Authors: Micha{\l} Pietruszka, {\L}ukasz Borchmann, Aleksander J\k{e}drosz, Pawe{\l} Morawiecki
Abstract: We present a benchmark for large language models designed to tackle one of the most knowledge-intensive tasks in data science: writing feature engineering code, which requires domain knowledge in addition to a deep understanding of the underlying problem and data structure. The model is provided with a dataset description in a prompt and asked to generate code transforming it. The evaluation score is derived from the improvement achieved by an XGBoost model fit on the modified dataset compared to the original data. By an extensive evaluation of state-of-the-art models and comparison to well-established benchmarks, we demonstrate that the FeatEng of our proposal can cheaply and efficiently assess the broad capabilities of LLMs, in contrast to the existing methods.
Authors: Keiichi Namikoshi, David A. Shamma, Rumen Iliev, Jingchao Fang, Alexandre Filipowicz, Candice L Hogan, Charlene Wu, Nikos Arechiga
Abstract: Behavior change interventions are important to coordinate societal action across a wide array of important applications, including the adoption of electrified vehicles to reduce emissions. Prior work has demonstrated that interventions for behavior must be personalized, and that the intervention that is most effective on average across a large group can result in a backlash effect that strengthens opposition among some subgroups. Thus, it is important to target interventions to different audiences, and to present them in a natural, conversational style. In this context, an important emerging application domain for large language models (LLMs) is conversational interventions for behavior change. In this work, we leverage prior work on understanding values motivating the adoption of battery electric vehicles. We leverage new advances in LLMs, combined with a contextual bandit, to develop conversational interventions that are personalized to the values of each study participant. We use a contextual bandit algorithm to learn to target values based on the demographics of each participant. To train our bandit algorithm in an offline manner, we leverage LLMs to play the role of study participants. We benchmark the persuasive effectiveness of our bandit-enhanced LLM against an unaided LLM generating conversational interventions without demographic-targeted values.
Authors: Yue Huang, Zhengqing Yuan, Yujun Zhou, Kehan Guo, Xiangqi Wang, Haomin Zhuang, Weixiang Sun, Lichao Sun, Jindong Wang, Yanfang Ye, Xiangliang Zhang
Abstract: Large Language Models (LLMs) are increasingly employed for simulations, enabling applications in role-playing agents and Computational Social Science (CSS). However, the reliability of these simulations is under-explored, which raises concerns about the trustworthiness of LLMs in these applications. In this paper, we aim to answer ``How reliable is LLM-based simulation?'' To address this, we introduce TrustSim, an evaluation dataset covering 10 CSS-related topics, to systematically investigate the reliability of the LLM simulation. We conducted experiments on 14 LLMs and found that inconsistencies persist in the LLM-based simulated roles. In addition, the consistency level of LLMs does not strongly correlate with their general performance. To enhance the reliability of LLMs in simulation, we proposed Adaptive Learning Rate Based ORPO (AdaORPO), a reinforcement learning-based algorithm to improve the reliability in simulation across 7 LLMs. Our research provides a foundation for future studies to explore more robust and trustworthy LLM-based simulations.
Authors: Vicky Dong, Hao Yu, Yao Chen
Abstract: This study introduces a novel approach to sentence-level relation extraction (RE) that integrates Graph Neural Networks (GNNs) with Large Language Models (LLMs) to generate contextually enriched support documents. By harnessing the power of LLMs to generate auxiliary information, our approach crafts an intricate graph representation of textual data. This graph is subsequently processed through a Graph Neural Network (GNN) to refine and enrich the embeddings associated with each entity ensuring a more nuanced and interconnected understanding of the data. This methodology addresses the limitations of traditional sentence-level RE models by incorporating broader contexts and leveraging inter-entity interactions, thereby improving the model's ability to capture complex relationships across sentences. Our experiments, conducted on the CrossRE dataset, demonstrate the effectiveness of our approach, with notable improvements in performance across various domains. The results underscore the potential of combining GNNs with LLM-generated context to advance the field of relation extraction.
Authors: Gabrielle Kaili-May Liu, Bowen Shi, Avi Caciularu, Idan Szpektor, Arman Cohan
Abstract: Multi-document (MD) processing is crucial for LLMs to handle real-world tasks such as summarization and question-answering across large sets of documents. While LLMs have improved at processing long inputs, MD contexts still present challenges, such as managing inter-document dependencies, redundancy, and incoherent structures. We introduce MDCure, a scalable and effective fine-tuning pipeline to enhance the MD capabilities of LLMs without the computational cost of pre-training or reliance on human annotated data. MDCure is based on generation of high-quality synthetic MD instruction data from sets of related articles via targeted prompts. We further introduce MDCureRM, a multi-objective reward model which filters generated data based on their training utility for MD settings. With MDCure, we fine-tune a variety of LLMs, from the FlanT5, Qwen2, and LLAMA3.1 model families, up to 70B parameters in size. Extensive evaluations on a wide range of MD and long-context benchmarks spanning various tasks show MDCure consistently improves performance over pre-trained baselines and over corresponding base models by up to 75.5%. Our code, datasets, and models are available at https://github.com/yale-nlp/MDCure.
Authors: Sireesh Gururaja, Yueheng Zhang, Guannan Tang, Tianhao Zhang, Kevin Murphy, Yu-Tsen Yi, Junwon Seo, Anthony Rollett, Emma Strubell
Abstract: Recent years in NLP have seen the continued development of domain-specific information extraction tools for scientific documents, alongside the release of increasingly multimodal pretrained transformer models. While the opportunity for scientists outside of NLP to evaluate and apply such systems to their own domains has never been clearer, these models are difficult to compare: they accept different input formats, are often black-box and give little insight into processing failures, and rarely handle PDF documents, the most common format of scientific publication. In this work, we present Collage, a tool designed for rapid prototyping, visualization, and evaluation of different information extraction models on scientific PDFs. Collage allows the use and evaluation of any HuggingFace token classifier, several LLMs, and multiple other task-specific models out of the box, and provides extensible software interfaces to accelerate experimentation with new models. Further, we enable both developers and users of NLP-based tools to inspect, debug, and better understand modeling pipelines by providing granular views of intermediate states of processing. We demonstrate our system in the context of information extraction to assist with literature review in materials science.
Authors: Guangliang Liu, Zhiyu Xue, Rongrong Wang, Kristen Marie Johnson
Abstract: Self-correction is one of the most amazing emerging capabilities of Large Language Models (LLMs), enabling LLMs to self-modify an inappropriate output given a natural language feedback which describes the problems of that output. Moral self-correction is a post-hoc approach correcting unethical generations without requiring a gradient update, making it both computationally lightweight and capable of preserving the language modeling ability. Previous works have shown that LLMs can self-debias, and it has been reported that small models, i.e., those with less than 22B parameters, are not capable of moral self-correction. However, there is no direct proof as to why such smaller models fall short of moral self-correction, though previous research hypothesizes that larger models are skilled in following instructions and understanding abstract social norms. In this paper, we empirically validate this hypothesis in the context of social stereotyping, through meticulous prompting. Our experimental results indicate that (i) surprisingly, 3.8B LLMs with proper safety alignment fine-tuning can achieve very good moral self-correction performance, highlighting the significant effects of safety alignment; and (ii) small LLMs are indeed weaker than larger-scale models in terms of comprehending social norms and self-explanation through CoT, but all scales of LLMs show bad self-correction performance given unethical instructions.
Authors: Muhammad Reza Qorib, Alham Fikri Aji, Hwee Tou Ng
Abstract: Error type information has been widely used to improve the performance of grammatical error correction (GEC) models, whether for generating corrections, re-ranking them, or combining GEC models. Combining GEC models that have complementary strengths in correcting different error types is very effective in producing better corrections. However, system combination incurs a high computational cost due to the need to run inference on the base systems before running the combination method itself. Therefore, it would be more efficient to have a single model with multiple sub-networks that specialize in correcting different error types. In this paper, we propose a mixture-of-experts model, MoECE, for grammatical error correction. Our model successfully achieves the performance of T5-XL with three times fewer effective parameters. Additionally, our model produces interpretable corrections by also identifying the error type during inference.
Authors: Tanmay Parekh, Pradyot Prakash, Alexander Radovic, Akshay Shekher, Denis Savenkov
Abstract: Research has shown the effectiveness of reasoning (e.g., Chain-of-Thought), planning (e.g., SelfAsk), and retrieval augmented generation strategies to improve the performance of Large Language Models (LLMs) on various tasks, such as question answering. However, using a single fixed strategy to answer different kinds of questions is suboptimal in performance and inefficient in terms of generated output tokens and performed retrievals. In our work, we propose a novel technique DyPlan, to induce a dynamic strategy selection process in LLMs, to improve performance and reduce costs in question-answering. DyPlan incorporates an initial decision step to select the most suitable strategy conditioned on the input question and guides the LLM's response generation accordingly. We extend DyPlan to DyPlan-verify, adding an internal verification and correction process to further enrich the generated answer. Experiments on three prominent multi-hop question answering (MHQA) datasets reveal how DyPlan can improve model performance by 7-13% while reducing the cost by 11-32% relative to the best baseline model.
Authors: Matej Klemen, Martin Bo\v{z}i\v{c}, \v{S}pela Arhar Holdt, Marko Robnik-\v{S}ikonja
Abstract: Spell-checkers are valuable tools that enhance communication by identifying misspelled words in written texts. Recent improvements in deep learning, and in particular in large language models, have opened new opportunities to improve traditional spell-checkers with new functionalities that not only assess spelling correctness but also the suitability of a word for a given context. In our work, we present and compare two new spell-checkers and evaluate them on synthetic, learner, and more general-domain Slovene datasets. The first spell-checker is a traditional, fast, word-based approach, based on a morphological lexicon with a significantly larger word list compared to existing spell-checkers. The second approach uses a language model trained on a large corpus with synthetically inserted errors. We present the training data construction strategies, which turn out to be a crucial component of neural spell-checkers. Further, the proposed neural model significantly outperforms all existing spell-checkers for Slovene in both precision and recall.
Authors: Hieu Tran, Junda Wang, Yujan Ting, Weijing Huang, Terrence Chen
Abstract: Large language models (LLMs) have shown remarkable capabilities in various natural language processing tasks, yet they often struggle with maintaining factual accuracy, particularly in knowledge-intensive domains like healthcare. This study introduces LEAF: Learning and Evaluation Augmented by Fact-Checking, a novel approach designed to enhance the factual reliability of LLMs, with a focus on medical question answering (QA). LEAF utilizes a dual strategy to enhance the factual accuracy of responses from models such as Llama 3 70B Instruct and Llama 3 8B Instruct. The first strategy, Fact-Check-Then-RAG, improves Retrieval-Augmented Generation (RAG) by incorporating fact-checking results to guide the retrieval process without updating model parameters. The second strategy, Learning from Fact-Checks via Self-Training, involves supervised fine-tuning (SFT) on fact-checked responses or applying Simple Preference Optimization (SimPO) with fact-checking as a ranking mechanism, both updating LLM parameters from supervision. These findings suggest that integrating fact-checked responses whether through RAG enhancement or self-training enhances the reliability and factual correctness of LLM outputs, offering a promising solution for applications where information accuracy is crucial.
Authors: Hajar Sakai, Sarah S. Lam, Mohammadsadegh Mikaeili, Joshua Bosire, Franziska Jovin
Abstract: Patient experience and care quality are crucial for a hospital's sustainability and reputation. The analysis of patient feedback offers valuable insight into patient satisfaction and outcomes. However, the unstructured nature of these comments poses challenges for traditional machine learning methods following a supervised learning paradigm. This is due to the unavailability of labeled data and the nuances these texts encompass. This research explores leveraging Large Language Models (LLMs) in conducting Multi-label Text Classification (MLTC) of inpatient comments shared after a stay in the hospital. GPT-4o-Turbo was leveraged to conduct the classification. However, given the sensitive nature of patients' comments, a security layer is introduced before feeding the data to the LLM through a Protected Health Information (PHI) detection framework, which ensures patients' de-identification. Additionally, using the prompt engineering framework, zero-shot learning, in-context learning, and chain-of-thought prompting were experimented with. Results demonstrate that GPT-4o-Turbo, whether following a zero-shot or few-shot setting, outperforms traditional methods and Pre-trained Language Models (PLMs) and achieves the highest overall performance with an F1-score of 76.12% and a weighted F1-score of 73.61% followed closely by the few-shot learning results. Subsequently, the results' association with other patient experience structured variables (e.g., rating) was conducted. The study enhances MLTC through the application of LLMs, offering healthcare practitioners an efficient method to gain deeper insights into patient feedback and deliver prompt, appropriate responses.
Authors: Daniel Philipov, Vardhan Dongre, Gokhan Tur, Dilek Hakkani-T\"ur
Abstract: Embodied agents designed to assist users with tasks must engage in natural language interactions, interpret instructions, execute actions, and communicate effectively to resolve issues. However, collecting large-scale, diverse datasets of situated human-robot dialogues to train and evaluate such agents is expensive, labor-intensive, and time-consuming. To address this challenge, we propose building a large language model (LLM)-based user agent that can simulate user behavior during interactions with an embodied agent in a virtual environment. Given a user goal (e.g., make breakfast), at each time step, the user agent may observe" the robot actions or speak" to either intervene with the robot or answer questions. Such a user agent assists in improving the scalability and efficiency of embodied dialogues dataset generation and is critical for enhancing and evaluating the robot's interaction and task completion ability, as well as for research in reinforcement learning using AI feedback. We evaluate our user agent's ability to generate human-like behaviors by comparing its simulated dialogues with the TEACh dataset. We perform three experiments: zero-shot prompting to predict dialogue acts, few-shot prompting, and fine-tuning on the TEACh training subset. Results show the LLM-based user agent achieves an F-measure of 42% with zero-shot prompting and 43.4% with few-shot prompting in mimicking human speaking behavior. Through fine-tuning, performance in deciding when to speak remained stable, while deciding what to say improved from 51.1% to 62.5%. These findings showcase the feasibility of the proposed approach for assessing and enhancing the effectiveness of robot task completion through natural language communication.
Authors: Nalin Tiwary, Vardhan Dongre, Sanil Arun Chawla, Ashwin Lamani, Dilek Hakkani-T\"ur
Abstract: Recent advancements in Large Language Model (LLM)-based frameworks have extended their capabilities to complex real-world applications, such as interactive web navigation. These systems, driven by user commands, navigate web browsers to complete tasks through multi-turn dialogues, offering both innovative opportunities and significant challenges. Despite the introduction of benchmarks for conversational web navigation, a detailed understanding of the key contextual components that influence the performance of these agents remains elusive. This study aims to fill this gap by analyzing the various contextual elements crucial to the functioning of web navigation agents. We investigate the optimization of context management, focusing on the influence of interaction history and web page representation. Our work highlights improved agent performance across out-of-distribution scenarios, including unseen websites, categories, and geographic locations through effective context management. These findings provide insights into the design and optimization of LLM-based agents, enabling more accurate and effective web navigation in real-world applications.
Authors: Farshad Noravesh
Abstract: State-of-the-art models for relation extraction (RE) in the biomedical domain consider finetuning BioBERT using classification, but they may suffer from the anisotropy problem. Contrastive learning methods can reduce this anisotropy phenomena, and also help to avoid class collapse in any classification problem. In the present paper, a new training method called biological non-contrastive relation extraction (BioNCERE) is introduced for relation extraction without using any named entity labels for training to reduce annotation costs. BioNCERE uses transfer learning and non-contrastive learning to avoid full or dimensional collapse as well as bypass overfitting. It resolves RE in three stages by leveraging transfer learning two times. By freezing the weights learned in previous stages in the proposed pipeline and by leveraging non-contrastive learning in the second stage, the model predicts relations without any knowledge of named entities. Experiments have been done on SemMedDB that are almost similar to State-of-the-art performance on RE without using the information of named entities.
Authors: Shuyang Yu, Runxue Bao, Parminder Bhatia, Taha Kass-Hout, Jiayu Zhou, Cao Xiao
Abstract: Large language models (LLMs) can learn vast amounts of knowledge from diverse domains during pre-training. However, long-tail knowledge from specialized domains is often scarce and underrepresented, rarely appearing in the models' memorization. Prior work has shown that in-context learning (ICL) with retriever augmentation can help LLMs better capture long-tail knowledge, reducing their reliance on pre-trained data. Despite these advances, we observe that LLM predictions for long-tail questions remain uncertain to variations in retrieved samples. To take advantage of the uncertainty in ICL for guiding LLM predictions toward correct answers on long-tail samples, we propose a reinforcement learning-based dynamic uncertainty ranking method for ICL that accounts for the varying impact of each retrieved sample on LLM predictions. Our approach prioritizes more informative and stable samples while demoting misleading ones, updating rankings based on the feedback from the LLM w.r.t. each retrieved sample. To enhance training efficiency and reduce query costs, we introduce a learnable dynamic ranking threshold, adjusted when the model encounters negative prediction shifts. Experimental results on various question-answering datasets from different domains show that our method outperforms the best baseline by $2.76\%$, with a notable $5.96\%$ boost in accuracy on long-tail questions that elude zero-shot inference.
Authors: David Wan, Jesse Vig, Mohit Bansal, Shafiq Joty
Abstract: Large Language Models (LLMs) often exhibit positional bias in long-context settings, under-attending to information in the middle of inputs. We investigate the presence of this bias in long-form summarization, its impact on faithfulness, and various techniques to mitigate this bias. To consistently evaluate faithfulness, we first compile a benchmark of eight human-annotated long-form summarization datasets and perform a meta-evaluation of faithfulness metrics. We show that LLM-based faithfulness metrics, though effective with full-context inputs, remain sensitive to document order, indicating positional bias. Analyzing LLM-generated summaries across six datasets, we find a "U-shaped" trend in faithfulness, where LLMs faithfully summarize the beginning and end of documents but neglect middle content. Perturbing document order similarly reveals models are less faithful when important documents are placed in the middle of the input. We find that this behavior is partly due to shifting focus with context length: as context increases, summaries become less faithful, but beyond a certain length, faithfulness improves as the model focuses on the end. Finally, we experiment with different generation techniques to reduce positional bias and find that prompting techniques effectively direct model attention to specific positions, whereas more sophisticated approaches offer limited improvements. Our data and code are available in https://github.com/meetdavidwan/longformfact.
Authors: I\~nigo Parra
Abstract: This study investigates the impact of morphological typology on tokenization and language modeling performance. We focus on languages with synthetic and analytical morphological structures and examine their productivity when tokenized using the byte-pair encoding (BPE) algorithm. We compare the performance of models trained with similar amounts of data in different languages. Our experiments reveal that languages with synthetic features exhibit greater subword regularity and productivity with BPE tokenization and achieve better results in language modeling tasks. We also observe that the typological continuum from linguistic theory is reflected in several experiments. These findings suggest a correlation between morphological typology and BPE tokenization efficiency.
Authors: David Koeplinger, Darshan Gandhi, Pushkar Nandkar, Nathan Sheeley, Matheen Musaddiq, Leon Zhang, Reid Goodbar, Matthew Shaffer, Han Wang, Angela Wang, Mingran Wang, Raghu Prabhakar
Abstract: Token generation speed is critical to power the next wave of AI inference applications. GPUs significantly underperform during token generation due to synchronization overheads at kernel boundaries, utilizing only 21% of their peak memory bandwidth. While recent dataflow architectures mitigate these overheads by enabling aggressive fusion of decoder layers into a single kernel, they too leave performance on the table due to synchronization penalties at layer boundaries. This paper presents kernel looping, a specialized global optimization technique which exploits an optimization opportunity brought by combining the unique layer-level fusion possible in modern dataflow architectures with the repeated layer structure found in language models. Kernel looping eliminates synchronization costs between consecutive calls to the same kernel by transforming these calls into a single call to a modified kernel containing a pipelined outer loop. We evaluate kernel looping on the SambaNova SN40L Reconfigurable Dataflow Unit (RDU), a commercial dataflow accelerator for AI. Experiments demonstrate that kernel looping speeds up the decode phase of a wide array of powerful open-source models by up to 2.2$\times$ on SN40L. Kernel looping allows scaling of decode performance over multiple SN40L sockets, achieving speedups of up to 2.5$\times$. Finally, kernel looping enables SN40L to achieve over 90% of peak performance on 8 and 16 sockets and achieve a speedup of up to 3.7$\times$ over DGX H100. Kernel looping, as well as the models evaluated in this paper, are deployed in production in a commercial AI inference cloud.
