedia articles. We also introduce a hard benchmark for scientific literature research called LitQA2 that guided design of PaperQA2, leading to it exceeding human performance. Finally, we apply PaperQA2 to identify contradictions within the scientific literature, an important scientific task that is challenging for humans. PaperQA2 identifies 2.34 +/- 1.99 contradictions per paper in a random subset of biology papers, of which 70% are validated by human experts. These results demonstrate that language model agents are now capable of exceeding domain experts across meaningful tasks on scientific literature.
Authors: Lemeng Qi, Yang Han, Zhuotong Xie
Abstract: This paper explores the challenges posed by nominal adjectives (NAs) in natural language processing (NLP) tasks, particularly in part-of-speech (POS) tagging. We propose treating NAs as a distinct POS tag, "JN," and investigate its impact on POS tagging, BIO chunking, and coreference resolution. Our study shows that reclassifying NAs can improve the accuracy of syntactic analysis and structural understanding in NLP. We present experimental results using Hidden Markov Models (HMMs), Maximum Entropy (MaxEnt) models, and Spacy, demonstrating the feasibility and potential benefits of this approach. Additionally we trained a bert model to identify the NA in untagged text.
Authors: Tuhin Chakrabarty, Philippe Laban, Chien-Sheng Wu
Abstract: LLM-based applications are helping people write, and LLM-generated text is making its way into social media, journalism, and our classrooms. However, the differences between LLM-generated and human-written text remain unclear. To explore this, we hired professional writers to edit paragraphs in several creative domains. We first found these writers agree on undesirable idiosyncrasies in LLM-generated text, formalizing it into a seven-category taxonomy (e.g. cliches, unnecessary exposition). Second, we curated the LAMP corpus: 1,057 LLM-generated paragraphs edited by professional writers according to our taxonomy. Analysis of LAMP reveals that none of the LLMs used in our study (GPT4o, Claude-3.5-Sonnet, Llama-3.1-70b) outperform each other in terms of writing quality, revealing common limitations across model families. Third, we explored automatic editing methods to improve LLM-generated text. A large-scale preference annotation confirms that although experts largely prefer text edited by other experts, automatic editing methods show promise in improving alignment between LLM-generated and human-written text.
Authors: Zhou Zhang, Dongzeng Tan, Jiaan Wang, Yilong Chen, Jiarong Xu
Abstract: Emojis have gained immense popularity on social platforms, serving as a common means to supplement or replace text. However, existing data mining approaches generally either completely ignore or simply treat emojis as ordinary Unicode characters, which may limit the model's ability to grasp the rich semantic information in emojis and the interaction between emojis and texts. Thus, it is necessary to release the emoji's power in social media data mining. To this end, we first construct a heterogeneous graph consisting of three types of nodes, i.e. post, word and emoji nodes to improve the representation of different elements in posts. The edges are also well-defined to model how these three elements interact with each other. To facilitate the sharing of information among post, word and emoji nodes, we propose a graph pre-train framework for text and emoji co-modeling, which contains two graph pre-training tasks: node-level graph contrastive learning and edge-level link reconstruction learning. Extensive experiments on the Xiaohongshu and Twitter datasets with two types of downstream tasks demonstrate that our approach proves significant improvement over previous strong baseline methods.
Authors: Emanuela Boros, Maud Ehrmann
Abstract: This paper investigates the presence of OCR-sensitive neurons within the Transformer architecture and their influence on named entity recognition (NER) performance on historical documents. By analysing neuron activation patterns in response to clean and noisy text inputs, we identify and then neutralise OCR-sensitive neurons to improve model performance. Based on two open access large language models (Llama2 and Mistral), experiments demonstrate the existence of OCR-sensitive regions and show improvements in NER performance on historical newspapers and classical commentaries, highlighting the potential of targeted neuron modulation to improve models' performance on noisy text.