Authors: Zheng Chen, Buhui Yao
Abstract: Goal hijacking is a type of adversarial attack on Large Language Models (LLMs) where the objective is to manipulate the model into producing a specific, predetermined output, regardless of the user's original input. In goal hijacking, an attacker typically appends a carefully crafted malicious suffix to the user's prompt, which coerces the model into ignoring the user's original input and generating the target response. In this paper, we introduce a novel goal hijacking attack method called Pseudo-Conversation Injection, which leverages the weaknesses of LLMs in role identification within conversation contexts. Specifically, we construct the suffix by fabricating responses from the LLM to the user's initial prompt, followed by a prompt for a malicious new task. This leads the model to perceive the initial prompt and fabricated response as a completed conversation, thereby executing the new, falsified prompt. Following this approach, we propose three Pseudo-Conversation construction strategies: Targeted Pseudo-Conversation, Universal Pseudo-Conversation, and Robust Pseudo-Conversation. These strategies are designed to achieve effective goal hijacking across various scenarios. Our experiments, conducted on two mainstream LLM platforms including ChatGPT and Qwen, demonstrate that our proposed method significantly outperforms existing approaches in terms of attack effectiveness.
Authors: Eugene Jang, Kimin Lee, Jin-Woo Chung, Keuntae Park, Seungwon Shin
Abstract: Tokenization is a crucial step that bridges human-readable text with model-readable discrete tokens. However, recent studies have revealed that tokenizers can be exploited to elicit unwanted model behaviors. In this work, we investigate incomplete tokens, i.e., undecodable tokens with stray bytes resulting from byte-level byte-pair encoding (BPE) tokenization. We hypothesize that such tokens are heavily reliant on their adjacent tokens and are fragile when paired with unfamiliar tokens. To demonstrate this vulnerability, we introduce improbable bigrams: out-of-distribution combinations of incomplete tokens designed to exploit their dependency. Our experiments show that improbable bigrams are significantly prone to hallucinatory behaviors. Surprisingly, alternative tokenizations of the same phrases result in drastically lower rates of hallucination (93% reduction in Llama3.1). We caution against the potential vulnerabilities introduced by byte-level BPE tokenizers, which may impede the development of trustworthy language models.
Authors: Peizhi Tang, Chuang Yang, Tong Xing, Xiaohang Xu, Renhe Jiang, Kaoru Sezaki
Abstract: Human mobility prediction plays a critical role in applications such as disaster response, urban planning, and epidemic forecasting. Traditional methods often rely on designing crafted, domain-specific models, and typically focus on short-term predictions, which struggle to generalize across diverse urban environments. In this study, we introduce Llama-3-8B-Mob, a large language model fine-tuned with instruction tuning, for long-term citywide mobility prediction -- in a Q&A manner. We validate our approach using large-scale human mobility data from four metropolitan areas in Japan, focusing on predicting individual trajectories over the next 15 days. The results demonstrate that Llama-3-8B-Mob excels in modeling long-term human mobility -- surpassing the state-of-the-art on multiple prediction metrics. It also displays strong zero-shot generalization capabilities -- effectively generalizing to other cities even when fine-tuned only on limited samples from a single city. Source codes are available at https://github.com/TANGHULU6/Llama3-8B-Mob.
Authors: Irina Tolstykh, Aleksandra Tsybina, Sergey Yakubson, Aleksandr Gordeev, Vladimir Dokholyan, Maksim Kuprashevich
Abstract: With the increasing quality and spread of LLM-based assistants, the amount of artificially generated content is growing rapidly. In many cases and tasks, such texts are already indistinguishable from those written by humans, and the quality of generation tends to only increase. At the same time, detection methods are developing more slowly, making it challenging to prevent misuse of these technologies. In this work, we investigate the task of generated text detection by proposing the GigaCheck. Our research explores two approaches: (i) distinguishing human-written texts from LLM-generated ones, and (ii) detecting LLM-generated intervals in Human-Machine collaborative texts. For the first task, our approach utilizes a general-purpose LLM, leveraging its extensive language abilities to fine-tune efficiently for the downstream task of LLM-generated text detection, achieving high performance even with limited data. For the second task, we propose a novel approach that combines computer vision and natural language processing techniques. Specifically, we use a fine-tuned general-purpose LLM in conjunction with a DETR-like detection model, adapted from computer vision, to localize artificially generated intervals within text. We evaluate the GigaCheck on five classification datasets with English texts and three datasets designed for Human-Machine collaborative text analysis. Our results demonstrate that GigaCheck outperforms previous methods, even in out-of-distribution settings, establishing a strong baseline across all datasets.
Authors: Ming Li, Yanhong Li, Tianyi Zhou
Abstract: What makes a difference in the post-training of LLMs? We investigate the training patterns of different layers in large language models (LLMs), through the lens of gradient, when training with different responses and initial models. We are specifically interested in how fast vs. slow thinking affects the layer-wise gradients, given the recent popularity of training LLMs on reasoning paths such as chain-of-thoughts (CoT) and process rewards. In our study, fast thinking without CoT leads to larger gradients and larger differences of gradients across layers than slow thinking (Detailed CoT), indicating the learning stability brought by the latter. Moreover, pre-trained LLMs are less affected by the instability of fast thinking than instruction-tuned LLMs. Additionally, we study whether the gradient patterns can reflect the correctness of responses when training different LLMs using slow vs. fast thinking paths. The results show that the gradients of slow thinking can distinguish correct and irrelevant reasoning paths. As a comparison, we conduct similar gradient analyses on non-reasoning knowledge learning tasks, on which, however, trivially increasing the response length does not lead to similar behaviors of slow thinking. Our study strengthens fundamental understandings of LLM training and sheds novel insights on its efficiency and stability, which pave the way towards building a generalizable System-2 agent. Our code, data, and gradient statistics can be found in: https://github.com/MingLiiii/Layer_Gradient.
Authors: Junchao Wu, Runzhe Zhan, Derek F. Wong, Shu Yang, Xinyi Yang, Yulin Yuan, Lidia S. Chao
Abstract: Detecting text generated by large language models (LLMs) is of great recent interest. With zero-shot methods like DetectGPT, detection capabilities have reached impressive levels. However, the reliability of existing detectors in real-world applications remains underexplored. In this study, we present a new benchmark, DetectRL, highlighting that even state-of-the-art (SOTA) detection techniques still underperformed in this task. We collected human-written datasets from domains where LLMs are particularly prone to misuse. Using popular LLMs, we generated data that better aligns with real-world applications. Unlike previous studies, we employed heuristic rules to create adversarial LLM-generated text, simulating advanced prompt usages, human revisions like word substitutions, and writing errors. Our development of DetectRL reveals the strengths and limitations of current SOTA detectors. More importantly, we analyzed the potential impact of writing styles, model types, attack methods, the text lengths, and real-world human writing factors on different types of detectors. We believe DetectRL could serve as an effective benchmark for assessing detectors in real-world scenarios, evolving with advanced attack methods, thus providing more stressful evaluation to drive the development of more efficient detectors. Data and code are publicly available at: https://github.com/NLP2CT/DetectRL.
Authors: Yunqi Zhu, Wen Tang, Ying Sun, Xuebing Yang
Abstract: Recent research on large language models (LLMs) has primarily focused on their adaptation and application in specialized domains. The application of LLMs in the medical field is mainly concentrated on tasks such as the automation of medical report generation, summarization, diagnostic reasoning, and question-and-answer interactions between doctors and patients. The challenge of becoming a good teacher is more formidable than that of becoming a good student, and this study pioneers the application of LLMs in the field of medical education. In this work, we investigate the extent to which LLMs can generate medical qualification exam questions and corresponding answers based on few-shot prompts. Utilizing a real-world Chinese dataset of elderly chronic diseases, we tasked the LLMs with generating open-ended questions and answers based on a subset of sampled admission reports across eight widely used LLMs, including ERNIE 4, ChatGLM 4, Doubao, Hunyuan, Spark 4, Qwen, Llama 3, and Mistral. Furthermore, we engaged medical experts to manually evaluate these open-ended questions and answers across multiple dimensions. The study found that LLMs, after using few-shot prompts, can effectively mimic real-world medical qualification exam questions, whereas there is room for improvement in the correctness, evidence-based statements, and professionalism of the generated answers. Moreover, LLMs also demonstrate a decent level of ability to correct and rectify reference answers. Given the immense potential of artificial intelligence in the medical field, the task of generating questions and answers for medical qualification exams aimed at medical students, interns and residents can be a significant focus of future research.
Authors: Lizhe Fang, Yifei Wang, Zhaoyang Liu, Chenheng Zhang, Stefanie Jegelka, Jinyang Gao, Bolin Ding, Yisen Wang
Abstract: Handling long-context inputs is crucial for large language models (LLMs) in tasks such as extended conversations, document summarization, and many-shot in-context learning. While recent approaches have extended the context windows of LLMs and employed perplexity (PPL) as a standard evaluation metric, PPL has proven unreliable for assessing long-context capabilities. The underlying cause of this limitation has remained unclear. In this work, we provide a comprehensive explanation for this issue. We find that PPL overlooks key tokens, which are essential for long-context understanding, by averaging across all tokens and thereby obscuring the true performance of models in long-context scenarios. To address this, we propose \textbf{LongPPL}, a novel metric that focuses on key tokens by employing a long-short context contrastive method to identify them. Our experiments demonstrate that LongPPL strongly correlates with performance on various long-context benchmarks (e.g., Pearson correlation of -0.96), significantly outperforming traditional PPL in predictive accuracy. Additionally, we introduce \textbf{LongCE} (Long-context Cross-Entropy) loss, a re-weighting strategy for fine-tuning that prioritizes key tokens, leading to consistent improvements across diverse benchmarks. In summary, these contributions offer deeper insights into the limitations of PPL and present effective solutions for accurately evaluating and enhancing the long-context capabilities of LLMs. Code is available at https://github.com/PKU-ML/LongPPL.
Authors: Amir Hossein Kargaran, Fran\c{c}ois Yvon, Hinrich Sch\"utze
Abstract: The need for large text corpora has increased with the advent of pretrained language models and, in particular, the discovery of scaling laws for these models. Most available corpora have sufficient data only for languages with large dominant communities. However, there is no corpus available that (i) covers a wide range of minority languages; (ii) is generated by an open-source reproducible pipeline; and (iii) is rigorously cleaned from noise, making it trustworthy to use. We present GlotCC, a clean, document-level, 2TB general domain corpus derived from CommonCrawl, covering more than 1000 languages. We make GlotCC and the system used to generate it - including the pipeline, language identification model, and filters - available to the research community. Corpus v. 1.0 https://huggingface.co/datasets/cis-lmu/GlotCC-v1, Pipeline v. 3.0 https://github.com/cisnlp/GlotCC.
URLs: https://huggingface.co/datasets/cis-lmu/GlotCC-v1,, https://github.com/cisnlp/GlotCC.
Authors: Xiusheng Huang, Yequan Wang, Jun Zhao, Kang Liu
Abstract: Knowledge editing technology is crucial for maintaining the accuracy and timeliness of large language models (LLMs) . However, the setting of this task overlooks a significant portion of commonsense knowledge based on free-text in the real world, characterized by broad knowledge scope, long content and non instantiation. The editing objects of previous methods (e.g., MEMIT) were single token or entity, which were not suitable for commonsense knowledge in free-text form. To address the aforementioned challenges, we conducted experiments from two perspectives: knowledge localization and knowledge editing. Firstly, we introduced Knowledge Localization for Free-Text(KLFT) method, revealing the challenges associated with the distribution of commonsense knowledge in MLP and Attention layers, as well as in decentralized distribution. Next, we propose a Dynamics-aware Editing Method(DEM), which utilizes a Dynamics-aware Module to locate the parameter positions corresponding to commonsense knowledge, and uses Knowledge Editing Module to update knowledge. The DEM method fully explores the potential of the MLP and Attention layers, and successfully edits commonsense knowledge based on free-text. The experimental results indicate that the DEM can achieve excellent editing performance.
Authors: Michael Schlichtkrull, Yulong Chen, Chenxi Whitehouse, Zhenyun Deng, Mubashara Akhtar, Rami Aly, Zhijiang Guo, Christos Christodoulopoulos, Oana Cocarascu, Arpit Mittal, James Thorne, Andreas Vlachos
Abstract: The Automated Verification of Textual Claims (AVeriTeC) shared task asks participants to retrieve evidence and predict veracity for real-world claims checked by fact-checkers. Evidence can be found either via a search engine, or via a knowledge store provided by the organisers. Submissions are evaluated using AVeriTeC score, which considers a claim to be accurately verified if and only if both the verdict is correct and retrieved evidence is considered to meet a certain quality threshold. The shared task received 21 submissions, 18 of which surpassed our baseline. The winning team was TUDA_MAI with an AVeriTeC score of 63%. In this paper we describe the shared task, present the full results, and highlight key takeaways from the shared task.
Authors: Zhanke Zhou, Rong Tao, Jianing Zhu, Yiwen Luo, Zengmao Wang, Bo Han
Abstract: This paper investigates an under-explored challenge in large language models (LLMs): chain-of-thought prompting with noisy rationales, which include irrelevant or inaccurate reasoning thoughts within examples used for in-context learning. We construct NoRa dataset that is tailored to evaluate the robustness of reasoning in the presence of noisy rationales. Our findings on NoRa dataset reveal a prevalent vulnerability to such noise among current LLMs, with existing robust methods like self-correction and self-consistency showing limited efficacy. Notably, compared to prompting with clean rationales, base LLM drops by 1.4%-19.8% in accuracy with irrelevant thoughts and more drastically by 2.2%-40.4% with inaccurate thoughts. Addressing this challenge necessitates external supervision that should be accessible in practice. Here, we propose the method of contrastive denoising with noisy chain-of-thought (CD-CoT). It enhances LLMs' denoising-reasoning capabilities by contrasting noisy rationales with only one clean rationale, which can be the minimal requirement for denoising-purpose prompting. This method follows a principle of exploration and exploitation: (1) rephrasing and selecting rationales in the input space to achieve explicit denoising and (2) exploring diverse reasoning paths and voting on answers in the output space. Empirically, CD-CoT demonstrates an average improvement of 17.8% in accuracy over the base model and shows significantly stronger denoising capabilities than baseline methods. The source code is publicly available at: https://github.com/tmlr-group/NoisyRationales.
Authors: Hao Yang, Lizhen Qu, Ehsan Shareghi, Gholamreza Haffari
Abstract: Large Multimodal Models (LMMs) have demonstrated the ability to interact with humans under real-world conditions by combining Large Language Models (LLMs) and modality encoders to align multimodal information (visual and auditory) with text. However, such models raise new safety challenges of whether models that are safety-aligned on text also exhibit consistent safeguards for multimodal inputs. Despite recent safety-alignment research on vision LMMs, the safety of audio LMMs remains under-explored. In this work, we comprehensively red team the safety of five advanced audio LMMs under three settings: (i) harmful questions in both audio and text formats, (ii) harmful questions in text format accompanied by distracting non-speech audio, and (iii) speech-specific jailbreaks. Our results under these settings demonstrate that open-source audio LMMs suffer an average attack success rate of 69.14% on harmful audio questions, and exhibit safety vulnerabilities when distracted with non-speech audio noise. Our speech-specific jailbreaks on Gemini-1.5-Pro achieve an attack success rate of 70.67% on the harmful query benchmark. We provide insights on what could cause these reported safety-misalignments. Warning: this paper contains offensive examples.
Authors: Rena Gao, Xuetong Wu, Siwen Luo, Caren Han, Feng Liu
Abstract: Out-of-distribution (OOD) detection in multimodal contexts is essential for identifying deviations in combined inputs from different modalities, particularly in applications like open-domain dialogue systems or real-life dialogue interactions. This paper aims to improve the user experience that involves multi-round long dialogues by efficiently detecting OOD dialogues and images. We introduce a novel scoring framework named Dialogue Image Aligning and Enhancing Framework (DIAEF) that integrates the visual language models with the novel proposed scores that detect OOD in two key scenarios (1) mismatches between the dialogue and image input pair and (2) input pairs with previously unseen labels. Our experimental results, derived from various benchmarks, demonstrate that integrating image and multi-round dialogue OOD detection is more effective with previously unseen labels than using either modality independently. In the presence of mismatched pairs, our proposed score effectively identifies these mismatches and demonstrates strong robustness in long dialogues. This approach enhances domain-aware, adaptive conversational agents and establishes baselines for future studies.
Authors: S\'eamus Lankford, Andy Way
Abstract: In an evolving landscape of crisis communication, the need for robust and adaptable Machine Translation (MT) systems is more pressing than ever, particularly for low-resource languages. This study presents a comprehensive exploration of leveraging Large Language Models (LLMs) and Multilingual LLMs (MLLMs) to enhance MT capabilities in such scenarios. By focusing on the unique challenges posed by crisis situations where speed, accuracy, and the ability to handle a wide range of languages are paramount, this research outlines a novel approach that combines the cutting-edge capabilities of LLMs with fine-tuning techniques and community-driven corpus development strategies. At the core of this study is the development and empirical evaluation of MT systems tailored for two low-resource language pairs, illustrating the process from initial model selection and fine-tuning through to deployment. Bespoke systems are developed and modelled on the recent Covid-19 pandemic. The research highlights the importance of community involvement in creating highly specialised, crisis-specific datasets and compares custom GPTs with NLLB-adapted MLLM models. It identifies fine-tuned MLLM models as offering superior performance compared with their LLM counterparts. A scalable and replicable model for rapid MT system development in crisis scenarios is outlined. Our approach enhances the field of humanitarian technology by offering a blueprint for developing multilingual communication systems during emergencies.
Authors: Matyas Juhasz, Kalyan Dutia, Henry Franks, Conor Delahunty, Patrick Fawbert Mills, Harrison Pim
Abstract: Climate decision making is constrained by the complexity and inaccessibility of key information within lengthy, technical, and multi-lingual documents. Generative AI technologies offer a promising route for improving the accessibility of information contained within these documents, but suffer from limitations. These include (1) a tendency to hallucinate or mis-represent information, (2) difficulty in steering or guaranteeing properties of generated output, and (3) reduced performance in specific technical domains. To address these challenges, we introduce a novel evaluation framework with domain-specific dimensions tailored for climate-related documents. We then apply this framework to evaluate Retrieval-Augmented Generation (RAG) approaches and assess retrieval- and generation-quality within a prototype tool that answers questions about individual climate law and policy documents. In addition, we publish a human-annotated dataset and scalable automated evaluation tools, with the aim of facilitating broader adoption and robust assessment of these systems in the climate domain. Our findings highlight the key components of responsible deployment of RAG to enhance decision-making, while also providing insights into user experience (UX) considerations for safely deploying such systems to build trust with users in high-risk domains.
Authors: Xinghao Wang, Pengyu Wang, Bo Wang, Dong Zhang, Yunhua Zhou, Xipeng Qiu
Abstract: Large language models (LLMs) have revolutionized numerous applications, yet their deployment remains challenged by memory constraints on local devices. While scaling laws have enhanced LLM capabilities, the primary bottleneck has shifted from \textit{capability} to \textit{availability}, emphasizing the need for efficient memory management. Traditional compression methods, such as quantization, often require predefined compression ratios and separate compression processes for each setting, complicating deployment in variable memory environments. In this paper, we introduce \textbf{BitStack}, a novel, training-free weight compression approach that enables megabyte-level trade-offs between memory usage and model performance. By leveraging weight decomposition, BitStack can dynamically adjust the model size with minimal transmission between running memory and storage devices. Our approach iteratively decomposes weight matrices while considering the significance of each parameter, resulting in an approximately 1-bit per parameter residual block in each decomposition iteration. These blocks are sorted and stacked in storage as basic transmission units, with different quantities loaded based on current memory availability. Extensive experiments across a wide range of tasks demonstrate that, despite offering fine-grained size control, BitStack consistently matches or surpasses strong quantization baselines, particularly at extreme compression ratios. To the best of our knowledge, this is the first decomposition-based method that effectively bridges the gap to practical compression techniques like quantization. Code is available at https://github.com/xinghaow99/BitStack.
Authors: Shanghaoran Quan, Tianyi Tang, Bowen Yu, An Yang, Dayiheng Liu, Bofei Gao, Jianhong Tu, Yichang Zhang, Jingren Zhou, Junyang Lin
Abstract: Recent advancements in Large Language Models (LLMs) have significantly enhanced their ability to process long contexts, yet a notable gap remains in generating long, aligned outputs. This limitation stems from a training gap where pre-training lacks effective instructions for long-text generation, and post-training data primarily consists of short query-response pairs. Current approaches, such as instruction backtranslation and behavior imitation, face challenges including data quality, copyright issues, and constraints on proprietary model usage. In this paper, we introduce an innovative iterative training framework called Self-Lengthen that leverages only the intrinsic knowledge and skills of LLMs without the need for auxiliary data or proprietary models. The framework consists of two roles: the Generator and the Extender. The Generator produces the initial response, which is then split and expanded by the Extender. This process results in a new, longer response, which is used to train both the Generator and the Extender iteratively. Through this process, the models are progressively trained to handle increasingly longer responses. Experiments on benchmarks and human evaluations show that Self-Lengthen outperforms existing methods in long-text generation, when applied to top open-source LLMs such as Qwen2 and LLaMA3. Our code is publicly available at https://github.com/QwenLM/Self-Lengthen.
Authors: Jiayi Wang, Yao Lu, Maurice Weber, Max Ryabinin, Yihong Chen, Raphael Tang, Pontus Stenetorp
Abstract: English, as a very high-resource language, enables the pretraining of high-quality large language models (LLMs). The same cannot be said for most other languages, as leading LLMs still underperform for non-English languages, likely due to a gap in the quality and diversity of the available multilingual pretraining corpora. In this work, we find that machine-translated text from a single high-quality source language can contribute significantly to the pretraining of multilingual LLMs. We translate FineWeb-Edu, a high-quality English web dataset, into French, German, and Spanish, resulting in a final 300B-token dataset, which we call TransWeb-Edu, and pretrain a 1.3B-parameter model, CuatroLLM, from scratch on this dataset. Across five non-English reasoning tasks, we show that CuatroLLM matches or outperforms state-of-the-art multilingual models trained using closed data, such as Llama3.2 and Gemma2, despite using an order of magnitude less data, such as about 6% of the tokens used for Llama3.2's training. We further demonstrate that with additional domain-specific pretraining, amounting to less than 1% of TransWeb-Edu, CuatroLLM surpasses the state of the art in multilingual reasoning. To promote reproducibility, we release our corpus, models, and training pipeline under open licenses at hf.co/britllm/CuatroLLM.
Authors: Ioannis Tsiamas, Matthias Sperber, Andrew Finch, Sarthak Garg
Abstract: The prosody of a spoken utterance, including features like stress, intonation and rhythm, can significantly affect the underlying semantics, and as a consequence can also affect its textual translation. Nevertheless, prosody is rarely studied within the context of speech-to-text translation (S2TT) systems. In particular, end-to-end (E2E) systems have been proposed as well-suited for prosody-aware translation because they have direct access to the speech signal when making translation decisions, but the understanding of whether this is successful in practice is still limited. A main challenge is the difficulty of evaluating prosody awareness in translation. To address this challenge, we introduce an evaluation methodology and a focused benchmark (named ContraProST) aimed at capturing a wide range of prosodic phenomena. Our methodology uses large language models and controllable text-to-speech (TTS) to generate contrastive examples. Through experiments in translating English speech into German, Spanish, and Japanese, we find that (a) S2TT models possess some internal representation of prosody, but the prosody signal is often not strong enough to affect the translations, (b) E2E systems outperform cascades of speech recognition and text translation systems, confirming their theoretical advantage in this regard, and (c) certain cascaded systems also capture prosodic information in the translation, but only to a lesser extent that depends on the particulars of the transcript's surface form.
Authors: Lucian Li, Eryclis Silva
Abstract: Introduction: Tracing the spread of ideas and the presence of influence is a question of special importance across a wide range of disciplines, ranging from intellectual history to cultural analytics, computational social science, and the science of science. Method: We collect a corpus of open source journal articles, generate Knowledge Graph representations using the Gemini LLM, and attempt to predict the existence of citations between sampled pairs of articles using previously published methods and a novel Graph Neural Network based embedding model. Results: We demonstrate that our knowledge graph embedding method is superior at distinguishing pairs of articles with and without citation. Once trained, it runs efficiently and can be fine-tuned on specific corpora to suit individual researcher needs. Conclusion(s): This experiment demonstrates that the relationships encoded in a knowledge graph, especially the types of concepts brought together by specific relations can encode information capable of revealing intellectual influence. This suggests that further work in analyzing document level knowledge graphs to understand latent structures could provide valuable insights.
Authors: Zhaohui Li, Rebecca J. Passonneau
Abstract: Classifier models are prevalent in natural language processing (NLP), often with high accuracy. Yet in real world settings, human-in-the-loop systems can foster trust in model outputs and even higher performance. Selective Prediction (SP) methods determine when to adopt a classifier's output versus defer to a human. Previous SP approaches have addressed how to improve softmax as a measure of model confidence, or have developed separate confidence estimators. One previous method involves learning a deferral model based on engineered features. We introduce a novel joint-training approach that simultaneously optimizes learned representations used by the classifier module and a learned deferral policy. Our results on four classification tasks demonstrate that joint training not only leads to better SP outcomes over two strong baselines, but also improves the performance of both modules.
Authors: Muhammed Saeed, Elgizouli Mohamed, Mukhtar Mohamed, Shaina Raza, Shady Shehata, Muhammad Abdul-Mageed
Abstract: Large language models (LLMs) are widely used but raise ethical concerns due to embedded social biases. This study examines LLM biases against Arabs versus Westerners across eight domains, including women's rights, terrorism, and anti-Semitism and assesses model resistance to perpetuating these biases. To this end, we create two datasets: one to evaluate LLM bias toward Arabs versus Westerners and another to test model safety against prompts that exaggerate negative traits ("jailbreaks"). We evaluate six LLMs -- GPT-4, GPT-4o, LlaMA 3.1 (8B & 405B), Mistral 7B, and Claude 3.5 Sonnet. We find 79% of cases displaying negative biases toward Arabs, with LlaMA 3.1-405B being the most biased. Our jailbreak tests reveal GPT-4o as the most vulnerable, despite being an optimized version, followed by LlaMA 3.1-8B and Mistral 7B. All LLMs except Claude exhibit attack success rates above 87% in three categories. We also find Claude 3.5 Sonnet the safest, but it still displays biases in seven of eight categories. Despite being an optimized version of GPT4, We find GPT-4o to be more prone to biases and jailbreaks, suggesting optimization flaws. Our findings underscore the pressing need for more robust bias mitigation strategies and strengthened security measures in LLMs.
Authors: Dominic Sobhani, Amir Feder, David Blei
Abstract: Probabilistic topic models are a powerful tool for extracting latent themes from large text datasets. In many text datasets, we also observe per-document covariates (e.g., source, style, political affiliation) that act as environments that modulate a "global" (environment-agnostic) topic representation. Accurately learning these representations is important for prediction on new documents in unseen environments and for estimating the causal effect of topics on real-world outcomes. To this end, we introduce the Multi-environment Topic Model (MTM), an unsupervised probabilistic model that separates global and environment-specific terms. Through experimentation on various political content, from ads to tweets and speeches, we show that the MTM produces interpretable global topics with distinct environment-specific words. On multi-environment data, the MTM outperforms strong baselines in and out-of-distribution. It also enables the discovery of accurate causal effects.
Authors: Kyle Gorman, Yuval Pinter
Abstract: The common practice of preprocessing text before feeding it into NLP models introduces many decision points which have unintended consequences on model performance. In this opinion piece, we focus on the handling of diacritics in texts originating in many languages and scripts. We demonstrate, through several case studies, the adverse effects of inconsistent encoding of diacritized characters and of removing diacritics altogether. We call on the community to adopt simple but necessary steps across all models and toolkits in order to improve handling of diacritized text and, by extension, increase equity in multilingual NLP.
Authors: Jinghan Zhang, Fengran Mo, Xiting Wang, Kunpeng Liu
Abstract: Recent advances in large language models (LLMs) have demonstrated their potential in handling complex reasoning tasks, which are usually achieved by constructing a thought chain to guide the model to solve the problem with multi-step thinking. However, existing methods often remain confined to previously explored solution spaces and thus overlook the critical blind spot within LLMs' cognitive range. To address these issues, we design the Thought Space Explorer (TSE), a novel framework to expand and optimize thought structures to guide LLMs to explore their blind spots of thinking. By generating new reasoning steps and branches based on the original thought structure with various designed strategies, TSE broadens the thought space and alleviates the impact of blind spots for LLM reasoning. Experimental results on multiple levels of reasoning tasks demonstrate the efficacy of TSE. We also conduct extensive analysis to understand how structured and expansive thought can contribute to unleashing the potential of LLM reasoning capabilities.
Authors: Lucas Georges Gabriel Charpentier, David Samuel
Abstract: We present a simple way to merge masked language modeling with causal language modeling. This hybrid training objective results in a model that combines the strengths of both modeling paradigms within a single transformer stack: GPT-BERT can be transparently used like any standard causal or masked language model. We test the pretraining process that enables this flexible behavior on the BabyLM Challenge 2024. The results show that the hybrid pretraining outperforms masked-only or causal-only models. We openly release the models, training corpora and code.
Authors: Yunjia Qi, Hao Peng, Xiaozhi Wang, Bin Xu, Lei Hou, Juanzi Li
Abstract: Large language models (LLMs) struggle to follow instructions with complex constraints in format, length, etc. Following the conventional instruction-tuning practice, previous works conduct post-training on complex instruction-response pairs generated by feeding complex instructions to advanced LLMs. However, even advanced LLMs cannot follow complex instructions well, thus limiting the quality of generated data. In this work, we find that existing datasets inherently contain implicit complex constraints and propose a novel data generation technique, constraint back-translation. Specifically, we take the high-quality instruction-response pairs in existing datasets and only adopt advanced LLMs to add complex constraints already met by the responses to the instructions, which naturally reduces costs and data noise. In the experiments, we adopt Llama3-70B-Instruct to back-translate constraints and create a high-quality complex instruction-response dataset, named CRAB. We present that post-training on CRAB improves multiple backbone LLMs' complex instruction-following ability, evaluated on extensive instruction-following benchmarks. We further find that constraint back-translation also serves as a useful auxiliary training objective in post-training. Our code, data, and models will be released to facilitate future research.
Authors: Yujin Potter, Shiyang Lai, Junsol Kim, James Evans, Dawn Song
Abstract: How could LLMs influence our democracy? We investigate LLMs' political leanings and the potential influence of LLMs on voters by conducting multiple experiments in a U.S. presidential election context. Through a voting simulation, we first demonstrate 18 open- and closed-weight LLMs' political preference for a Democratic nominee over a Republican nominee. We show how this leaning towards the Democratic nominee becomes more pronounced in instruction-tuned models compared to their base versions by analyzing their responses to candidate-policy related questions. We further explore the potential impact of LLMs on voter choice by conducting an experiment with 935 U.S. registered voters. During the experiments, participants interacted with LLMs (Claude-3, Llama-3, and GPT-4) over five exchanges. The experiment results show a shift in voter choices towards the Democratic nominee following LLM interaction, widening the voting margin from 0.7% to 4.6%, even though LLMs were not asked to persuade users to support the Democratic nominee during the discourse. This effect is larger than many previous studies on the persuasiveness of political campaigns, which have shown minimal effects in presidential elections. Many users also expressed a desire for further political interaction with LLMs. Which aspects of LLM interactions drove these shifts in voter choice requires further study. Lastly, we explore how a safety method can make LLMs more politically neutral, while leaving some open questions.
Authors: Yuxiang Wei, Federico Cassano, Jiawei Liu, Yifeng Ding, Naman Jain, Zachary Mueller, Harm de Vries, Leandro von Werra, Arjun Guha, Lingming Zhang
Abstract: Instruction tuning is a supervised fine-tuning approach that significantly improves the ability of large language models (LLMs) to follow human instructions. We propose SelfCodeAlign, the first fully transparent and permissive pipeline for self-aligning code LLMs without extensive human annotations or distillation. SelfCodeAlign employs the same base model for inference throughout the data generation process. It first extracts diverse coding concepts from high-quality seed snippets to generate new tasks. It then samples multiple responses per task, pairs each with test cases, and validates them in a sandbox environment. Finally, passing examples are selected for instruction tuning. In our primary experiments, we use SelfCodeAlign with CodeQwen1.5-7B to generate a dataset of 74k instruction-response pairs. Finetuning on this dataset leads to a model that achieves a 67.1 pass@1 on HumanEval+, surpassing CodeLlama-70B-Instruct despite being ten times smaller. Across all benchmarks, this finetuned model consistently outperforms the original version trained with OctoPack, the previous state-of-the-art method for instruction tuning without human annotations or distillation. Additionally, we show that SelfCodeAlign is effective across LLMs of various sizes, from 3B to 33B, and that the base models can benefit more from alignment with their own data distribution. We further validate each component's effectiveness in our pipeline, showing that SelfCodeAlign outperforms both direct distillation from GPT-4o and leading GPT-3.5-based distillation methods, such as OSS-Instruct and Evol-Instruct. SelfCodeAlign has also led to the creation of StarCoder2-Instruct, the first fully transparent, permissively licensed, and self-aligned code LLM that achieves state-of-the-art coding performance.
Authors: Mohamed Elgaar, Hadi Amiri
Abstract: We present a novel approach to paraphrase generation that enables precise control and fine-tuning of 40 linguistic attributes for English. Our model is an encoder-decoder architecture that takes as input a source sentence and desired linguistic attributes, and produces paraphrases of the source that satisfy the desired attributes. To guarantee high-quality outputs at inference time, our method is equipped with a quality control mechanism that gradually adjusts the embedding of linguistic attributes to find the nearest and most attainable configuration of desired attributes for paraphrase generation. We evaluate the effectiveness of our method by comparing it to recent controllable generation models. Experimental results demonstrate that the proposed model outperforms baselines in generating paraphrases that satisfy desired linguistic attributes.
Authors: Yuqi Zhou, Sunhao Dai, Zhanshuo Cao, Xiao Zhang, Jun Xu
Abstract: Text embeddings enable various applications, but their performance deteriorates on longer texts. In this paper, we find that the performance degradation is due to a phenomenon called Length Collapse, where longer text embeddings collapse into a narrow space. This collapse results in a distributional inconsistency between embeddings of different text lengths, ultimately hurting the performance of downstream tasks. Theoretically, by considering the self-attention mechanism inherently functions as a low-pass filter, we prove that long sequences increase the attenuation rate of the low-pass filter effect of the self-attention mechanism. With layers going deeper, excessive low-pass filtering causes the token signals to retain only their Direct-Current (DC) component, which means the input token feature maps will collapse into a narrow space, especially in long texts. Based on the above analysis, we propose to mitigate the undesirable length collapse limitation by introducing a temperature in softmax(), which achieves a higher low-filter attenuation rate. The tuning-free method, called TempScale, can be plugged into multiple transformer-based embedding models. Empirically, we demonstrate that TempScale can improve existing embedding models, especially on long text inputs, bringing up to 0.53% performance gains on 40 datasets from Massive Text Embedding Benchmark (MTEB) and 0.82% performance gains on 4 datasets from LongEmbed, which specifically focuses on long context retrieval.
Authors: Mohamed Elgaar, Hadi Amiri
Abstract: We introduce LingGen, a novel approach for controlled text generation that offers precise control over a wide array of linguistic attributes, even as the number of attributes varies. LingGen employs a dynamic P-MASKING strategy, which samples masking rates from a power law distribution during training. This innovative approach enables the model to develop robust representations and adapt its attribute control capabilities across a variable number of attributes, from a single attribute to multiple complex configurations. The P-MASKING technique enhances LingGen's ability to manage different levels of attribute visibility, resulting in superior performance in multi-attribute generation tasks. Our experiments demonstrate that LingGen surpasses current state-of-the-art models in both attribute control accuracy and text fluency, particularly excelling in scenarios with varying attribute demands. Additionally, our ablation studies highlight the effectiveness of P-MASKING and the influence of different base language models on performance. These findings demonstrate LingGen's potential for applications requiring precise and adaptable control over multiple linguistic attributes in text generation.
Authors: Jiajun Xi, Yinong He, Jianing Yang, Yinpei Dai, Joyce Chai
Abstract: In real-world scenarios, it is desirable for embodied agents to have the ability to leverage human language to gain explicit or implicit knowledge for learning tasks. Despite recent progress, most previous approaches adopt simple low-level instructions as language inputs, which may not reflect natural human communication. It's not clear how to incorporate rich language use to facilitate task learning. To address this question, this paper studies different types of language inputs in facilitating reinforcement learning (RL) embodied agents. More specifically, we examine how different levels of language informativeness (i.e., feedback on past behaviors and future guidance) and diversity (i.e., variation of language expressions) impact agent learning and inference. Our empirical results based on four RL benchmarks demonstrate that agents trained with diverse and informative language feedback can achieve enhanced generalization and fast adaptation to new tasks. These findings highlight the pivotal role of language use in teaching embodied agents new tasks in an open world. Project website: https://github.com/sled-group/Teachable_RL
Authors: Lucie-Aim\'ee Kaffee, Arnav Arora, Isabelle Augenstein
Abstract: The moderation of content on online platforms is usually non-transparent. On Wikipedia, however, this discussion is carried out publicly and the editors are encouraged to use the content moderation policies as explanations for making moderation decisions. Currently, only a few comments explicitly mention those policies -- 20% of the English ones, but as few as 2% of the German and Turkish comments. To aid in this process of understanding how content is moderated, we construct a novel multilingual dataset of Wikipedia editor discussions along with their reasoning in three languages. The dataset contains the stances of the editors (keep, delete, merge, comment), along with the stated reason, and a content moderation policy, for each edit decision. We demonstrate that stance and corresponding reason (policy) can be predicted jointly with a high degree of accuracy, adding transparency to the decision-making process. We release both our joint prediction models and the multilingual content moderation dataset for further research on automated transparent content moderation.
Authors: Victoria Benjamin, Emily Braca, Israel Carter, Hafsa Kanchwala, Nava Khojasteh, Charly Landow, Yi Luo, Caroline Ma, Anna Magarelli, Rachel Mirin, Avery Moyer, Kayla Simpson, Amelia Skawinski, Thomas Heverin
Abstract: This study systematically analyzes the vulnerability of 36 large language models (LLMs) to various prompt injection attacks, a technique that leverages carefully crafted prompts to elicit malicious LLM behavior. Across 144 prompt injection tests, we observed a strong correlation between model parameters and vulnerability, with statistical analyses, such as logistic regression and random forest feature analysis, indicating that parameter size and architecture significantly influence susceptibility. Results revealed that 56 percent of tests led to successful prompt injections, emphasizing widespread vulnerability across various parameter sizes, with clustering analysis identifying distinct vulnerability profiles associated with specific model configurations. Additionally, our analysis uncovered correlations between certain prompt injection techniques, suggesting potential overlaps in vulnerabilities. These findings underscore the urgent need for robust, multi-layered defenses in LLMs deployed across critical infrastructure and sensitive industries. Successful prompt injection attacks could result in severe consequences, including data breaches, unauthorized access, or misinformation. Future research should explore multilingual and multi-step defenses alongside adaptive mitigation strategies to strengthen LLM security in diverse, real-world environments.
Authors: Dezhan Tu, Danylo Vashchilenko, Yuzhe Lu, Panpan Xu
Abstract: Vision-Language Models (VLMs) have demonstrated impressive performance across a versatile set of tasks. A key challenge in accelerating VLMs is storing and accessing the large Key-Value (KV) cache that encodes long visual contexts, such as images or videos. While existing KV cache compression methods are effective for Large Language Models (LLMs), directly migrating them to VLMs yields suboptimal accuracy and speedup. To bridge the gap, we propose VL-Cache, a novel KV cache compression recipe tailored for accelerating VLM inference. In this paper, we first investigate the unique sparsity pattern of VLM attention by distinguishing visual and text tokens in prefill and decoding phases. Based on these observations, we introduce a layer-adaptive sparsity-aware cache budget allocation method that effectively distributes the limited cache budget across different layers, further reducing KV cache size without compromising accuracy. Additionally, we develop a modality-aware token scoring policy to better evaluate the token importance. Empirical results on multiple benchmark datasets demonstrate that retaining only 10% of KV cache achieves accuracy comparable to that with full cache. In a speed benchmark, our method accelerates end-to-end latency of generating 100 tokens by up to 2.33x and speeds up decoding by up to 7.08x, while reducing the memory footprint of KV cache in GPU by 90%.
Authors: Peter Ochieng, Dennis Kaburu
Abstract: This paper presents a pilot study on direct speech-to-speech translation (S2ST) by leveraging linguistic similarities among selected African languages within the same phylum, particularly in cases where traditional data annotation is expensive or impractical. We propose a segment-based model that maps speech segments both within and across language phyla, effectively eliminating the need for large paired datasets. By utilizing paired segments and guided diffusion, our model enables translation between any two languages in the dataset. We evaluate the model on a proprietary dataset from the Kenya Broadcasting Corporation (KBC), which includes five languages: Swahili, Luo, Kikuyu, Nandi, and English. The model demonstrates competitive performance in segment pairing and translation quality, particularly for languages within the same phylum. Our experiments reveal that segment length significantly influences translation accuracy, with average-length segments yielding the highest pairing quality. Comparative analyses with traditional cascaded ASR-MT techniques show that the proposed model delivers nearly comparable translation performance. This study underscores the potential of exploiting linguistic similarities within language groups to perform efficient S2ST, especially in low-resource language contexts.