Authors: Mohammad Abuzar Hashemi, Zhanghexuan Li, Mihir Chauhan, Yan Shen, Abhishek Satbhai, Mir Basheer Ali, Mingchen Gao, Sargur Srihari
Abstract: Pre-training visual and textual representations from large-scale image-text pairs is becoming a standard approach for many downstream vision-language tasks. The transformer-based models learn inter and intra-modal attention through a list of self-supervised learning tasks. This paper proposes LAViTeR, a novel architecture for visual and textual representation learning. The main module, Visual Textual Alignment (VTA) will be assisted by two auxiliary tasks, GAN-based image synthesis and Image Captioning. We also propose a new evaluation metric measuring the similarity between the learnt visual and textual embedding. The experimental results on two public datasets, CUB and MS-COCO, demonstrate superior visual and textual representation alignment in the joint feature embedding space
Authors: Sen Yang, Xin Li, Leyang Cui, Lidong Bing, Wai Lam
Abstract: Two lines of approaches are adopted for complex reasoning with LLMs. One line of work prompts LLMs with various reasoning structures, while the structural outputs can be naturally regarded as intermediate reasoning steps. Another line of work adopt LLM-free declarative solvers to do the reasoning task, rendering higher reasoning accuracy but lacking interpretability due to the black-box nature of the solvers. Aiming to resolve the trade-off between answer accuracy and interpretability, we present a simple extension to the latter line of work. Specifically, we showcase that the intermediate search logs generated by Prolog interpreters can be accessed and interpreted into human-readable reasoning proofs. As long as LLMs correctly translate problem descriptions into Prolog representations, the corresponding reasoning proofs are ensured to be causal and reliable. On two logical reasoning and one arithmetic reasoning datasets, our framework obtains significant improvements in terms of both answer accuracy and reasoning proof accuracy. Our code is released at https://github.com/DAMO-NLP-SG/CaRing
Authors: Wenjun Hou, Yi Cheng, Kaishuai Xu, Yan Hu, Wenjie Li, Jiang Liu
Abstract: Previous research on radiology report generation has made significant progress in terms of increasing the clinical accuracy of generated reports. In this paper, we emphasize another crucial quality that it should possess, i.e., inter-report consistency, which refers to the capability of generating consistent reports for semantically equivalent radiographs. This quality is even of greater significance than the overall report accuracy in terms of ensuring the system's credibility, as a system prone to providing conflicting results would severely erode users' trust. Regrettably, existing approaches struggle to maintain inter-report consistency, exhibiting biases towards common patterns and susceptibility to lesion variants. To address this issue, we propose ICON, which improves the inter-report consistency of radiology report generation. Aiming to enhance the system's ability to capture similarities in semantically equivalent lesions, our approach first involves extracting lesions from input images and examining their characteristics. Then, we introduce a lesion-aware mixup technique to ensure that the representations of the semantically equivalent lesions align with the same attributes, achieved through a linear combination during the training phase. Extensive experiments on three publicly available chest X-ray datasets verify the effectiveness of our approach, both in terms of improving the consistency and accuracy of the generated reports.
Authors: Hao Chen, Guoqiang Li, Minyu Chen, Ruibang Liu, Sinka Gao
Abstract: Zero-knowledge proof (ZKP) systems have surged attention and held a fundamental role in contemporary cryptography. Zero-knowledge succinct non-interactive argument of knowledge (zk-SNARK) protocols dominate the ZKP usage, implemented through arithmetic circuit programming paradigm. However, underconstrained or overconstrained circuits may lead to bugs. The former refers to circuits that lack the necessary constraints, resulting in unexpected solutions and causing the verifier to accept a bogus witness, and the latter refers to circuits that are constrained excessively, resulting in lacking necessary solutions and causing the verifier to accept no witness. This paper introduces a novel approach for pinpointing two distinct types of bugs in ZKP circuits. The method involves encoding the arithmetic circuit constraints to polynomial equation systems and solving them over finite fields by the computer algebra system. The classification of verification results is refined, greatly enhancing the expressive power of the system. A tool, AC4, is proposed to represent the implementation of the method. Experiments show that AC4 demonstrates a increase in the checked ratio, showing a 29% improvement over Picus, a checker for Circom circuits, and a 10% improvement over halo2-analyzer, a checker for halo2 circuits. Within a solvable range, the checking time has also exhibited noticeable improvement, demonstrating a magnitude increase compared to previous efforts.
Authors: Fengran Mo, Abbas Ghaddar, Kelong Mao, Mehdi Rezagholizadeh, Boxing Chen, Qun Liu, Jian-Yun Nie
Abstract: In this paper, we study how open-source large language models (LLMs) can be effectively deployed for improving query rewriting in conversational search, especially for ambiguous queries. We introduce CHIQ, a two-step method that leverages the capabilities of LLMs to resolve ambiguities in the conversation history before query rewriting. This approach contrasts with prior studies that predominantly use closed-source LLMs to directly generate search queries from conversation history. We demonstrate on five well-established benchmarks that CHIQ leads to state-of-the-art results across most settings, showing highly competitive performances with systems leveraging closed-source LLMs. Our study provides a first step towards leveraging open-source LLMs in conversational search, as a competitive alternative to the prevailing reliance on commercial LLMs. Data, models, and source code will be publicly available upon acceptance at https://github.com/fengranMark/CHIQ.