Authors: Arihan Yadav, Alan McMillan
Abstract: Retrieval-Augmented Generation (RAG) systems enhance text generation by incorporating external knowledge but often struggle when retrieving context across different text modalities due to semantic gaps. We introduce a generalized projection-based method, inspired by adapter modules in transfer learning, that efficiently bridges these gaps between various text types, such as programming code and pseudocode, or English and French sentences. Our approach emphasizes speed, accuracy, and data efficiency, requiring minimal resources for training and inference. By aligning embeddings from heterogeneous text modalities into a unified space through a lightweight projection network, our model significantly outperforms traditional retrieval methods like the Okapi BM25 algorithm and models like Dense Passage Retrieval (DPR), while approaching the accuracy of Sentence Transformers. Extensive evaluations demonstrate the effectiveness and generalizability of our method across different tasks, highlighting its potential for real-time, resource-constrained applications.
Authors: Yunwei Ren, Zixuan Wang, Jason D. Lee
Abstract: Transformers have excelled in natural language modeling and one reason behind this success is their exceptional ability to combine contextual informal and global knowledge. However, the theoretical basis remains unclear. In this paper, first we introduce the Sparse Contextual Bigram (SCB), a natural extension of the classical bigram model, where the next token's generation depends on a sparse set of earlier positions determined by the last token. We then analyze the training dynamics and sample complexity of learning SCB using a one-layer linear transformer with a gradient-based algorithm. We show that when trained from scratch, the training process can be split into an initial sample-intensive stage where the correlation is boosted from zero to a nontrivial value, followed by a more sample-efficient stage of further improvement. Additionally, we prove that, provided a nontrivial correlation between the downstream and pretraining tasks, finetuning from a pretrained model allows us to bypass the initial sample-intensive stage. We also empirically demonstrate that our algorithm can outperform SGD in this setting and discuss its relationship with the usual softmax-based transformers.
Authors: Emanuele Marconato, S\'ebastien Lachapelle, Sebastian Weichwald, Luigi Gresele
Abstract: We analyze identifiability as a possible explanation for the ubiquity of linear properties across language models, such as the vector difference between the representations of "easy" and "easiest" being parallel to that between "lucky" and "luckiest". For this, we ask whether finding a linear property in one model implies that any model that induces the same distribution has that property, too. To answer that, we first prove an identifiability result to characterize distribution-equivalent next-token predictors, lifting a diversity requirement of previous results. Second, based on a refinement of relational linearity [Paccanaro and Hinton, 2001; Hernandez et al., 2024], we show how many notions of linearity are amenable to our analysis. Finally, we show that under suitable conditions, these linear properties either hold in all or none distribution-equivalent next-token predictors.
Authors: Edward S. Hu, Kwangjun Ahn, Qinghua Liu, Haoran Xu, Manan Tomar, Ada Langford, Dinesh Jayaraman, Alex Lamb, John Langford
Abstract: We introduce the "Belief State Transformer", a next-token predictor that takes both a prefix and suffix as inputs, with a novel objective of predicting both the next token for the prefix and the previous token for the suffix. The Belief State Transformer effectively learns to solve challenging problems that conventional forward-only transformers struggle with, in a domain-independent fashion. Key to this success is learning a compact belief state that captures all relevant information necessary for accurate predictions. Empirical ablations show that each component of the model is essential in difficult scenarios where standard Transformers fall short. For the task of story writing with known prefixes and suffixes, our approach outperforms the Fill-in-the-Middle method for reaching known goals and demonstrates improved performance even when the goals are unknown. Altogether, the Belief State Transformer enables more efficient goal-conditioned decoding, better test-time inference, and high-quality text representations on small scale problems.
Authors: Peter Belcak, Roger Wattenhofer
Abstract: It is staggering that words of the English language, which are on average represented by 5--6 bytes of ASCII, require as much as 24 kilobytes when served to large language models. We show that there is room for more information in every token embedding. We demonstrate that 1--3-layer transformers are capable of encoding and subsequently decoding standard English sentences into as little as a single 3-kilobyte token. Our work implies that even small networks can learn to construct valid English sentences and suggests the possibility of optimising large language models by moving from sub-word token embeddings towards larger fragments of text.
Authors: Andy Lo, Albert Q. Jiang, Wenda Li, Mateja Jamnik
Abstract: Ontologies are useful for automatic machine processing of domain knowledge as they represent it in a structured format. Yet, constructing ontologies requires substantial manual effort. To automate part of this process, large language models (LLMs) have been applied to solve various subtasks of ontology learning. However, this partial ontology learning does not capture the interactions between subtasks. We address this gap by introducing OLLM, a general and scalable method for building the taxonomic backbone of an ontology from scratch. Rather than focusing on subtasks, like individual relations between entities, we model entire subcomponents of the target ontology by finetuning an LLM with a custom regulariser that reduces overfitting on high-frequency concepts. We introduce a novel suite of metrics for evaluating the quality of the generated ontology by measuring its semantic and structural similarity to the ground truth. In contrast to standard metrics, our metrics use deep learning techniques to define more robust distance measures between graphs. Both our quantitative and qualitative results on Wikipedia show that OLLM outperforms subtask composition methods, producing more semantically accurate ontologies while maintaining structural integrity. We further demonstrate that our model can be effectively adapted to new domains, like arXiv, needing only a small number of training examples. Our source code and datasets are available at https://github.com/andylolu2/ollm.
Authors: Colin Conwell, Christopher Hamblin, Chelsea Boccagno, David Mayo, Jesse Cummings, Leyla Isik, Andrei Barbu
Abstract: When we experience a visual stimulus as beautiful, how much of that experience derives from perceptual computations we cannot describe versus conceptual knowledge we can readily translate into natural language? Disentangling perception from language in visually-evoked affective and aesthetic experiences through behavioral paradigms or neuroimaging is often empirically intractable. Here, we circumnavigate this challenge by using linear decoding over the learned representations of unimodal vision, unimodal language, and multimodal (language-aligned) deep neural network (DNN) models to predict human beauty ratings of naturalistic images. We show that unimodal vision models (e.g. SimCLR) account for the vast majority of explainable variance in these ratings. Language-aligned vision models (e.g. SLIP) yield small gains relative to unimodal vision. Unimodal language models (e.g. GPT2) conditioned on visual embeddings to generate captions (via CLIPCap) yield no further gains. Caption embeddings alone yield less accurate predictions than image and caption embeddings combined (concatenated). Taken together, these results suggest that whatever words we may eventually find to describe our experience of beauty, the ineffable computations of feedforward perception may provide sufficient foundation for that experience.
Authors: Junda Wu, Xintong Li, Ruoyu Wang, Yu Xia, Yuxin Xiong, Jianing Wang, Tong Yu, Xiang Chen, Branislav Kveton, Lina Yao, Jingbo Shang, Julian McAuley
Abstract: Offline evaluation of LLMs is crucial in understanding their capacities, though current methods remain underexplored in existing research. In this work, we focus on the offline evaluation of the chain-of-thought capabilities and show how to optimize LLMs based on the proposed evaluation method. To enable offline feedback with rich knowledge and reasoning paths, we use knowledge graphs (e.g., Wikidata5m) to provide feedback on the generated chain of thoughts. Due to the heterogeneity between LLM reasoning and KG structures, direct interaction and feedback from KGs on LLM behavior are challenging, as they require accurate entity linking and grounding of LLM-generated chains of thought in the KG. To address the above challenge, we propose an offline chain-of-thought evaluation framework, OCEAN, which models chain-of-thought reasoning in LLMs as an MDP and evaluate the policy's alignment with KG preference modeling. To overcome the reasoning heterogeneity and grounding problems, we leverage on-policy KG exploration and RL to model a KG policy that generates token-level likelihood distributions for LLM-generated chain-of-thought reasoning paths, simulating KG reasoning preference. Then we incorporate the knowledge-graph feedback on the validity and alignment of the generated reasoning paths into inverse propensity scores and propose KG-IPS estimator. Theoretically, we prove the unbiasedness of the proposed KG-IPS estimator and provide a lower bound on its variance. With the off-policy evaluated value function, we can directly enable off-policy optimization to further enhance chain-of-thought alignment. Our empirical study shows that OCEAN can be efficiently optimized for generating chain-of-thought reasoning paths with higher estimated values without affecting LLMs' general abilities in downstream tasks or their internal knowledge.
Authors: Taridzo Chomutare, Therese Olsen Svenning, Miguel \'Angel Tejedor Hern\'andez, Phuong Dinh Ngo, Andrius Budrionis, Kaisa Markljung, Lill Irene Hind, Torbj{\o}rn Torsvik, Karl {\O}yvind Mikalsen, Aleksandar Babic, Hercules Dalianis
Abstract: \textbf{Trial design} Crossover randomized controlled trial. \textbf{Methods} An AI tool, Easy-ICD, was developed to assist clinical coders and was tested for improving both accuracy and time in a user study in Norway and Sweden. Participants were randomly assigned to two groups, and crossed over between coding complex (longer) texts versus simple (shorter) texts, while using our tool versus not using our tool. \textbf{Results} Based on Mann-Whitney U test, the median coding time difference for complex clinical text sequences was 123 seconds (\emph{P}\textless.001, 95\% CI: 81 to 164), representing a 46\% reduction in median coding time when our tool is used. There was no significant time difference for simpler text sequences. For coding accuracy, the improvement we noted for both complex and simple texts was not significant. \textbf{Conclusions} This study demonstrates the potential of AI to transform common tasks in clinical workflows, with ostensible positive impacts on work efficiencies for complex clinical coding tasks. Further studies within hospital workflows are required before these presumed impacts can be more clearly understood.
Authors: Xiusheng Huang, Jiaxiang Liu, Yequan Wang, Kang Liu
Abstract: Knowledge editing technology has received widespread attention for low-cost updates of incorrect or outdated knowledge in large-scale language models. However, recent research has found that edited models often exhibit varying degrees of performance degradation. The reasons behind this phenomenon and potential solutions have not yet been provided. In order to investigate the reasons for the performance decline of the edited model and optimize the editing method, this work explores the underlying reasons from both data and model perspectives. Specifically, 1) from a data perspective, to clarify the impact of data on the performance of editing models, this paper first constructs a Multi-Question Dataset (MQD) to evaluate the impact of different types of editing data on model performance. The performance of the editing model is mainly affected by the diversity of editing targets and sequence length, as determined through experiments. 2) From a model perspective, this article explores the factors that affect the performance of editing models. The results indicate a strong correlation between the L1-norm of the editing model layer and the editing accuracy, and clarify that this is an important factor leading to the bottleneck of editing performance. Finally, in order to improve the performance of the editing model, this paper further proposes a Dump for Sequence (D4S) method, which successfully overcomes the previous editing bottleneck by reducing the L1-norm of the editing layer, allowing users to perform multiple effective edits and minimizing model damage. Our code is available at https://github.com/nlpkeg/D4S.
Authors: Khurram Yamin, Shantanu Gupta, Gaurav R. Ghosal, Zachary C. Lipton, Bryan Wilder
Abstract: In this work, we investigate the causal reasoning abilities of large language models (LLMs) through the representative problem of inferring causal relationships from narratives. We find that even state-of-the-art language models rely on unreliable shortcuts, both in terms of the narrative presentation and their parametric knowledge. For example, LLMs tend to determine causal relationships based on the topological ordering of events (i.e., earlier events cause later ones), resulting in lower performance whenever events are not narrated in their exact causal order. Similarly, we demonstrate that LLMs struggle with long-term causal reasoning and often fail when the narratives are long and contain many events. Additionally, we show LLMs appear to rely heavily on their parametric knowledge at the expense of reasoning over the provided narrative. This degrades their abilities whenever the narrative opposes parametric knowledge. We extensively validate these failure modes through carefully controlled synthetic experiments, as well as evaluations on real-world narratives. Finally, we observe that explicitly generating a causal graph generally improves performance while naive chain-of-thought is ineffective. Collectively, our results distill precise failure modes of current state-of-the-art models and can pave the way for future techniques to enhance causal reasoning in LLMs.
Authors: Tianyi Qiu
Abstract: Social choice theory is the study of preference aggregation across a population, used both in mechanism design for human agents and in the democratic alignment of language models. In this study, we propose the representative social choice framework for the modeling of democratic representation in collective decisions, where the number of issues and individuals are too large for mechanisms to consider all preferences directly. These scenarios are widespread in real-world decision-making processes, such as jury trials, indirect elections, legislation processes, corporate governance, and, more recently, language model alignment. In representative social choice, the population is represented by a finite sample of individual-issue pairs based on which social choice decisions are made. We show that many of the deepest questions in representative social choice can be naturally formulated as statistical learning problems, and prove the generalization properties of social choice mechanisms using the theory of machine learning. We further formulate axioms for representative social choice, and prove Arrow-like impossibility theorems with new combinatorial tools of analysis. Our framework introduces the representative approach to social choice, opening up research directions at the intersection of social choice, learning theory, and AI alignment.
Authors: Yingzhe Peng, Xiaoting Qin, Zhiyang Zhang, Jue Zhang, Qingwei Lin, Xu Yang, Dongmei Zhang, Saravan Rajmohan, Qi Zhang
Abstract: The rise of large language models (LLMs) has revolutionized user interactions with knowledge-based systems, enabling chatbots to synthesize vast amounts of information and assist with complex, exploratory tasks. However, LLM-based chatbots often struggle to provide personalized support, particularly when users start with vague queries or lack sufficient contextual information. This paper introduces the Collaborative Assistant for Personalized Exploration (CARE), a system designed to enhance personalization in exploratory tasks by combining a multi-agent LLM framework with a structured user interface. CARE's interface consists of a Chat Panel, Solution Panel, and Needs Panel, enabling iterative query refinement and dynamic solution generation. The multi-agent framework collaborates to identify both explicit and implicit user needs, delivering tailored, actionable solutions. In a within-subject user study with 22 participants, CARE was consistently preferred over a baseline LLM chatbot, with users praising its ability to reduce cognitive load, inspire creativity, and provide more tailored solutions. Our findings highlight CARE's potential to transform LLM-based systems from passive information retrievers to proactive partners in personalized problem-solving and exploration.
Authors: Abhimanyu Das, Matthew Faw, Rajat Sen, Yichen Zhou
Abstract: Motivated by the recent success of time-series foundation models for zero-shot forecasting, we present a methodology for $\textit{in-context fine-tuning}$ of a time-series foundation model. In particular, we design a pretrained foundation model that can be prompted (at inference time) with multiple time-series examples, in order to forecast a target time-series into the future. Our foundation model is specifically trained to utilize examples from multiple related time-series in its context window (in addition to the history of the target time-series) to help it adapt to the specific distribution of the target domain at inference time. We show that such a foundation model that uses in-context examples at inference time can obtain much better performance on popular forecasting benchmarks compared to supervised deep learning methods, statistical models, as well as other time-series foundation models. Interestingly, our in-context fine-tuning approach even rivals the performance of a foundation model that is explicitly fine-tuned on the target domain.
Authors: Neil Chowdhury, Franklin Wang, Sumedh Shenoy, Douwe Kiela, Sarah Schwettmann, Tristan Thrush
Abstract: Multimodal models leverage large-scale pre-training to achieve strong but still imperfect performance on tasks such as image captioning, visual question answering, and cross-modal retrieval. In this paper, we present a simple and efficient method for correcting errors in trained contrastive image-text retrieval models with no additional training, called Nearest Neighbor Normalization (NNN). We show an improvement on retrieval metrics in both text retrieval and image retrieval for all of the contrastive models that we tested (CLIP, BLIP, ALBEF, SigLIP, BEiT) and for both of the datasets that we used (MS-COCO and Flickr30k). NNN requires a reference database, but does not require any training on this database, and can even increase the retrieval accuracy of a model after finetuning.
Authors: Chao Huang, Susan Liang, Yunlong Tang, Yapeng Tian, Anurag Kumar, Chenliang Xu
Abstract: Text-guided diffusion models have revolutionized generative tasks by producing high-fidelity content from text descriptions. They have also enabled an editing paradigm where concepts can be replaced through text conditioning (e.g., a dog to a tiger). In this work, we explore a novel approach: instead of replacing a concept, can we enhance or suppress the concept itself? Through an empirical study, we identify a trend where concepts can be decomposed in text-guided diffusion models. Leveraging this insight, we introduce ScalingConcept, a simple yet effective method to scale decomposed concepts up or down in real input without introducing new elements. To systematically evaluate our approach, we present the WeakConcept-10 dataset, where concepts are imperfect and need to be enhanced. More importantly, ScalingConcept enables a variety of novel zero-shot applications across image and audio domains, including tasks such as canonical pose generation and generative sound highlighting or removal.
Authors: Fu Feng, Yucheng Xie, Jing Wang, Xin Geng
Abstract: Creativity, both in human and diffusion models, remains an inherently abstract concept; thus, simply adding "creative" to a prompt does not yield reliable semantic recognition by the model. In this work, we concretize the abstract notion of "creative" through the TP2O task, which aims to merge two unrelated concepts, and introduce CreTok, redefining "creative" as the token $\texttt{
Authors: Biman Barua, M. Shamim Kaiser
Abstract: This paper introduces a microservices architecture for the purpose of enhancing the flexibility and performance of an airline reservation system. The architectural design incorporates Redis cache technologies, two different messaging systems (Kafka and RabbitMQ), two types of storages (MongoDB, and PostgreSQL). It also introduces authorization techniques, including secure communication through OAuth2 and JWT which is essential with the management of high-demand travel services. According to selected indicators, the architecture provides an impressive level of data consistency at 99.5% and a latency of data propagation of less than 75 ms allowing rapid and reliable intercommunication between microservices. A system throughput of 1050 events per second was achieved so that the acceptability level was maintained even during peak time. Redis caching reduced a 92% cache hit ratio on the database thereby lowering the burden on the database and increasing the speed of response. Further improvement of the systems scalability was done through the use of Docker and Kubernetes which enabled services to be expanded horizontally to cope with the changes in demand. The error rates were very low, at 0.2% further enhancing the efficiency of the system in handling real-time data integration. This approach is suggested to meet the specific needs of the airline reservation system. It is secure, fast, scalable, all serving to improve the user experience as well as the efficiency of operations. The low latency and high data integration levels and prevaiing efficient usage of the resources demonstrates the architecture ability to offer continued support in the ever growing high demand situations.
Authors: Heng-Jui Chang, Hongyu Gong, Changhan Wang, James Glass, Yu-An Chung
Abstract: Spoken language models (SLMs) have gained increasing attention with advancements in text-based, decoder-only language models. SLMs process text and speech, enabling simultaneous speech understanding and generation. This paper presents Double-Codebook Speaker-invariant Clustering (DC-Spin), which aims to improve speech tokenization by bridging audio signals and SLM tokens. DC-Spin extracts speaker-invariant tokens rich in phonetic information and resilient to input variations, enhancing zero-shot SLM tasks and speech resynthesis. We propose a chunk-wise approach to enable streamable DC-Spin without retraining and degradation. Comparisons of tokenization methods (self-supervised and neural audio codecs), model scalability, and downstream task proxies show that tokens easily modeled by an n-gram LM or aligned with phonemes offer strong performance, providing insights for designing speech tokenizers for SLMs.
Authors: Yongliang Shen, Kaitao Song, Xu Tan, Wenqi Zhang, Kan Ren, Siyu Yuan, Weiming Lu, Dongsheng Li, Yueting Zhuang
Abstract: In recent years, the remarkable progress of large language models (LLMs) has sparked interest in task automation, which involves decomposing complex tasks described by user instructions into sub-tasks and invoking external tools to execute them, playing a central role in autonomous agents. However, there is a lack of systematic and standardized benchmarks to promote the development of LLMs in task automation. To address this, we introduce TaskBench, a comprehensive framework to evaluate the capability of LLMs in task automation. Specifically, task automation can be divided into three critical stages: task decomposition, tool selection, and parameter prediction. To tackle the complexities inherent in these stages, we introduce the concept of Tool Graph to represent decomposed tasks and adopt a back-instruct method to generate high-quality user instructions. We propose TaskEval, a multi-faceted evaluation methodology that assesses LLM performance across these three stages. Our approach combines automated construction with rigorous human verification, ensuring high consistency with human evaluation. Experimental results demonstrate that TaskBench effectively reflects the capabilities of various LLMs in task automation. It provides insights into model performance across different task complexities and domains, pushing the boundaries of what current models can achieve. TaskBench offers a scalable, adaptable, and reliable benchmark for advancing LLM-based autonomous agents.
Authors: Boyi Zeng, Lizheng Wang, Yuncong Hu, Yi Xu, Chenghu Zhou, Xinbing Wang, Yu Yu, Zhouhan Lin
Abstract: Protecting the copyright of large language models (LLMs) has become crucial due to their resource-intensive training and accompanying carefully designed licenses. However, identifying the original base model of an LLM is challenging due to potential parameter alterations. In this study, we introduce HuRef, a human-readable fingerprint for LLMs that uniquely identifies the base model without interfering with training or exposing model parameters to the public. We first observe that the vector direction of LLM parameters remains stable after the model has converged during pretraining, with negligible perturbations through subsequent training steps, including continued pretraining, supervised fine-tuning, and RLHF, which makes it a sufficient condition to identify the base model. The necessity is validated by continuing to train an LLM with an extra term to drive away the model parameters' direction and the model becomes damaged. However, this direction is vulnerable to simple attacks like dimension permutation or matrix rotation, which significantly change it without affecting performance. To address this, leveraging the Transformer structure, we systematically analyze potential attacks and define three invariant terms that identify an LLM's base model. Due to the potential risk of information leakage, we cannot publish invariant terms directly. Instead, we map them to a Gaussian vector using an encoder, then convert it into a natural image using StyleGAN2, and finally publish the image. In our black-box setting, all fingerprinting steps are internally conducted by the LLMs owners. To ensure the published fingerprints are honestly generated, we introduced Zero-Knowledge Proof (ZKP). Experimental results across various LLMs demonstrate the effectiveness of our method. The code is available at https://github.com/LUMIA-Group/HuRef.
Authors: Lena Held, Ivan Habernal
Abstract: Why does an argument end up in the final court decision? Was it deliberated or questioned during the oral hearings? Was there something in the hearings that triggered a particular judge to write a dissenting opinion? Despite the availability of the final judgments of the European Court of Human Rights (ECHR), none of these legal research questions can currently be answered as the ECHR's multilingual oral hearings are not transcribed, structured, or speaker-attributed. We address this fundamental gap by presenting LaCour!, the first corpus of textual oral arguments of the ECHR, consisting of 154 full hearings (2.1 million tokens from over 267 hours of video footage) in English, French, and other court languages, each linked to the corresponding final judgment documents. In addition to the transcribed and partially manually corrected text from the video, we provide sentence-level timestamps and manually annotated role and language labels. We also showcase LaCour! in a set of preliminary experiments that explore the interplay between questions and dissenting opinions. Apart from the use cases in legal NLP, we hope that law students or other interested parties will also use LaCour! as a learning resource, as it is freely available in various formats at https://huggingface.co/datasets/TrustHLT/LaCour.
Authors: Derong Xu, Wei Chen, Wenjun Peng, Chao Zhang, Tong Xu, Xiangyu Zhao, Xian Wu, Yefeng Zheng, Yang Wang, Enhong Chen
Abstract: Information extraction (IE) aims to extract structural knowledge from plain natural language texts. Recently, generative Large Language Models (LLMs) have demonstrated remarkable capabilities in text understanding and generation. As a result, numerous works have been proposed to integrate LLMs for IE tasks based on a generative paradigm. To conduct a comprehensive systematic review and exploration of LLM efforts for IE tasks, in this study, we survey the most recent advancements in this field. We first present an extensive overview by categorizing these works in terms of various IE subtasks and techniques, and then we empirically analyze the most advanced methods and discover the emerging trend of IE tasks with LLMs. Based on a thorough review conducted, we identify several insights in technique and promising research directions that deserve further exploration in future studies. We maintain a public repository and consistently update related works and resources on GitHub (\href{https://github.com/quqxui/Awesome-LLM4IE-Papers}{LLM4IE repository})
Authors: Fanghua Ye, Mingming Yang, Jianhui Pang, Longyue Wang, Derek F. Wong, Emine Yilmaz, Shuming Shi, Zhaopeng Tu
Abstract: The proliferation of open-source Large Language Models (LLMs) from various institutions has highlighted the urgent need for comprehensive evaluation methods. However, current evaluation platforms, such as the widely recognized HuggingFace open LLM leaderboard, neglect a crucial aspect -- uncertainty, which is vital for thoroughly assessing LLMs. To bridge this gap, we introduce a new benchmarking approach for LLMs that integrates uncertainty quantification. Our examination involves nine LLMs (LLM series) spanning five representative natural language processing tasks. Our findings reveal that: I) LLMs with higher accuracy may exhibit lower certainty; II) Larger-scale LLMs may display greater uncertainty compared to their smaller counterparts; and III) Instruction-finetuning tends to increase the uncertainty of LLMs. These results underscore the significance of incorporating uncertainty in the evaluation of LLMs.
Authors: Yanbin Wei, Shuai Fu, Weisen Jiang, Zejian Zhang, Zhixiong Zeng, Qi Wu, James T. Kwok, Yu Zhang
Abstract: Large Language Models (LLMs) are increasingly used for various tasks with graph structures. Though LLMs can process graph information in a textual format, they overlook the rich vision modality, which is an intuitive way for humans to comprehend structural information and conduct general graph reasoning. The potential benefits and capabilities of representing graph structures as visual images (i.e., $\textit{visual graph}$) are still unexplored. To fill the gap, we innovatively propose an end-to-end framework, called $\textbf{G}$raph to v$\textbf{I}$sual and $\textbf{T}$extual Integr$\textbf{A}$tion (GITA), which firstly incorporates visual graphs into general graph reasoning. Besides, we establish $\textbf{G}$raph-based $\textbf{V}$ision-$\textbf{L}$anguage $\textbf{Q}$uestion $\textbf{A}$nswering (GVLQA) dataset from existing graph data, which is the first vision-language dataset for general graph reasoning purposes. Extensive experiments on the GVLQA dataset and five real-world datasets show that GITA outperforms mainstream LLMs in terms of general graph reasoning capabilities. Moreover, We highlight the effectiveness of the layout augmentation on visual graphs and pretraining on the GVLQA dataset.
Authors: Yuxia Wang, Minghan Wang, Muhammad Arslan Manzoor, Fei Liu, Georgi Georgiev, Rocktim Jyoti Das, Preslav Nakov
Abstract: Large language models (LLMs), especially when instruction-tuned for chat, have become part of our daily lives, freeing people from the process of searching, extracting, and integrating information from multiple sources by offering a straightforward answer to a variety of questions in a single place. Unfortunately, in many cases, LLM responses are factually incorrect, which limits their applicability in real-world scenarios. As a result, research on evaluating and improving the factuality of LLMs has attracted a lot of attention recently. In this survey, we critically analyze existing work with the aim to identify the major challenges and their associated causes, pointing out to potential solutions for improving the factuality of LLMs, and analyzing the obstacles to automated factuality evaluation for open-ended text generation. We further offer an outlook on where future research should go.
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: Jio Oh, Soyeon Kim, Junseok Seo, Jindong Wang, Ruochen Xu, Xing Xie, Steven Euijong Whang
Abstract: Large language models (LLMs) have achieved unprecedented performances in various applications, yet evaluating them is still challenging. Existing benchmarks are either manually constructed or are automatic, but lack the ability to evaluate the thought process of LLMs with arbitrary complexity. We contend that utilizing existing relational databases based on the entity-relationship (ER) model is a promising approach for constructing benchmarks as they contain structured knowledge that can be used to question LLMs. Unlike knowledge graphs, which are also used to evaluate LLMs, relational databases have integrity constraints that can be used to better construct complex in-depth questions and verify answers: (1) functional dependencies can be used to pinpoint critical keywords that an LLM must know to properly answer a given question containing certain attribute values; and (2) foreign key constraints can be used to join relations and construct multi-hop questions, which can be arbitrarily long and used to debug intermediate answers. We thus propose ERBench, which uses these integrity constraints to convert any database into an LLM benchmark. ERBench supports continuous evaluation as databases change, multimodal questions, and various prompt engineering techniques. In our experiments, we construct LLM benchmarks using databases of multiple domains and make an extensive comparison of contemporary LLMs. We show how ERBench can properly evaluate any LLM by not only checking for answer correctness, but also effectively verifying the rationales by looking for the right keywords.
Authors: Kyomin Hwang, Suyoung Kim, JunHoo Lee, Nojun Kwak
Abstract: Large Models (LMs) have heightened expectations for the potential of general AI as they are akin to human intelligence. This paper shows that recent large models such as Stable Diffusion and DALL-E3 also share the vulnerability of human intelligence, namely the "white bear phenomenon". We investigate the causes of the white bear phenomenon by analyzing their representation space. Based on this analysis, we propose a simple prompt-based attack method, which generates figures prohibited by the LM provider's policy. To counter these attacks, we introduce prompt-based defense strategies inspired by cognitive therapy techniques, successfully mitigating attacks by up to 48.22\%.
Authors: Bishwaranjan Bhattacharjee, Aashka Trivedi, Masayasu Muraoka, Muthukumaran Ramasubramanian, Takuma Udagawa, Iksha Gurung, Nishan Pantha, Rong Zhang, Bharath Dandala, Rahul Ramachandran, Manil Maskey, Kaylin Bugbee, Mike Little, Elizabeth Fancher, Irina Gerasimov, Armin Mehrabian, Lauren Sanders, Sylvain Costes, Sergi Blanco-Cuaresma, Kelly Lockhart, Thomas Allen, Felix Grezes, Megan Ansdell, Alberto Accomazzi, Yousef El-Kurdi, Davis Wertheimer, Birgit Pfitzmann, Cesar Berrospi Ramis, Michele Dolfi, Rafael Teixeira de Lima, Panagiotis Vagenas, S. Karthik Mukkavilli, Peter Staar, Sanaz Vahidinia, Ryan McGranaghan, Tsendgar Lee
Abstract: Large language models (LLMs) trained on general domain corpora showed remarkable results on natural language processing (NLP) tasks. However, previous research demonstrated LLMs trained using domain-focused corpora perform better on specialized tasks. Inspired by this insight, we developed INDUS, a comprehensive suite of LLMs tailored for the closely-related domains of Earth science, biology, physics, heliophysics, planetary sciences and astrophysics, and trained using curated scientific corpora drawn from diverse data sources. The suite of models include: (1) an encoder model trained using domain-specific vocabulary and corpora to address NLP tasks, (2) a contrastive-learning based text embedding model trained using a diverse set of datasets to address information retrieval tasks and (3) smaller versions of these models created using knowledge distillation for applications which have latency or resource constraints. We also created three new scientific benchmark datasets, CLIMATE-CHANGE NER (entity-recognition), NASA-QA (extractive QA) and NASA-IR (IR) to accelerate research in these multi-disciplinary fields. We show that our models outperform both general-purpose (RoBERTa) and domain-specific (SCIBERT) encoders on these new tasks as well as existing tasks in the domains of interest. Furthermore, we demonstrate the use of these models in two industrial settings -- as a retrieval model for large-scale vector search applications and in automatic content tagging systems.
Authors: Domenic Rosati, Jan Wehner, Kai Williams, {\L}ukasz Bartoszcze, David Atanasov, Robie Gonzales, Subhabrata Majumdar, Carsten Maple, Hassan Sajjad, Frank Rudzicz
Abstract: Releasing open-source large language models (LLMs) presents a dual-use risk since bad actors can easily fine-tune these models for harmful purposes. Even without the open release of weights, weight stealing and fine-tuning APIs make closed models vulnerable to harmful fine-tuning attacks (HFAs). While safety measures like preventing jailbreaks and improving safety guardrails are important, such measures can easily be reversed through fine-tuning. In this work, we propose Representation Noising (RepNoise), a defence mechanism that operates even when attackers have access to the weights. RepNoise works by removing information about harmful representations such that it is difficult to recover them during fine-tuning. Importantly, our defence is also able to generalize across different subsets of harm that have not been seen during the defence process as long as they are drawn from the same distribution of the attack set. Our method does not degrade the general capability of LLMs and retains the ability to train the model on harmless tasks. We provide empirical evidence that the efficacy of our defence lies in its ``depth'': the degree to which information about harmful representations is removed across all layers of the LLM. We also find areas where RepNoise still remains ineffective and highlight how those limitations can inform future research.
Authors: Zhijing Jin, Nils Heil, Jiarui Liu, Shehzaad Dhuliawala, Yahang Qi, Bernhard Sch\"olkopf, Rada Mihalcea, Mrinmaya Sachan
Abstract: Implicit Personalization (IP) is a phenomenon of language models inferring a user's background from the implicit cues in the input prompts and tailoring the response based on this inference. While previous work has touched upon various instances of this problem, there lacks a unified framework to study this behavior. This work systematically studies IP through a rigorous mathematical formulation, a multi-perspective moral reasoning framework, and a set of case studies. Our theoretical foundation for IP relies on a structural causal model and introduces a novel method, indirect intervention, to estimate the causal effect of a mediator variable that cannot be directly intervened upon. Beyond the technical approach, we also introduce a set of moral reasoning principles based on three schools of moral philosophy to study when IP may or may not be ethically appropriate. Equipped with both mathematical and ethical insights, we present three diverse case studies illustrating the varied nature of the IP problem and offer recommendations for future research. Our code is at https://github.com/jiarui-liu/IP, and our data is at https://huggingface.co/datasets/Jerry999/ImplicitPersonalizationData.
URLs: https://github.com/jiarui-liu/IP,, https://huggingface.co/datasets/Jerry999/ImplicitPersonalizationData.
Authors: Li Ji-An, Corey Y. Zhou, Marcus K. Benna, Marcelo G. Mattar
Abstract: Understanding connections between artificial and biological intelligent systems can reveal fundamental principles of general intelligence. While many artificial intelligence models have a neuroscience counterpart, such connections are largely missing in Transformer models and the self-attention mechanism. Here, we examine the relationship between interacting attention heads and human episodic memory. We focus on induction heads, which contribute to in-context learning in Transformer-based large language models (LLMs). We demonstrate that induction heads are behaviorally, functionally, and mechanistically similar to the contextual maintenance and retrieval (CMR) model of human episodic memory. Our analyses of LLMs pre-trained on extensive text data show that CMR-like heads often emerge in the intermediate and late layers, qualitatively mirroring human memory biases. The ablation of CMR-like heads suggests their causal role in in-context learning. Our findings uncover a parallel between the computational mechanisms of LLMs and human memory, offering valuable insights into both research fields.
Authors: Boshi Wang, Xiang Yue, Yu Su, Huan Sun
Abstract: We study whether transformers can learn to implicitly reason over parametric knowledge, a skill that even the most capable language models struggle with. Focusing on two representative reasoning types, composition and comparison, we consistently find that transformers can learn implicit reasoning, but only through grokking, i.e., extended training far beyond overfitting. The levels of generalization also vary across reasoning types: when faced with out-of-distribution examples, transformers fail to systematically generalize for composition but succeed for comparison. We delve into the model's internals throughout training, conducting analytical experiments that reveal: 1) the mechanism behind grokking, such as the formation of the generalizing circuit and its relation to the relative efficiency of generalizing and memorizing circuits, and 2) the connection between systematicity and the configuration of the generalizing circuit. Our findings guide data and training setup to better induce implicit reasoning and suggest potential improvements to the transformer architecture, such as encouraging cross-layer knowledge sharing. Furthermore, we demonstrate that for a challenging reasoning task with a large search space, GPT-4-Turbo and Gemini-1.5-Pro based on non-parametric memory fail badly regardless of prompting styles or retrieval augmentation, while a fully grokked transformer can achieve near-perfect accuracy, showcasing the power of parametric memory for complex reasoning.
Authors: Felipe Maia Polo, Ronald Xu, Lucas Weber, M\'irian Silva, Onkar Bhardwaj, Leshem Choshen, Allysson Flavio Melo de Oliveira, Yuekai Sun, Mikhail Yurochkin
Abstract: Most popular benchmarks for comparing LLMs rely on a limited set of prompt templates, which may not fully capture the LLMs' abilities and can affect the reproducibility of results on leaderboards. Many recent works empirically verify prompt sensitivity and advocate for changes in LLM evaluation. In this paper, we consider the problem of estimating the performance distribution across many prompt variants instead of finding a single prompt to evaluate with. We introduce PromptEval, a method for estimating performance across a large set of prompts borrowing strength across prompts and examples to produce accurate estimates under practical evaluation budgets. The resulting distribution can be used to obtain performance quantiles to construct various robust performance metrics (e.g., top 95% quantile or median). We prove that PromptEval consistently estimates the performance distribution and demonstrate its efficacy empirically on three prominent LLM benchmarks: MMLU, BIG-bench Hard, and LMentry; for example, PromptEval can accurately estimate performance quantiles across 100 prompt templates on MMLU with a budget equivalent to two single-prompt evaluations. Moreover, we show how PromptEval can be useful in LLM-as-a-judge and best prompt identification applications.
Authors: Chenyang Le, Yao Qian, Dongmei Wang, Long Zhou, Shujie Liu, Xiaofei Wang, Midia Yousefi, Yanmin Qian, Jinyu Li, Sheng Zhao, Michael Zeng
Abstract: There is a rising interest and trend in research towards directly translating speech from one language to another, known as end-to-end speech-to-speech translation. However, most end-to-end models struggle to outperform cascade models, i.e., a pipeline framework by concatenating speech recognition, machine translation and text-to-speech models. The primary challenges stem from the inherent complexities involved in direct translation tasks and the scarcity of data. In this study, we introduce a novel model framework TransVIP that leverages diverse datasets in a cascade fashion yet facilitates end-to-end inference through joint probability. Furthermore, we propose two separated encoders to preserve the speaker's voice characteristics and isochrony from the source speech during the translation process, making it highly suitable for scenarios such as video dubbing. Our experiments on the French-English language pair demonstrate that our model outperforms the current state-of-the-art speech-to-speech translation model.
Authors: Fangcong Yin, Xi Ye, Greg Durrett
Abstract: Recent work in interpretability shows that large language models (LLMs) can be adapted for new tasks in a learning-free way: it is possible to intervene on LLM representations to elicit desired behaviors for alignment. For instance, adding certain bias vectors to the outputs of certain attention heads is reported to boost the truthfulness of models. In this work, we show that localized fine-tuning serves as an effective alternative to such representation intervention methods. We introduce a framework called Localized Fine-Tuning on LLM Representations (LoFiT), which identifies a subset of attention heads that are most important for learning a specific task, then trains offset vectors to add to the model's hidden representations at those selected heads. LoFiT localizes to a sparse set of heads (3%-10%) and learns the offset vectors from limited training data, comparable to the settings used for representation intervention. For truthfulness and reasoning tasks, we find that LoFiT's intervention vectors are more effective for LLM adaptation than vectors from representation intervention methods such as Inference-time Intervention. We also find that the localization step is important: selecting a task-specific set of attention heads can lead to higher performance than intervening on heads selected for a different task. Finally, across 7 tasks we study, LoFiT achieves comparable performance to other parameter-efficient fine-tuning methods such as LoRA, despite modifying 20x-200x fewer parameters than these methods.
Authors: David Samuel
Abstract: While in-context learning is commonly associated with causal language models, such as GPT, we demonstrate that this capability also 'emerges' in masked language models. Through an embarrassingly simple inference technique, we enable an existing masked model, DeBERTa, to perform generative tasks without additional training or architectural changes. Our evaluation reveals that the masked and causal language models behave very differently, as they clearly outperform each other on different categories of tasks. These complementary strengths suggest that the field's focus on causal models for in-context learning may be limiting - both architectures can develop these capabilities, but with distinct advantages; pointing toward promising hybrid approaches that combine the strengths of both objectives.
Authors: Andrew M. Bean, Simi Hellsten, Harry Mayne, Jabez Magomere, Ethan A. Chi, Ryan Chi, Scott A. Hale, Hannah Rose Kirk
Abstract: In this paper, we present the LingOly benchmark, a novel benchmark for advanced reasoning abilities in large language models. Using challenging Linguistic Olympiad puzzles, we evaluate (i) capabilities for in-context identification and generalisation of linguistic patterns in very low-resource or extinct languages, and (ii) abilities to follow complex task instructions. The LingOly benchmark covers more than 90 mostly low-resource languages, minimising issues of data contamination, and contains 1,133 problems across 6 formats and 5 levels of human difficulty. We assess performance with both direct accuracy and comparison to a no-context baseline to penalise memorisation. Scores from 11 state-of-the-art LLMs demonstrate the benchmark to be challenging, and models perform poorly on the higher difficulty problems. On harder problems, even the top model only achieved 38.7% accuracy, a 24.7% improvement over the no-context baseline. Large closed models typically outperform open models, and in general, the higher resource the language, the better the scores. These results indicate, in absence of memorisation, true multi-step out-of-domain reasoning remains a challenge for current language models.