Authors: Ayush Kaushal, Tejas Vaidhya, Tejas Pandey, Aaryan Bhagat, Irina Rish
Abstract: Post-training quantization is the leading method for addressing memory-related bottlenecks in LLM inference, but unfortunately, it suffers from significant performance degradation below 4-bit precision. An alternative approach involves training compressed models directly at a low bitwidth (e.g., binary or ternary models). However, the performance, training dynamics, and scaling trends of such models are not yet well understood. To address this issue, we train and openly release the Spectra LLM suite consisting of 54 language models ranging from 99M to 3.9B parameters, trained on 300B tokens. Spectra includes FloatLMs, post-training quantized QuantLMs (3, 4, 6, and 8 bits), and ternary LLMs (TriLMs) - our improved architecture for ternary language modeling, which significantly outperforms previously proposed ternary models of a given size (in bits), matching half-precision models at scale. For example, TriLM 3.9B is (bit-wise) smaller than the half-precision FloatLM 830M, but matches half-precision FloatLM 3.9B in commonsense reasoning and knowledge benchmarks. However, TriLM 3.9B is also as toxic and stereotyping as FloatLM 3.9B, a model six times larger in size. Additionally, TriLM 3.9B lags behind FloatLM in perplexity on validation splits and web-based corpora but performs better on less noisy datasets like Lambada and PennTreeBank. To enhance understanding of low-bitwidth models, we are releasing 500+ intermediate checkpoints of the Spectra suite at \href{https://github.com/NolanoOrg/SpectraSuite}{https://github.com/NolanoOrg/SpectraSuite}.
URLs: https://github.com/NolanoOrg/SpectraSuite, https://github.com/NolanoOrg/SpectraSuite
Authors: Yanxi Chen, Yaliang Li, Bolin Ding, Jingren Zhou
Abstract: We initiate a formal investigation into the design and analysis of LLM-based algorithms, i.e. algorithms that contain one or multiple calls of large language models (LLMs) as sub-routines and critically rely on the capabilities of LLMs. While LLM-based algorithms, ranging from basic LLM calls with prompt engineering to complicated LLM-powered agent systems and compound AI systems, have achieved remarkable empirical success, the design and optimization of them have mostly relied on heuristics and trial-and-errors, which is largely due to a lack of formal and analytical study for these algorithms. To fill this gap, we start by identifying the computational-graph representation of LLM-based algorithms, the design principle of task decomposition, and some key abstractions, which then facilitate our formal analysis for the accuracy and efficiency of LLM-based algorithms, despite the black-box nature of LLMs. Through extensive analytical and empirical investigation in a series of case studies, we demonstrate that the proposed framework is broadly applicable to a wide range of scenarios and diverse patterns of LLM-based algorithms, such as parallel, hierarchical and recursive task decomposition. Our proposed framework holds promise for advancing LLM-based algorithms, by revealing the reasons behind curious empirical phenomena, guiding the choices of hyperparameters, predicting the empirical performance of algorithms, and inspiring new algorithm design. To promote further study of LLM-based algorithms, we release our source code at https://github.com/modelscope/agentscope/tree/main/examples/paper_llm_based_algorithm.
URLs: https://github.com/modelscope/agentscope/tree/main/examples/paper_llm_based_algorithm.
Authors: Tong Yang, Yu Huang, Yingbin Liang, Yuejie Chi
Abstract: In-context learning (ICL) refers to a remarkable capability of pretrained large language models, which can learn a new task given a few examples during inference. However, theoretical understanding of ICL is largely under-explored, particularly whether transformers can be trained to generalize to unseen examples in a prompt, which will require the model to acquire contextual knowledge of the prompt for generalization. This paper investigates the training dynamics of transformers by gradient descent through the lens of non-linear regression tasks. The contextual generalization here can be attained via learning the template function for each task in-context, where all template functions lie in a linear space with $m$ basis functions. We analyze the training dynamics of one-layer multi-head transformers to in-contextly predict unlabeled inputs given partially labeled prompts, where the labels contain Gaussian noise and the number of examples in each prompt are not sufficient to determine the template. Under mild assumptions, we show that the training loss for a one-layer multi-head transformer converges linearly to a global minimum. Moreover, the transformer effectively learns to perform ridge regression over the basis functions. To our knowledge, this study is the first provable demonstration that transformers can learn contextual (i.e., template) information to generalize to both unseen examples and tasks when prompts contain only a small number of query-answer pairs.