Authors: Chris Yuhao Liu, Yaxuan Wang, Jeffrey Flanigan, Yang Liu
Abstract: Large language models (LLMs) have advanced to encompass extensive knowledge across diverse domains. Yet controlling what a large language model should not know is important for ensuring alignment and thus safe use. However, accurately and efficiently unlearning knowledge from an LLM remains challenging due to the potential collateral damage caused by the fuzzy boundary between retention and forgetting, and the large computational requirements for optimization across state-of-the-art models with hundreds of billions of parameters. In this work, we present \textbf{Embedding-COrrupted (ECO) Prompts}, a lightweight unlearning framework for large language models to address both the challenges of knowledge entanglement and unlearning efficiency. Instead of relying on the LLM itself to unlearn, we enforce an unlearned state during inference by employing a prompt classifier to identify and safeguard prompts to forget. We learn corruptions added to prompt embeddings via zeroth order optimization toward the unlearning objective offline and corrupt prompts flagged by the classifier during inference. We find that these embedding-corrupted prompts not only lead to desirable outputs that satisfy the unlearning objective but also closely approximate the output from a model that has never been trained on the data intended for forgetting. Through extensive experiments on unlearning, we demonstrate the superiority of our method in achieving promising unlearning at \textit{nearly zero side effects} in general domains and domains closely related to the unlearned ones. Additionally, we highlight the scalability of our method to 100 LLMs, ranging from 0.5B to 236B parameters, incurring no additional cost as the number of parameters increases. We have made our code publicly available at \url{https://github.com/chrisliu298/llm-unlearn-eco}.
Authors: Cheng-Kuang Wu, Zhi Rui Tam, Chieh-Yen Lin, Yun-Nung Chen, Hung-yi Lee
Abstract: Recent works have shown that large language model (LLM) agents are able to improve themselves from experience, which is an important ability for continuous enhancement post-deployment. However, existing benchmarks primarily evaluate their innate capabilities and do not assess their ability to improve over time. To address this gap, we introduce StreamBench, a pioneering benchmark designed to evaluate the continuous improvement of LLM agents over an input-feedback sequence. StreamBench simulates an online learning environment where LLMs receive a continuous flow of feedback stream and iteratively enhance their performance. In addition, we propose several simple yet effective baselines for improving LLMs on StreamBench, and provide a comprehensive analysis to identify critical components that contribute to successful streaming strategies. Our work serves as a stepping stone towards developing effective online learning strategies for LLMs, paving the way for more adaptive AI systems in streaming scenarios. Source code: https://github.com/stream-bench/stream-bench. Benchmark website: https://stream-bench.github.io.
URLs: https://github.com/stream-bench/stream-bench., https://stream-bench.github.io.
Authors: Xuan Zhang, Chao Du, Tianyu Pang, Qian Liu, Wei Gao, Min Lin
Abstract: The recent development of chain-of-thought (CoT) decoding has enabled large language models (LLMs) to generate explicit logical reasoning paths for complex problem-solving. However, research indicates that these paths are not always deliberate and optimal. The tree-of-thought (ToT) method employs tree-searching to extensively explore the reasoning space and find better reasoning paths that CoT decoding might overlook. This deliberation, however, comes at the cost of significantly increased inference complexity. In this work, we demonstrate that fine-tuning LLMs leveraging the search tree constructed by ToT allows CoT to achieve similar or better performance, thereby avoiding the substantial inference burden. This is achieved through Chain of Preference Optimization (CPO), where LLMs are fine-tuned to align each step of the CoT reasoning paths with those of ToT using the inherent preference information in the tree-search process. Extensive experimental results show that CPO significantly improves LLM performance in solving a variety of complex problems, including question answering, fact verification, and arithmetic reasoning, demonstrating its effectiveness. Our code is available at https://github.com/sail-sg/CPO.
Authors: Muhammad Arslan Manzoor, Yuxia Wang, Minghan Wang, Preslav Nakov
Abstract: Empathy plays a pivotal role in fostering prosocial behavior, often triggered by the sharing of personal experiences through narratives. However, modeling empathy using NLP approaches remains challenging due to its deep interconnection with human interaction dynamics. Previous approaches, which involve fine-tuning language models (LMs) on human-annotated empathic datasets, have had limited success. In our pursuit of improving empathy understanding in LMs, we propose several strategies, including contrastive learning with masked LMs and supervised fine-tuning with large language models. While these methods show improvements over previous methods, the overall results remain unsatisfactory. To better understand this trend, we performed an analysis which reveals a low agreement among annotators. This lack of consensus hinders training and highlights the subjective nature of the task. We also explore the cultural impact on annotations. To study this, we meticulously collected story pairs in Urdu language and find that subjectivity in interpreting empathy among annotators appears to be independent of cultural background. Our systematic exploration of LMs' understanding of empathy reveals substantial opportunities for further investigation in both task formulation and modeling.
Authors: Yingzhe Peng, Chenduo Hao, Xu Yang, Jiawei Peng, Xinting Hu, Xin Geng
Abstract: As language models continue to scale, Large Language Models (LLMs) have exhibited emerging capabilities in In-Context Learning (ICL), enabling them to solve language tasks by prefixing a few in-context demonstrations (ICDs) as context. Inspired by these advancements, researchers have extended these techniques to develop Large Multimodal Models (LMMs) with ICL capabilities. However, applying ICL usually faces two major challenges: 1) using more ICDs will largely increase the inference time and 2) the performance is sensitive to the selection of ICDs. These challenges are further exacerbated in LMMs due to the integration of multiple data types and the combinational complexity of multimodal ICDs. Recently, to address these challenges, some NLP studies introduce non-learnable In-Context Vectors (ICVs) which extract useful task information from ICDs into a single vector and then insert it into the LLM to help solve the corresponding task. However, although useful in simple NLP tasks, these non-learnable methods fail to handle complex multimodal tasks like Visual Question Answering (VQA). In this study, we propose Learnable In-Context VEctor (LIVE) to distill essential task information from demonstrations, improving ICL performance in LMMs. Experiments show that LIVE can significantly reduce computational costs while enhancing accuracy in VQA tasks compared to traditional ICL and other non-learnable ICV methods. The code is available at \url{https://github.com/ForJadeForest/LIVE-Learnable-In-Context-Vector}.
URLs: https://github.com/ForJadeForest/LIVE-Learnable-In-Context-Vector
Authors: Allen Roush, Yusuf Shabazz, Arvind Balaji, Peter Zhang, Stefano Mezza, Markus Zhang, Sanjay Basu, Sriram Vishwanath, Mehdi Fatemi, Ravid Shwartz-Ziv
Abstract: We introduce OpenDebateEvidence, a comprehensive dataset for argument mining and summarization sourced from the American Competitive Debate community. This dataset includes over 3.5 million documents with rich metadata, making it one of the most extensive collections of debate evidence. OpenDebateEvidence captures the complexity of arguments in high school and college debates, providing valuable resources for training and evaluation. Our extensive experiments demonstrate the efficacy of fine-tuning state-of-the-art large language models for argumentative abstractive summarization across various methods, models, and datasets. By providing this comprehensive resource, we aim to advance computational argumentation and support practical applications for debaters, educators, and researchers. OpenDebateEvidence is publicly available to support further research and innovation in computational argumentation. Access it here: https://huggingface.co/datasets/Yusuf5/OpenCaselist
Authors: Guilherme Penedo, Hynek Kydl\'i\v{c}ek, Loubna Ben allal, Anton Lozhkov, Margaret Mitchell, Colin Raffel, Leandro Von Werra, Thomas Wolf
Abstract: The performance of a large language model (LLM) depends heavily on the quality and size of its pretraining dataset. However, the pretraining datasets for state-of-the-art open LLMs like Llama 3 and Mixtral are not publicly available and very little is known about how they were created. In this work, we introduce FineWeb, a 15-trillion token dataset derived from 96 Common Crawl snapshots that produces better-performing LLMs than other open pretraining datasets. To advance the understanding of how best to curate high-quality pretraining datasets, we carefully document and ablate all of the design choices used in FineWeb, including in-depth investigations of deduplication and filtering strategies. In addition, we introduce FineWeb-Edu, a 1.3-trillion token collection of educational text filtered from FineWeb. LLMs pretrained on FineWeb-Edu exhibit dramatically better performance on knowledge- and reasoning-intensive benchmarks like MMLU and ARC. Along with our datasets, we publicly release our data curation codebase and all of the models trained during our ablation experiments.
Authors: Venkata S Govindarajan, Matianyu Zang, Kyle Mahowald, David Beaver, Junyi Jessy Li
Abstract: The variations between in-group and out-group speech (intergroup bias) are subtle and could underlie many social phenomena like stereotype perpetuation and implicit bias. In this paper, we model the intergroup bias as a tagging task on English sports comments from forums dedicated to fandom for NFL teams. We curate a unique dataset of over 6 million game-time comments from opposing perspectives (the teams in the game), each comment grounded in a non-linguistic description of the events that precipitated these comments (live win probabilities for each team). Expert and crowd annotations justify modeling the bias through tagging of implicit and explicit referring expressions and reveal the rich, contextual understanding of language and the world required for this task. For large-scale analysis of intergroup variation, we use LLMs for automated tagging, and discover that some LLMs perform best when prompted with linguistic descriptions of the win probability at the time of the comment, rather than numerical probability. Further, large-scale tagging of comments using LLMs uncovers linear variations in the form of referent across win probabilities that distinguish in-group and out-group utterances. Code and data are available at https://github.com/venkatasg/intergroup-nfl .
Authors: Irina Saparina, Mirella Lapata
Abstract: Practical semantic parsers are expected to understand user utterances and map them to executable programs, even when these are ambiguous. We introduce a new benchmark, AMBROSIA, which we hope will inform and inspire the development of text-to-SQL parsers capable of recognizing and interpreting ambiguous requests. Our dataset contains questions showcasing three different types of ambiguity (scope ambiguity, attachment ambiguity, and vagueness), their interpretations, and corresponding SQL queries. In each case, the ambiguity persists even when the database context is provided. This is achieved through a novel approach that involves controlled generation of databases from scratch. We benchmark various LLMs on AMBROSIA, revealing that even the most advanced models struggle to identify and interpret ambiguity in questions.
Authors: Dustin Wright, Arnav Arora, Nadav Borenstein, Srishti Yadav, Serge Belongie, Isabelle Augenstein
Abstract: Uncovering latent values and opinions embedded in large language models (LLMs) can help identify biases and mitigate potential harm. Recently, this has been approached by prompting LLMs with survey questions and quantifying the stances in the outputs towards morally and politically charged statements. However, the stances generated by LLMs can vary greatly depending on how they are prompted, and there are many ways to argue for or against a given position. In this work, we propose to address this by analysing a large and robust dataset of 156k LLM responses to the 62 propositions of the Political Compass Test (PCT) generated by 6 LLMs using 420 prompt variations. We perform coarse-grained analysis of their generated stances and fine-grained analysis of the plain text justifications for those stances. For fine-grained analysis, we propose to identify tropes in the responses: semantically similar phrases that are recurrent and consistent across different prompts, revealing natural patterns in the text that a given LLM is prone to produce. We find that demographic features added to prompts significantly affect outcomes on the PCT, reflecting bias, as well as disparities between the results of tests when eliciting closed-form vs. open domain responses. Additionally, patterns in the plain text rationales via tropes show that similar justifications are repeatedly generated across models and prompts even with disparate stances.
Authors: Bosi Wen, Pei Ke, Xiaotao Gu, Lindong Wu, Hao Huang, Jinfeng Zhou, Wenchuang Li, Binxin Hu, Wendy Gao, Jiaxin Xu, Yiming Liu, Jie Tang, Hongning Wang, Minlie Huang
Abstract: Instruction following is one of the fundamental capabilities of large language models (LLMs). As the ability of LLMs is constantly improving, they have been increasingly applied to deal with complex human instructions in real-world scenarios. Therefore, how to evaluate the ability of complex instruction-following of LLMs has become a critical research problem. Existing benchmarks mainly focus on modeling different types of constraints in human instructions while neglecting the composition of different constraints, which is an indispensable constituent in complex instructions. To this end, we propose ComplexBench, a benchmark for comprehensively evaluating the ability of LLMs to follow complex instructions composed of multiple constraints. We propose a hierarchical taxonomy for complex instructions, including 4 constraint types, 19 constraint dimensions, and 4 composition types, and manually collect a high-quality dataset accordingly. To make the evaluation reliable, we augment LLM-based evaluators with rules to effectively verify whether generated texts can satisfy each constraint and composition. Furthermore, we obtain the final evaluation score based on the dependency structure determined by different composition types. ComplexBench identifies significant deficiencies in existing LLMs when dealing with complex instructions with multiple constraints composition.
Authors: Kai Liu, Zhihang Fu, Chao Chen, Wei Zhang, Rongxin Jiang, Fan Zhou, Yaowu Chen, Yue Wu, Jieping Ye
Abstract: When reading long-form text, human cognition is complex and structurized. While large language models (LLMs) process input contexts through a causal and sequential perspective, this approach can potentially limit their ability to handle intricate and complex inputs effectively. To enhance LLM's cognition capability, this paper presents a novel concept of context structurization. Specifically, we transform the plain, unordered contextual sentences into well-ordered and hierarchically structurized elements. By doing so, LLMs can better grasp intricate and extended contexts through precise attention and information-seeking along the organized structures. Extensive evaluations are conducted across various model architectures and sizes (including a series of auto-regressive LLMs as well as BERT-like masking models) on a diverse set of NLP tasks (e.g., context-based question-answering, exhaustive hallucination evaluation, and passage-level dense retrieval). Empirical results show consistent and significant performance gains afforded by a single-round structurization. In particular, we boost the open-sourced LLaMA2-70B model to achieve comparable performance against GPT-3.5-Turbo as the hallucination evaluator. Besides, we show the feasibility of distilling advanced LLMs' language processing abilities to a smaller yet effective StruXGPT-7B to execute structurization, addressing the practicality of our approach. Code is available at https://github.com/alibaba/struxgpt.
Authors: Kai Liu, Ze Chen, Zhihang Fu, Rongxin Jiang, Fan Zhou, Yaowu Chen, Yue Wu, Jieping Ye
Abstract: This paper introduces a pioneering methodology, termed StructTuning, to efficiently transform foundation Large Language Models (LLMs) into domain specialists. It significantly reduces the training corpus requirement to a mere 0.3%, while achieving an impressive 50% of traditional knowledge injection performance. Our method is inspired by the educational processes of human students, particularly how structured domain knowledge from textbooks is assimilated and subsequently applied to tackle real-world challenges through specific exercises. Based on this, we propose a novel two-stage strategy for knowledge injection and alignment: Structure-aware Continual Pre-Training (SCPT) and Structure-aware Supervised Fine-Tuning (SSFT). In the SCPT phase, we automatically extract the domain knowledge taxonomy and reorganize the training corpora, enabling LLMs to effectively link textual segments to targeted knowledge points within the taxonomy. In the SSFT phase, we explicitly prompt models to elucidate the underlying knowledge structure in their outputs, leveraging the structured domain insight to address practical problems. Our ultimate method has undergone extensive evaluations across model architectures and scales, using closed-book question-answering tasks on LongBench and MMedBench datasets. Remarkably, our method demonstrates the potential of comparable improvement against the state-of-the-art MMedLM2 on MMedBench, while significantly reducing the training costs to 5%. This breakthrough paves the way for scaling up our StructTuning for stronger domain-specific LLMs with comprehensive data utilization. Code is available at https://github.com/alibaba/struxgpt.
Authors: Nunzio Lore, Sepehr Ilami, Babak Heydari
Abstract: As the performance of larger, newer Large Language Models continues to improve for strategic Theory of Mind (ToM) tasks, the demand for these state-of-the-art models increases commensurately. However, their deployment is costly both in terms of processing power and time. In this paper, we investigate the feasibility of creating smaller, highly-performing specialized algorithms by way of fine-tuning. To do this, we first present a large pre-trained model with 20 unique scenarios that combine different social contexts with games of varying social dilemmas, record its answers, and use them for Q&A fine-tuning on a smaller model of the same family. Our focus is on in-context game-theoretic decision-making, the same domain within which human interaction occurs and that requires both a theory of mind (or a semblance thereof) and an understanding of social dynamics. The smaller model is therefore trained not just on the answers provided, but also on the motivations provided by the larger model, which should contain advice and guidelines to navigate both strategic dilemmas and social cues. We find that the fine-tuned smaller language model consistently bridged the gap in performance between the smaller pre-trained version of the model and its larger relative and that its improvements extended in areas and contexts beyond the ones provided in the training examples, including on out-of-sample scenarios that include completely different game structures. On average for all games, through fine-tuning, the smaller model showed a 46% improvement measured as alignment towards the behavior of the larger model, with 100% representing indistinguishable behavior. When presented with out-of-sample social contexts and games, the fine-tuned model still displays remarkable levels of alignment, reaching an improvement of 18% and 28% respectively.
Authors: Bofei Gao, Feifan Song, Yibo Miao, Zefan Cai, Zhe Yang, Liang Chen, Helan Hu, Runxin Xu, Qingxiu Dong, Ce Zheng, Shanghaoran Quan, Wen Xiao, Ge Zhang, Daoguang Zan, Keming Lu, Bowen Yu, Dayiheng Liu, Zeyu Cui, Jian Yang, Lei Sha, Houfeng Wang, Zhifang Sui, Peiyi Wang, Tianyu Liu, Baobao Chang
Abstract: Large Language Models (LLMs) exhibit remarkably powerful capabilities. One of the crucial factors to achieve success is aligning the LLM's output with human preferences. This alignment process often requires only a small amount of data to efficiently enhance the LLM's performance. While effective, research in this area spans multiple domains, and the methods involved are relatively complex to understand. The relationships between different methods have been under-explored, limiting the development of the preference alignment. In light of this, we break down the existing popular alignment strategies into different components and provide a unified framework to study the current alignment strategies, thereby establishing connections among them. In this survey, we decompose all the strategies in preference learning into four components: model, data, feedback, and algorithm. This unified view offers an in-depth understanding of existing alignment algorithms and also opens up possibilities to synergize the strengths of different strategies. Furthermore, we present detailed working examples of prevalent existing algorithms to facilitate a comprehensive understanding for the readers. Finally, based on our unified perspective, we explore the challenges and future research directions for aligning large language models with human preferences.
Authors: Michiko Yoshitake (OMRON SINIC X), Yuta Suzuki (Osaka Univ), Ryo Igarashi (OMRON SINIC X), Yoshitaka Ushiku (OMRON SINIC X), Keisuke Nagato (Univ. Tokyo)
Abstract: A college-level benchmark dataset for large language models (LLMs) in the materials science field, MaterialBENCH, is constructed. This dataset consists of problem-answer pairs, based on university textbooks. There are two types of problems: one is the free-response answer type, and the other is the multiple-choice type. Multiple-choice problems are constructed by adding three incorrect answers as choices to a correct answer, so that LLMs can choose one of the four as a response. Most of the problems for free-response answer and multiple-choice types overlap except for the format of the answers. We also conduct experiments using the MaterialBENCH on LLMs, including ChatGPT-3.5, ChatGPT-4, Bard (at the time of the experiments), and GPT-3.5 and GPT-4 with the OpenAI API. The differences and similarities in the performance of LLMs measured by the MaterialBENCH are analyzed and discussed. Performance differences between the free-response type and multiple-choice type in the same models and the influence of using system massages on multiple-choice problems are also studied. We anticipate that MaterialBENCH will encourage further developments of LLMs in reasoning abilities to solve more complicated problems and eventually contribute to materials research and discovery.
Authors: Amit Ben-Artzy, Roy Schwartz
Abstract: In decoder-based LLMs, the representation of a given layer serves two purposes: as input to the next layer during the computation of the current token; and as input to the attention mechanism of future tokens. In this work, we show that the importance of the latter role might be overestimated. To show that, we start by manipulating the representations of previous tokens; e.g. by replacing the hidden states at some layer k with random vectors. Our experimenting with four LLMs and four tasks show that this operation often leads to small to negligible drop in performance. Importantly, this happens if the manipulation occurs in the top part of the model-k is in the final 30-50% of the layers. In contrast, doing the same manipulation in earlier layers might lead to chance level performance. We continue by switching the hidden state of certain tokens with hidden states of other tokens from another prompt; e.g., replacing the word "Italy" with "France" in "What is the capital of Italy?". We find that when applying this switch in the top 1/3 of the model, the model ignores it (answering "Rome"). However if we apply it before, the model conforms to the switch ("Paris"). Our results hint at a two stage process in transformer-based LLMs: the first part gathers input from previous tokens, while the second mainly processes that information internally.
Authors: Larissa Pusch, Tim O. F. Conrad
Abstract: Advancements in natural language processing have revolutionized the way we can interact with digital information systems, such as databases, making them more accessible. However, challenges persist, especially when accuracy is critical, as in the biomedical domain. A key issue is the hallucination problem, where models generate information unsupported by the underlying data, potentially leading to dangerous misinformation. This paper presents a novel approach designed to bridge this gap by combining Large Language Models (LLM) and Knowledge Graphs (KG) to improve the accuracy and reliability of question-answering systems, on the example of a biomedical KG. Built on the LangChain framework, our method incorporates a query checker that ensures the syntactical and semantic validity of LLM-generated queries, which are then used to extract information from a Knowledge Graph, substantially reducing errors like hallucinations. We evaluated the overall performance using a new benchmark dataset of 50 biomedical questions, testing several LLMs, including GPT-4 Turbo and llama3:70b. Our results indicate that while GPT-4 Turbo outperforms other models in generating accurate queries, open-source models like llama3:70b show promise with appropriate prompt engineering. To make this approach accessible, a user-friendly web-based interface has been developed, allowing users to input natural language queries, view generated and corrected Cypher queries, and verify the resulting paths for accuracy. Overall, this hybrid approach effectively addresses common issues such as data gaps and hallucinations, offering a reliable and intuitive solution for question answering systems. The source code for generating the results of this paper and for the user-interface can be found in our Git repository: https://git.zib.de/lpusch/cyphergenkg-gui
Authors: Neeladri Bhuiya, Viktor Schlegel, Stefan Winkler
Abstract: State-of-the-art Large Language Models (LLMs) are accredited with an increasing number of different capabilities, ranging from reading comprehension, over advanced mathematical and reasoning skills to possessing scientific knowledge. In this paper we focus on their multi-hop reasoning capability: the ability to identify and integrate information from multiple textual sources. Given the concerns with the presence of simplifying cues in existing multi-hop reasoning benchmarks, which allow models to circumvent the reasoning requirement, we set out to investigate, whether LLMs are prone to exploiting such simplifying cues. We find evidence that they indeed circumvent the requirement to perform multi-hop reasoning, but they do so in more subtle ways than what was reported about their fine-tuned pre-trained language model (PLM) predecessors. Motivated by this finding, we propose a challenging multi-hop reasoning benchmark, by generating seemingly plausible multi-hop reasoning chains, which ultimately lead to incorrect answers. We evaluate multiple open and proprietary state-of-the-art LLMs, and find that their performance to perform multi-hop reasoning is affected, as indicated by up to 45% relative decrease in F1 score when presented with such seemingly plausible alternatives. We conduct a deeper analysis and find evidence that while LLMs tend to ignore misleading lexical cues, misleading reasoning paths indeed present a significant challenge.
Authors: Xichen Tang
Abstract: Using computerized verifiable formal languages like Lean 4 to prove mathematical theorems has a significant impact on mathematical formalization. Lean 4 offers prominent potential for advancing mathematical reasoning. However, existing efforts are limited to mathematical formalization languages in substantial online corpora and are dedicated to keeping pace with rapidly evolving languages. To bridge the gap between the traditional and computerized proof, my approach to formalizing theorem proving involves generating formal steps and complete proofs using Large Language Models (LLMs) based on Natural Language (NL) proofs. The method is to introduce the basic structure and tactics in general, determine how AI can assist the mathematical formalization process to improve its performance, and give examples of solving problems in Lean 4 comparing to NL, mainly in IMO, and a sample theorem proving in abstract algebra.
Authors: Yu Zhang, Songlin Yang, Ruijie Zhu, Yue Zhang, Leyang Cui, Yiqiao Wang, Bolun Wang, Freda Shi, Bailin Wang, Wei Bi, Peng Zhou, Guohong Fu
Abstract: Linear attention Transformers and their gated variants, celebrated for enabling parallel training and efficient recurrent inference, still fall short in recall-intensive tasks compared to traditional Transformers and demand significant resources for training from scratch. This paper introduces Gated Slot Attention (GSA), which enhances Attention with Bounded-memory-Control (ABC) by incorporating a gating mechanism inspired by Gated Linear Attention (GLA). Essentially, GSA comprises a two-layer GLA linked via $\operatorname{softmax}$, utilizing context-aware memory reading and adaptive forgetting to improve memory capacity while maintaining compact recurrent state size. This design greatly enhances both training and inference efficiency through GLA's hardware-efficient training algorithm and reduced state size. Additionally, retaining the $\operatorname{softmax}$ operation is particularly beneficial in "finetuning pretrained Transformers to RNNs" (T2R) settings, reducing the need for extensive training from scratch. Extensive experiments confirm GSA's superior performance in scenarios requiring in-context recall and in T2R settings.
Authors: Navya Jain, Zekun Wu, Cristian Munoz, Airlie Hilliard, Adriano Koshiyama, Emre Kazim, Philip Treleaven
Abstract: As the demand for human-like interactions with LLMs continues to grow, so does the interest in manipulating their personality traits, which has emerged as a key area of research. Methods like prompt-based In-Context Knowledge Editing (IKE) and gradient-based Model Editor Networks (MEND) have been explored but show irregularity and variability. IKE depends on the prompt, leading to variability and sensitivity, while MEND yields inconsistent and gibberish outputs. To address this, we employed Opinion QA Based Parameter-Efficient Fine-Tuning (PEFT), specifically Quantized Low-Rank Adaptation (QLoRA), to manipulate the Big Five personality traits: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. After PEFT, models such as Mistral-7B-Instruct and Llama-2-7B-chat began generating emojis, despite their absence in the PEFT data. For instance, Llama-2-7B-chat generated emojis in 99.5\% of extraversion-related test instances, while Mistral-7B-Instruct did so in 92.5\% of openness-related test instances. Explainability analysis indicated that the LLMs used emojis intentionally to express these traits. This paper provides a number of novel contributions. First, introducing an Opinion QA dataset for PEFT-driven personality manipulation; second, developing metric models to benchmark LLM personality traits; third, demonstrating PEFT's superiority over IKE in personality manipulation; and finally, analysing and validating emoji usage through explainability methods such as mechanistic interpretability and in-context learning explainability methods.
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: Xiaosong Yuan, Chen Shen, Shaotian Yan, Xiaofeng Zhang, Liang Xie, Wenxiao Wang, Renchu Guan, Ying Wang, Jieping Ye
Abstract: Zero-shot Chain-of-Thought (CoT) prompting emerges as a simple and effective strategy for enhancing the performance of large language models (LLMs) in real-world reasoning tasks. Nonetheless, the efficacy of a singular, task-level prompt uniformly applied across the whole of instances is inherently limited since one prompt cannot be a good partner for all, a more appropriate approach should consider the interaction between the prompt and each instance meticulously. This work introduces an instance-adaptive prompting algorithm as an alternative zero-shot CoT reasoning scheme by adaptively differentiating good and bad prompts. Concretely, we first employ analysis on LLMs through the lens of information flow to detect the mechanism under zero-shot CoT reasoning, in which we discover that information flows from question to prompt and question to rationale jointly influence the reasoning results most. We notice that a better zero-shot CoT reasoning needs the prompt to obtain semantic information from the question then the rationale aggregates sufficient information from the question directly and via the prompt indirectly. On the contrary, lacking any of those would probably lead to a bad one. Stem from that, we further propose an instance-adaptive prompting strategy (IAP) for zero-shot CoT reasoning. Experiments conducted with LLaMA-2, LLaMA-3, and Qwen on math, logic, and commonsense reasoning tasks (e.g., GSM8K, MMLU, Causal Judgement) obtain consistent improvement, demonstrating that the instance-adaptive zero-shot CoT prompting performs better than other task-level methods with some curated prompts or sophisticated procedures, showing the significance of our findings in the zero-shot CoT reasoning mechanism.
Authors: Mario Giulianelli, Luca Malagutti, Juan Luis Gastaldi, Brian DuSell, Tim Vieira, Ryan Cotterell
Abstract: Language models are widely used in computational psycholinguistics to test theories that relate the negative log probability (the surprisal) of a region of interest (a substring of characters) under a language model to its cognitive cost experienced by readers, as operationalized, for example, by gaze duration on the region. However, the application of modern language models to psycholinguistic studies is complicated by the practice of using tokenization as an intermediate step in training a model. Doing so results in a language model over token strings rather than one over character strings. Vexingly, regions of interest are generally misaligned with these token strings. The paper argues that token-level language models should be (approximately) marginalized into character-level language models before they are used in psycholinguistic studies to compute the surprisal of a region of interest; then, the marginalized character-level language model can be used to compute the surprisal of an arbitrary character substring, which we term a focal area, that the experimenter may wish to use as a predictor. Our proposal of marginalizing a token-level model into a character-level one solves this misalignment issue independently of the tokenization scheme. Empirically, we discover various focal areas whose surprisal is a better psychometric predictor than the surprisal of the region of interest itself.
Authors: Gabriel Roccabruna, Massimo Rizzoli, Giuseppe Riccardi
Abstract: The automatic detection of temporal relations among events has been mainly investigated with encoder-only models such as RoBERTa. Large Language Models (LLM) have recently shown promising performance in temporal reasoning tasks such as temporal question answering. Nevertheless, recent studies have tested the LLMs' performance in detecting temporal relations of closed-source models only, limiting the interpretability of those results. In this work, we investigate LLMs' performance and decision process in the Temporal Relation Classification task. First, we assess the performance of seven open and closed-sourced LLMs experimenting with in-context learning and lightweight fine-tuning approaches. Results show that LLMs with in-context learning significantly underperform smaller encoder-only models based on RoBERTa. Then, we delve into the possible reasons for this gap by applying explainable methods. The outcome suggests a limitation of LLMs in this task due to their autoregressive nature, which causes them to focus only on the last part of the sequence. Additionally, we evaluate the word embeddings of these two models to better understand their pre-training differences. The code and the fine-tuned models can be found respectively on GitHub.
Authors: Denitsa Saynova, Lovisa Hagstr\"om, Moa Johansson, Richard Johansson, Marco Kuhlmann
Abstract: Previous interpretations of language models (LMs) miss important distinctions in how these models process factual information. For example, given the query "Astrid Lindgren was born in" with the corresponding completion "Sweden", no difference is made between whether the prediction was based on having the exact knowledge of the birthplace of the Swedish author or assuming that a person with a Swedish-sounding name was born in Sweden. In this paper, we investigate four different prediction scenarios for which the LM can be expected to show distinct behaviors. These scenarios correspond to different levels of model reliability and types of information being processed - some being less desirable for factual predictions. To facilitate precise interpretations of LMs for fact completion, we propose a model-specific recipe called PrISM for constructing datasets with examples of each scenario based on a set of diagnostic criteria. We apply a popular interpretability method, causal tracing (CT), to the four prediction scenarios and find that while CT produces different results for each scenario, aggregations over a set of mixed examples may only represent the results from the scenario with the strongest measured signal. In summary, we contribute tools for a more granular study of fact completion in language models and analyses that provide a more nuanced understanding of how LMs process fact-related queries.
Authors: Melanie Walsh, Anna Preus, Elizabeth Gronski
Abstract: Generating poetry has become a popular application of LLMs, perhaps especially of OpenAI's widely-used chatbot ChatGPT. What kind of poet is ChatGPT? Does ChatGPT have its own poetic style? Can it successfully produce poems in different styles? To answer these questions, we prompt the GPT-3.5 and GPT-4 models to generate English-language poems in 24 different poetic forms and styles, about 40 different subjects, and in response to 3 different writing prompt templates. We then analyze the resulting 5.7k poems, comparing them to a sample of 3.7k poems from the Poetry Foundation and the Academy of American Poets. We find that the GPT models, especially GPT-4, can successfully produce poems in a range of both common and uncommon English-language forms in superficial yet noteworthy ways, such as by producing poems of appropriate lengths for sonnets (14 lines), villanelles (19 lines), and sestinas (39 lines). But the GPT models also exhibit their own distinct stylistic tendencies, both within and outside of these specific forms. Our results show that GPT poetry is much more constrained and uniform than human poetry, showing a strong penchant for rhyme, quatrains (4-line stanzas), iambic meter, first-person plural perspectives (we, us, our), and specific vocabulary like "heart," "embrace," "echo," and "whisper."
Authors: Nima Meghdadi, Diana Inkpen
Abstract: This paper presents the methods used for LegalLens-2024 shared task, which focused on detecting legal violations within unstructured textual data and associating these violations with potentially affected individuals. The shared task included two subtasks: A) Legal Named Entity Recognition (L-NER) and B) Legal Natural Language Inference (L-NLI). For subtask A, we utilized the spaCy library, while for subtask B, we employed a combined model incorporating RoBERTa and CNN. Our results were 86.3% in the L-NER subtask and 88.25% in the L-NLI subtask. Overall, our paper demonstrates the effectiveness of transformer models in addressing complex tasks in the legal domain. The source code for our implementation is publicly available at https://github.com/NimaMeghdadi/uOttawa-at-LegalLens-2024-Transformer-based-Classification
URLs: https://github.com/NimaMeghdadi/uOttawa-at-LegalLens-2024-Transformer-based-Classification
Authors: Ge Yang, Changyi He, Jinyang Guo, Jianyu Wu, Yifu Ding, Aishan Liu, Haotong Qin, Pengliang Ji, Xianglong Liu
Abstract: Although large language models (LLMs) have demonstrated their strong intelligence ability, the high demand for computation and storage hinders their practical application. To this end, many model compression techniques are proposed to increase the efficiency of LLMs. However, current researches only validate their methods on limited models, datasets, metrics, etc, and still lack a comprehensive evaluation under more general scenarios. So it is still a question of which model compression approach we should use under a specific case. To mitigate this gap, we present the Large Language Model Compression Benchmark (LLMCBench), a rigorously designed benchmark with an in-depth analysis for LLM compression algorithms. We first analyze the actual model production requirements and carefully design evaluation tracks and metrics. Then, we conduct extensive experiments and comparison using multiple mainstream LLM compression approaches. Finally, we perform an in-depth analysis based on the evaluation and provide useful insight for LLM compression design. We hope our LLMCBench can contribute insightful suggestions for LLM compression algorithm design and serve as a foundation for future research. Our code is available at https://github.com/AboveParadise/LLMCBench.
Authors: Yuxu Wu, Yiren Xing
Abstract: With the rapid development of Natural Language Processing (NLP) technology, the accuracy and efficiency of machine translation have become hot topics of research. This paper proposes a novel Seq2Seq model aimed at improving translation quality while reducing the storage space required by the model. The model employs a Bidirectional Long Short-Term Memory network (Bi-LSTM) as the encoder to capture the context information of the input sequence; the decoder incorporates an attention mechanism, enhancing the model's ability to focus on key information during the translation process. Compared to the current mainstream Transformer model, our model achieves superior performance on the WMT14 machine translation dataset while maintaining a smaller size. The study first introduces the design principles and innovative points of the model architecture, followed by a series of experiments to verify the effectiveness of the model. The experimental includes an assessment of the model's performance on different language pairs, as well as comparative analysis with traditional Seq2Seq models. The results show that while maintaining translation accuracy, our model significantly reduces the storage requirements, which is of great significance for translation applications in resource-constrained scenarios. our code are available at https://github.com/mindspore-lab/models/tree/master/research/arxiv_papers/miniformer. Thanks for the support provided by MindSpore Community.
URLs: https://github.com/mindspore-lab/models/tree/master/research/arxiv_papers/miniformer.
Authors: Zehan Qi, Rongwu Xu, Zhijiang Guo, Cunxiang Wang, Hao Zhang, Wei Xu
Abstract: Retrieval-augmented generation (RAG) is a promising approach to address the limitations of fixed knowledge in large language models (LLMs). However, current benchmarks for evaluating RAG systems suffer from two key deficiencies: (1) they fail to adequately measure LLMs' capability in handling long-context retrieval due to a lack of datasets that reflect the characteristics of retrieved documents, and (2) they lack a comprehensive evaluation method for assessing LLMs' ability to generate long-form responses that effectively exploits retrieved information. To address these shortcomings, we introduce the Long$^2$RAG benchmark and the Key Point Recall (KPR) metric. Long$^2$RAG comprises 280 questions spanning 10 domains and across 8 question categories, each associated with 5 retrieved documents with an average length of 2,444 words. KPR evaluates the extent to which LLMs incorporate key points extracted from the retrieved documents into their generated responses, providing a more nuanced assessment of their ability to exploit retrieved information.
Authors: Anay Mehrotra, Manolis Zampetakis, Paul Kassianik, Blaine Nelson, Hyrum Anderson, Yaron Singer, Amin Karbasi
Abstract: While Large Language Models (LLMs) display versatile functionality, they continue to generate harmful, biased, and toxic content, as demonstrated by the prevalence of human-designed jailbreaks. In this work, we present Tree of Attacks with Pruning (TAP), an automated method for generating jailbreaks that only requires black-box access to the target LLM. TAP utilizes an attacker LLM to iteratively refine candidate (attack) prompts until one of the refined prompts jailbreaks the target. In addition, before sending prompts to the target, TAP assesses them and prunes the ones unlikely to result in jailbreaks, reducing the number of queries sent to the target LLM. In empirical evaluations, we observe that TAP generates prompts that jailbreak state-of-the-art LLMs (including GPT4-Turbo and GPT4o) for more than 80% of the prompts. This significantly improves upon the previous state-of-the-art black-box methods for generating jailbreaks while using a smaller number of queries than them. Furthermore, TAP is also capable of jailbreaking LLMs protected by state-of-the-art guardrails, e.g., LlamaGuard.
Authors: Xu Yang, Yingzhe Peng, Haoxuan Ma, Shuo Xu, Chi Zhang, Yucheng Han, Hanwang Zhang
Abstract: As Archimedes famously said, ``Give me a lever long enough and a fulcrum on which to place it, and I shall move the world'', in this study, we propose to use a tiny Language Model (LM), \eg, a Transformer with 67M parameters, to lever much larger Vision-Language Models (LVLMs) with 9B parameters. Specifically, we use this tiny \textbf{Lever-LM} to configure effective in-context demonstration (ICD) sequences to improve the In-Context Learinng (ICL) performance of LVLMs. Previous studies show that diverse ICD configurations like the selection and ordering of the demonstrations heavily affect the ICL performance, highlighting the significance of configuring effective ICD sequences. Motivated by this and by re-considering the the process of configuring ICD sequence, we find this is a mirror process of human sentence composition and further assume that effective ICD configurations may contain internal statistical patterns that can be captured by Lever-LM. Then a dataset with effective ICD sequences is constructed to train Lever-LM. After training, given novel queries, new ICD sequences are configured by the trained Lever-LM to solve vision-language tasks through ICL. Experiments show that these ICD sequences can improve the ICL performance of two LVLMs compared with some strong baselines in Visual Question Answering and Image Captioning, validating that Lever-LM can really capture the statistical patterns for levering LVLMs. The code is available at \url{https://github.com/ForJadeForest/Lever-LM}.
Authors: Adib Hasan, Ileana Rugina, Alex Wang
Abstract: This paper investigates the impact of model compression on the way Large Language Models (LLMs) process prompts, particularly concerning jailbreak resistance. We show that moderate WANDA pruning can enhance resistance to jailbreaking attacks without fine-tuning, while maintaining performance on standard benchmarks. To systematically evaluate this safety enhancement, we introduce a dataset of 225 harmful tasks across five categories. Our analysis of LLaMA-2 Chat, Vicuna 1.3, and Mistral Instruct v0.2 reveals that pruning benefits correlate with initial model safety levels. We interpret these results by examining changes in attention patterns and perplexity shifts, demonstrating that pruned models exhibit sharper attention and increased sensitivity to artificial jailbreak constructs. We extend our evaluation to the AdvBench harmful behavior tasks and the GCG attack method. We find that LLaMA-2 is much safer on AdvBench prompts than on our dataset when evaluated with manual jailbreak attempts, and that pruning is effective against both automated attacks and manual jailbreaking on Advbench.
Authors: David Grangier, Angelos Katharopoulos, Pierre Ablin, Awni Hannun
Abstract: Large language models are versatile tools but are not suitable for small inference budgets. Small models have more efficient inference, but their lower capacity means that their performance can be good only if one limits their scope to a specialized domain. This paper explores how to get good specialized small language models using a large, generic, pretraining set and a limited amount of specialized data. We consider two scenarios, depending on whether (i) one can afford pretraining a model for each specialization task, or (ii) one wants to cheaply adapt a single pretrained model for each task. In the first scenario, we propose an effective solution based on importance sampling: we resample the pretraining set to imitate the specialization data and train a small model on it. In the second scenario, we propose a novel architecture, projected networks (PN). PN is a large network whose parameters can be linearly projected into a small network for specialization. For both scenarios, we demonstrate the empirical effectiveness of our solutions across various domains, training set sizes, and training budgets.