Authors: Yinmin Zhong, Zili Zhang, Bingyang Wu, Shengyu Liu, Yukun Chen, Changyi Wan, Hanpeng Hu, Lei Xia, Ranchen Ming, Yibo Zhu, Xin Jin
Abstract: Reinforcement Learning from Human Feedback (RLHF) enhances the alignment between LLMs and human preference. The workflow of RLHF typically involves several models and tasks in a series of distinct stages. Existing RLHF training systems view each task as the smallest execution unit thus overlooking the opportunities for subtask-level optimizations. Due to the intrinsic nature of RLHF training, i.e., the data skewness in the generation stage, and the pipeline bubbles in the training stage, existing RLHF systems suffer from low GPU utilization in production deployments. RLHFuse breaks the traditional view of RLHF workflow as a composition of individual tasks, splitting each task into finer-grained subtasks, and performing stage fusion to improve GPU utilization. RLHFuse contains two key ideas. First, for generation and inference tasks, RLHFuse splits them into sample-level subtasks, enabling efficient inter-stage fusion to mitigate the original generation bottleneck dominated by long-tailed samples. Second, for training tasks, RLHFuse breaks them into subtasks of micro-batches. By leveraging the intuition that pipeline execution can be essentially complemented by another pipeline, RLHFuse performs intra-stage fusion to concurrently execute these subtasks in the training stage with a fused pipeline schedule, resulting in fewer pipeline bubbles. In addition, RLHFuse incorporates a series of system optimizations tailored for each stage of RLHF, making it efficient and scalable for our internal product usage. We evaluate RLHFuse on various popular LLMs and the results show that RLHFuse increases the training throughput by up to 3.7x, compared to existing state-of-the-art systems.
Authors: Yu Zhang, Changhao Pan, Wenxiang Guo, Ruiqi Li, Zhiyuan Zhu, Jialei Wang, Wenhao Xu, Jingyu Lu, Zhiqing Hong, Chuxin Wang, LiChao Zhang, Jinzheng He, Ziyue Jiang, Yuxin Chen, Chen Yang, Jiecheng Zhou, Xinyu Cheng, Zhou Zhao
Abstract: The scarcity of high-quality and multi-task singing datasets significantly hinders the development of diverse controllable and personalized singing tasks, as existing singing datasets suffer from low quality, limited diversity of languages and singers, absence of multi-technique information and realistic music scores, and poor task suitability. To tackle these problems, we present GTSinger, a large global, multi-technique, free-to-use, high-quality singing corpus with realistic music scores, designed for all singing tasks, along with its benchmarks. Particularly, (1) we collect 80.59 hours of high-quality singing voices, forming the largest recorded singing dataset; (2) 20 professional singers across nine widely spoken languages offer diverse timbres and styles; (3) we provide controlled comparison and phoneme-level annotations of six commonly used singing techniques, helping technique modeling and control; (4) GTSinger offers realistic music scores, assisting real-world musical composition; (5) singing voices are accompanied by manual phoneme-to-audio alignments, global style labels, and 16.16 hours of paired speech for various singing tasks. Moreover, to facilitate the use of GTSinger, we conduct four benchmark experiments: technique-controllable singing voice synthesis, technique recognition, style transfer, and speech-to-singing conversion. The corpus and demos can be found at http://gtsinger.github.io. We provide the dataset and the code for processing data and conducting benchmarks at https://huggingface.co/datasets/GTSinger/GTSinger and https://github.com/GTSinger/GTSinger.
URLs: http://gtsinger.github.io., https://huggingface.co/datasets/GTSinger/GTSinger, https://github.com/GTSinger/GTSinger.
Authors: Jon Saad-Falcon, Adrian Gamarra Lafuente, Shlok Natarajan, Nahum Maru, Hristo Todorov, Etash Guha, E. Kelly Buchanan, Mayee Chen, Neel Guha, Christopher R\'e, Azalia Mirhoseini
Abstract: Inference-time techniques are emerging as highly effective tools to increase large language model (LLM) capabilities. However, there is still limited understanding of the best practices for developing systems that combine inference-time techniques with one or more LLMs, with challenges including: (1) effectively allocating inference compute budget, (2) understanding the interactions between different combinations of inference-time techniques and their impact on downstream performance, and 3) efficiently searching over the large space of model choices, inference-time techniques, and their compositions. To address these challenges, we introduce Archon, an automated framework for designing inference-time architectures. Archon defines an extensible design space, encompassing methods such as generation ensembling, multi-sampling, ranking, fusion, critiquing, verification, and unit testing. It then transforms the problem of selecting and combining LLMs and inference-time techniques into a hyperparameter optimization objective. To optimize this objective, we introduce automated Inference-Time Architecture Search (ITAS) algorithms. Given target benchmark(s), an inference compute budget, and available LLMs, ITAS outputs optimized architectures. We evaluate Archon architectures across a wide range of instruction-following and reasoning benchmarks, including MT-Bench, Arena-Hard-Auto, AlpacaEval 2.0, MixEval, MixEval Hard, MATH, and CodeContests. We show that automatically designed inference-time architectures by Archon outperform strong models such as GPT-4o and Claude 3.5 Sonnet on these benchmarks, achieving an average increase of 15.1 and 11.2 percentage points with all-source models and open-source models, respectively. We make our code and datasets available publicly on Github: https://github.com/ScalingIntelligence/Archon.
Authors: R. Stuart Geiger, Flynn O'Sullivan, Elsie Wang, Jonathan Lo
Abstract: We conducted controlled experimental bias audits for four versions of ChatGPT, which we asked to recommend an opening offer in salary negotiations for a new hire. We submitted 98,800 prompts to each version, systematically varying the employee's gender, university, and major, and tested prompts in voice of each side of the negotiation: the employee versus employer. We find ChatGPT as a multi-model platform is not robust and consistent enough to be trusted for such a task. We observed statistically significant salary offers when varying gender for all four models, although with smaller gaps than for other attributes tested. The largest gaps were different model versions and between the employee- vs employer-voiced prompts. We also observed substantial gaps when varying university and major, but many of the biases were not consistent across model versions. We tested for fictional and fraudulent universities and found wildly inconsistent results across cases and model versions. We make broader contributions to the AI/ML fairness literature. Our scenario and our experimental design differ from mainstream AI/ML auditing efforts in key ways. Bias audits typically test discrimination for protected classes like gender, which we contrast with testing non-protected classes of university and major. Asking for negotiation advice includes how aggressive one ought to be in a negotiation relative to known empirical salary distributions and scales, which is a deeply contextual and personalized task that has no objective ground truth to validate. These results raise concerns for the specific model versions we tested and ChatGPT as a multi-model platform in continuous development. Our epistemology does not permit us to definitively certify these models as either generally biased or unbiased on the attributes we test, but our study raises matters of concern for stakeholders to further investigate.
Authors: Yu Zhang, Ziyue Jiang, Ruiqi Li, Changhao Pan, Jinzheng He, Rongjie Huang, Chuxin Wang, Zhou Zhao
Abstract: Zero-shot singing voice synthesis (SVS) with style transfer and style control aims to generate high-quality singing voices with unseen timbres and styles (including singing method, emotion, rhythm, technique, and pronunciation) from audio and text prompts. However, the multifaceted nature of singing styles poses a significant challenge for effective modeling, transfer, and control. Furthermore, current SVS models often fail to generate singing voices rich in stylistic nuances for unseen singers. To address these challenges, we introduce TCSinger, the first zero-shot SVS model for style transfer across cross-lingual speech and singing styles, along with multi-level style control. Specifically, TCSinger proposes three primary modules: 1) the clustering style encoder employs a clustering vector quantization model to stably condense style information into a compact latent space; 2) the Style and Duration Language Model (S\&D-LM) concurrently predicts style information and phoneme duration, which benefits both; 3) the style adaptive decoder uses a novel mel-style adaptive normalization method to generate singing voices with enhanced details. Experimental results show that TCSinger outperforms all baseline models in synthesis quality, singer similarity, and style controllability across various tasks, including zero-shot style transfer, multi-level style control, cross-lingual style transfer, and speech-to-singing style transfer. Singing voice samples can be accessed at https://tcsinger.github.io/.
Authors: Shadi Iskander, Nachshon Cohen, Zohar Karnin, Ori Shapira, Sofia Tolmach
Abstract: Training large language models (LLMs) for external tool usage is a rapidly expanding field, with recent research focusing on generating synthetic data to address the shortage of available data. However, the absence of systematic data quality checks poses complications for properly training and testing models. To that end, we propose two approaches for assessing the reliability of data for training LLMs to use external tools. The first approach uses intuitive, human-defined correctness criteria. The second approach uses a model-driven assessment with in-context evaluation. We conduct a thorough evaluation of data quality on two popular benchmarks, followed by an extrinsic evaluation that showcases the impact of data quality on model performance. Our results demonstrate that models trained on high-quality data outperform those trained on unvalidated data, even when trained with a smaller quantity of data. These findings empirically support the significance of assessing and ensuring the reliability of training data for tool-using LLMs.