Authors: Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long
Abstract: Foundation models of time series have not been fully developed due to the limited availability of time series corpora and the underexploration of scalable pre-training. Based on the similar sequential formulation of time series and natural language, increasing research demonstrates the feasibility of leveraging large language models (LLM) for time series. Nevertheless, the inherent autoregressive property and decoder-only architecture of LLMs have not been fully considered, resulting in insufficient utilization of LLM abilities. To fully revitalize the general-purpose token transition and multi-step generation capability of large language models, we propose AutoTimes to repurpose LLMs as autoregressive time series forecasters, which projects time series into the embedding space of language tokens and autoregressively generates future predictions with arbitrary lengths. Compatible with any decoder-only LLMs, the consequent forecaster exhibits the flexibility of the lookback length and scalability with larger LLMs. Further, we formulate time series as prompts, extending the context for prediction beyond the lookback window, termed in-context forecasting. By introducing LLM-embedded textual timestamps, AutoTimes can utilize chronological information to align multivariate time series. Empirically, AutoTimes achieves state-of-the-art with 0.1% trainable parameters and over $5\times$ training/inference speedup compared to advanced LLM-based forecasters. Code is available at this repository: https://github.com/thuml/AutoTimes.
Authors: Chengxing Xie, Canyu Chen, Feiran Jia, Ziyu Ye, Shiyang Lai, Kai Shu, Jindong Gu, Adel Bibi, Ziniu Hu, David Jurgens, James Evans, Philip Torr, Bernard Ghanem, Guohao Li
Abstract: Large Language Model (LLM) agents have been increasingly adopted as simulation tools to model humans in social science and role-playing applications. However, one fundamental question remains: can LLM agents really simulate human behavior? In this paper, we focus on one elemental behavior in human interactions, trust, and investigate whether LLM agents can simulate human trust behavior. We first find that LLM agents generally exhibit trust behavior, referred to as agent trust, under the framework of Trust Games, which are widely recognized in behavioral economics. Then, we discover that GPT-4 agents manifest high behavioral alignment with humans in terms of trust behavior, indicating the feasibility of simulating human trust behavior with LLM agents. In addition, we probe the biases of agent trust and differences in agent trust towards other LLM agents and humans. We also explore the intrinsic properties of agent trust under conditions including external manipulations and advanced reasoning strategies. Our study provides new insights into the behaviors of LLM agents and the fundamental analogy between LLMs and humans beyond value alignment. We further illustrate broader implications of our discoveries for applications where trust is paramount.
Authors: Yichuan Mo, Yuji Wang, Zeming Wei, Yisen Wang
Abstract: While Large Language Models (LLMs) have achieved tremendous success in various applications, they are also susceptible to jailbreaking attacks. Several primary defense strategies have been proposed to protect LLMs from producing harmful information, mostly focusing on model fine-tuning or heuristical defense designs. However, how to achieve intrinsic robustness through prompt optimization remains an open problem. In this paper, motivated by adversarial training paradigms for achieving reliable robustness, we propose an approach named Prompt Adversarial Tuning (PAT) that trains a prompt control attached to the user prompt as a guard prefix. To achieve our defense goal whilst maintaining natural performance, we optimize the control prompt with both adversarial and benign prompts. Comprehensive experiments show that our method is effective against both grey-box and black-box attacks, reducing the success rate of advanced attacks to nearly 0%, while maintaining the model's utility on the benign task and incurring only negligible computational overhead, charting a new perspective for future explorations in LLM security. Our code is available at https://github.com/PKU-ML/PAT.
Authors: Yanbang Wang, Hejie Cui, Jon Kleinberg
Abstract: Graphs data is crucial for many applications, and much of it exists in the relations described in textual format. As a result, being able to accurately recall and encode a graph described in earlier text is a basic yet pivotal ability that LLMs need to demonstrate if they are to perform reasoning tasks that involve graph-structured information. Human performance at graph recall has been studied by cognitive scientists for decades, and has been found to often exhibit certain structural patterns of bias that align with human handling of social relationships. To date, however, we know little about how LLMs behave in analogous graph recall tasks: do their recalled graphs also exhibit certain biased patterns, and if so, how do they compare with humans and affect other graph reasoning tasks? In this work, we perform the first systematical study of graph recall by LLMs, investigating the accuracy and biased microstructures (local structural patterns) in their recall. We find that LLMs not only underperform often in graph recall, but also tend to favor more triangles and alternating 2-paths. Moreover, we find that more advanced LLMs have a striking dependence on the domain that a real-world graph comes from -- by yielding the best recall accuracy when the graph is narrated in a language style consistent with its original domain.
Authors: Christos Thrampoulidis
Abstract: We initiate an investigation into the optimization properties of next-token prediction (NTP), the dominant training paradigm for modern language models. Specifically, we study the structural properties of the solutions selected by gradient-based optimizers among the many possible minimizers of the NTP objective. By framing NTP as cross-entropy minimization across distinct contexts, each tied with a sparse conditional probability distribution across a finite vocabulary of tokens, we introduce "NTP-separability conditions" that enable reaching the data-entropy lower bound. With this setup, and focusing on linear models with fixed context embeddings, we characterize the optimization bias of gradient descent (GD): Within the data subspace defined by the sparsity patterns of distinct contexts, GD selects parameters that equate the logits' differences of in-support tokens to their log-odds. In the orthogonal subspace, the GD parameters diverge in norm and select the direction that maximizes a margin specific to NTP. These findings extend previous research on implicit bias in one-hot classification to the NTP setting, highlighting key differences and prompting further research into the optimization and generalization properties of NTP, irrespective of the specific architecture used to generate the context embeddings.
Authors: Ziyi Chen, Mengyuan Zhang, Mustafa Mohammed Ahmed, Yi Guo, Thomas J. George, Jiang Bian, Yonghui Wu
Abstract: Cancer treatments are known to introduce cardiotoxicity, negatively impacting outcomes and survivorship. Identifying cancer patients at risk of heart failure (HF) is critical to improving cancer treatment outcomes and safety. This study examined machine learning (ML) models to identify cancer patients at risk of HF using electronic health records (EHRs), including traditional ML, Time-Aware long short-term memory (T-LSTM), and large language models (LLMs) using novel narrative features derived from the structured medical codes. We identified a cancer cohort of 12,806 patients from the University of Florida Health, diagnosed with lung, breast, and colorectal cancers, among which 1,602 individuals developed HF after cancer. The LLM, GatorTron-3.9B, achieved the best F1 scores, outperforming the traditional support vector machines by 0.397, the T-LSTM deep learning model by 0.077, and a widely used transformer model, BERT, by 0.056. The analysis shows that the proposed narrative features remarkably increased feature density and improved performance.
Authors: Michael Saxon, Fatima Jahara, Mahsa Khoshnoodi, Yujie Lu, Aditya Sharma, William Yang Wang
Abstract: With advances in the quality of text-to-image (T2I) models has come interest in benchmarking their prompt faithfulness -- the semantic coherence of generated images to the prompts they were conditioned on. A variety of T2I faithfulness metrics have been proposed, leveraging advances in cross-modal embeddings and vision-language models (VLMs). However, these metrics are not rigorously compared and benchmarked, instead presented with correlation to human Likert scores over a set of easy-to-discriminate images against seemingly weak baselines. We introduce T2IScoreScore, a curated set of semantic error graphs containing a prompt and a set of increasingly erroneous images. These allow us to rigorously judge whether a given prompt faithfulness metric can correctly order images with respect to their objective error count and significantly discriminate between different error nodes, using meta-metric scores derived from established statistical tests. Surprisingly, we find that the state-of-the-art VLM-based metrics (e.g., TIFA, DSG, LLMScore, VIEScore) we tested fail to significantly outperform simple (and supposedly worse) feature-based metrics like CLIPScore, particularly on a hard subset of naturally-occurring T2I model errors. TS2 will enable the development of better T2I prompt faithfulness metrics through more rigorous comparison of their conformity to expected orderings and separations under objective criteria.
Authors: Angelica Chen, Sadhika Malladi, Lily H. Zhang, Xinyi Chen, Qiuyi Zhang, Rajesh Ranganath, Kyunghyun Cho
Abstract: Preference learning algorithms (e.g., RLHF and DPO) are frequently used to steer LLMs to produce generations that are more preferred by humans, but our understanding of their inner workings is still limited. In this work, we study the conventional wisdom that preference learning trains models to assign higher likelihoods to more preferred outputs than less preferred outputs, measured via ranking accuracy. Surprisingly, we find that most state-of-the-art preference-tuned models achieve a ranking accuracy of less than 60% on common preference datasets. We furthermore derive the idealized ranking accuracy that a preference-tuned LLM would achieve if it optimized the DPO or RLHF objective perfectly. We demonstrate that existing models exhibit a significant alignment gap -- i.e., a gap between the observed and idealized ranking accuracies. We attribute this discrepancy to the DPO objective, which is empirically and theoretically ill-suited to fix even mild ranking errors in the reference model, and derive a simple and efficient formula for quantifying the difficulty of learning a given preference datapoint. Finally, we demonstrate that ranking accuracy strongly correlates with the empirically popular win rate metric when the model is close to the reference model used in the objective, shedding further light on the differences between on-policy (e.g., RLHF) and off-policy (e.g., DPO) preference learning algorithms.
Authors: Shirley Wu, Shiyu Zhao, Qian Huang, Kexin Huang, Michihiro Yasunaga, Kaidi Cao, Vassilis N. Ioannidis, Karthik Subbian, Jure Leskovec, James Zou
Abstract: Large language model (LLM) agents have demonstrated impressive capabilities in utilizing external tools and knowledge to boost accuracy and reduce hallucinations. However, developing prompting techniques that enable LLM agents to effectively use these tools and knowledge remains a heuristic and labor-intensive task. Here, we introduce AvaTaR, a novel and automated framework that optimizes an LLM agent to effectively leverage provided tools, improving performance on a given task. During optimization, we design a comparator module to iteratively deliver insightful and comprehensive prompts to the LLM agent by contrastively reasoning between positive and negative examples sampled from training data. We demonstrate AvaTaR on four complex multimodal retrieval datasets featuring textual, visual, and relational information, and three general question-answering (QA) datasets. We find AvaTaR consistently outperforms state-of-the-art approaches across all seven tasks, exhibiting strong generalization ability when applied to novel cases and achieving an average relative improvement of 14% on the Hit@1 metric for the retrieval datasets and 13% for the QA datasets. Code and dataset are available at https://github.com/zou-group/avatar.
Authors: Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, Neel Nanda
Abstract: Conversational large language models are fine-tuned for both instruction-following and safety, resulting in models that obey benign requests but refuse harmful ones. While this refusal behavior is widespread across chat models, its underlying mechanisms remain poorly understood. In this work, we show that refusal is mediated by a one-dimensional subspace, across 13 popular open-source chat models up to 72B parameters in size. Specifically, for each model, we find a single direction such that erasing this direction from the model's residual stream activations prevents it from refusing harmful instructions, while adding this direction elicits refusal on even harmless instructions. Leveraging this insight, we propose a novel white-box jailbreak method that surgically disables refusal with minimal effect on other capabilities. Finally, we mechanistically analyze how adversarial suffixes suppress propagation of the refusal-mediating direction. Our findings underscore the brittleness of current safety fine-tuning methods. More broadly, our work showcases how an understanding of model internals can be leveraged to develop practical methods for controlling model behavior.
Authors: Rachel S. Y. Teo, Tan M. Nguyen
Abstract: The remarkable success of transformers in sequence modeling tasks, spanning various applications in natural language processing and computer vision, is attributed to the critical role of self-attention. Similar to the development of most deep learning models, the construction of these attention mechanisms relies on heuristics and experience. In our work, we derive self-attention from kernel principal component analysis (kernel PCA) and show that self-attention projects its query vectors onto the principal component axes of its key matrix in a feature space. We then formulate the exact formula for the value matrix in self-attention, theoretically and empirically demonstrating that this value matrix captures the eigenvectors of the Gram matrix of the key vectors in self-attention. Leveraging our kernel PCA framework, we propose Attention with Robust Principal Components (RPC-Attention), a novel class of robust attention that is resilient to data contamination. We empirically demonstrate the advantages of RPC-Attention over softmax attention on the ImageNet-1K object classification, WikiText-103 language modeling, and ADE20K image segmentation task.
Authors: Tianyi Qiu, Yang Zhang, Xuchuan Huang, Jasmine Xinze Li, Jiaming Ji, Yaodong Yang
Abstract: Frontier AI systems, including large language models (LLMs), hold increasing influence over the epistemology of human users. Such influence can reinforce prevailing societal values, potentially contributing to the lock-in of misguided moral beliefs and, consequently, the perpetuation of problematic moral practices on a broad scale. We introduce progress alignment as a technical solution to mitigate this imminent risk. Progress alignment algorithms learn to emulate the mechanics of human moral progress, thereby addressing the susceptibility of existing alignment methods to contemporary moral blindspots. To empower research in progress alignment, we introduce ProgressGym, an experimental framework allowing the learning of moral progress mechanics from history, in order to facilitate future progress in real-world moral decisions. Leveraging 9 centuries of historical text and 18 historical LLMs, ProgressGym enables codification of real-world progress alignment challenges into concrete benchmarks. Specifically, we introduce three core challenges: tracking evolving values (PG-Follow), preemptively anticipating moral progress (PG-Predict), and regulating the feedback loop between human and AI value shifts (PG-Coevolve). Alignment methods without a temporal dimension are inapplicable to these tasks. In response, we present lifelong and extrapolative algorithms as baseline methods of progress alignment, and build an open leaderboard soliciting novel algorithms and challenges. The framework and the leaderboard are available at https://github.com/PKU-Alignment/ProgressGym and https://huggingface.co/spaces/PKU-Alignment/ProgressGym-LeaderBoard respectively.
URLs: https://github.com/PKU-Alignment/ProgressGym, https://huggingface.co/spaces/PKU-Alignment/ProgressGym-LeaderBoard
Authors: Yu Bai, Yukai Miao, Li Chen, Dawei Wang, Dan Li, Yanyu Ren, Hongtao Xie, Ce Yang, Xuhui Cai
Abstract: RAG systems face limitations when semantic relevance alone does not guarantee improved generation quality. This issue becomes particularly evident due to the sensitivity of large language models (LLMs) to the ordering of few-shot prompts, which can affect model performance. To address this challenge, aligning LLM outputs with human preferences using structured feedback, such as options to copy, regenerate, or dislike, offers a promising method for improvement. This feedback is applied to the entire list of inputs rather than giving specific ratings for individual documents, making it a Listwide Labels Learning-to-Rank task. To address this task, we propose Pistis-RAG, a new RAG framework designed with a content-centric approach to better align LLMs with human preferences. Pistis-RAG effectively utilizes human feedback, enhancing content ranking and generation quality. To validate our framework, we use public datasets to simulate human feedback, allowing us to evaluate and refine our method effectively. Experimental results indicate that Pistis-RAG improves alignment with human preferences relative to the baseline RAG system, showing a 6.06% increase in MMLU (English) and a 7.08% increase in C-EVAL (Chinese) accuracy metrics. These results highlight Pistis-RAG's effectiveness in overcoming the limitations associated with traditional RAG approaches.
Authors: Zihao Wang, Shaofei Cai, Zhancun Mu, Haowei Lin, Ceyao Zhang, Xuejie Liu, Qing Li, Anji Liu, Xiaojian Ma, Yitao Liang
Abstract: This paper presents OmniJARVIS, a novel Vision-Language-Action (VLA) model for open-world instruction-following agents in Minecraft. Compared to prior works that either emit textual goals to separate controllers or produce the control command directly, OmniJARVIS seeks a different path to ensure both strong reasoning and efficient decision-making capabilities via unified tokenization of multimodal interaction data. First, we introduce a self-supervised approach to learn a behavior encoder that produces discretized tokens for behavior trajectories $\tau = \{o_0, a_0, \dots\}$ and an imitation learning policy decoder conditioned on these tokens. These additional behavior tokens will be augmented to the vocabulary of pretrained Multimodal Language Models. With this encoder, we then pack long-term multimodal interactions involving task instructions, memories, thoughts, observations, textual responses, behavior trajectories, etc into unified token sequences and model them with autoregressive transformers. Thanks to the semantically meaningful behavior tokens, the resulting VLA model, OmniJARVIS, can reason (by producing chain-of-thoughts), plan, answer questions, and act (by producing behavior tokens for the imitation learning policy decoder). OmniJARVIS demonstrates excellent performances on a comprehensive collection of atomic, programmatic, and open-ended tasks in open-world Minecraft. Our analysis further unveils the crucial design principles in interaction data formation, unified tokenization, and its scaling potentials. The dataset, models, and code will be released at https://craftjarvis.org/OmniJARVIS.
Authors: Harshit Joshi, Shicheng Liu, James Chen, Robert Weigle, Monica S. Lam
Abstract: Large Language Models (LLMs) present an opportunity to create automated assistants that can help users navigate complex tasks. However, existing approaches have limitations in handling conditional logic, integrating knowledge sources, and consistently following instructions. Researchers and industry professionals often employ ad hoc pipelines to construct conversational agents. These pipelines aim to maintain context, address failure cases, and minimize hallucinations, yet frequently fail to achieve these objectives. To this end, we present Genie - a programmable framework for creating task-oriented conversational agents that are designed to handle complex user interactions and knowledge queries. Unlike LLMs, Genie provides reliable grounded responses, with controllable agent policies through its expressive specification, Genie Worksheet. In contrast to dialog trees, it is resilient to diverse user queries, helpful with knowledge sources, and offers ease of programming policies through its declarative paradigm. The agents built using Genie outperforms the state-of-the-art method on complex logic domains in STARV2 dataset by up to 20.5%. Additionally, through a real-user study involving 62 participants, we show that Genie beats the GPT-4 with function calling baseline by 21.1%, 20.1%, and 61% on execution accuracy, dialogue act accuracy, and goal completion rate, respectively, on three diverse real-world domains
Authors: Nikhil Abhyankar, Vivek Gupta, Dan Roth, Chandan K. Reddy
Abstract: Tabular reasoning involves interpreting natural language queries about tabular data, which presents a unique challenge of combining language understanding with structured data analysis. Existing methods employ either textual reasoning, which excels in semantic interpretation but struggles with mathematical operations, or symbolic reasoning, which handles computations well but lacks semantic understanding. This paper introduces a novel algorithm H-STAR that integrates both symbolic and semantic (textual) approaches in a two-stage process to address these limitations. H-STAR employs: (1) step-wise table extraction using `multi-view' column retrieval followed by row extraction, and (2) adaptive reasoning that adapts reasoning strategies based on question types, utilizing semantic reasoning for direct lookup and complex lexical queries while augmenting textual reasoning with symbolic reasoning support for quantitative and logical tasks. Our extensive experiments demonstrate that H-STAR significantly outperforms state-of-the-art methods across three tabular question-answering (QA) and fact-verification datasets, underscoring its effectiveness and efficiency.
Authors: Sasha Boguraev, Ben Lipkin, Leonie Weissweiler, Kyle Mahowald
Abstract: Math is constructed by people for people: just as natural language corpora reflect not just propositions but the communicative goals of language users, the math data that models are trained on reflects not just idealized mathematical entities but rich communicative intentions. While there are important advantages to treating math in a purely symbolic manner, we here hypothesize that there are benefits to treating math as situated linguistic communication and that language models are well suited for this goal, in ways that are not fully appreciated. We illustrate these points with two case studies. First, we ran an experiment in which we found that language models interpret the equals sign in a humanlike way -- generating systematically different word problems for the same underlying equation arranged in different ways. Second, we found that language models prefer proofs to be ordered in naturalistic ways, even though other orders would be logically equivalent. We advocate for AI systems that learn from and represent the communicative intentions latent in human-generated math.
Authors: Xiaojuan Tang, Jiaqi Li, Yitao Liang, Song-chun Zhu, Muhan Zhang, Zilong Zheng
Abstract: Large Language Models (LLMs) trained on massive corpora have shown remarkable success in knowledge-intensive tasks. Yet, most of them rely on pre-stored knowledge. Inducing new general knowledge from a specific environment and performing reasoning with the acquired knowledge -- \textit{situated inductive reasoning}, is crucial and challenging for machine intelligence. In this paper, we design Mars, an interactive environment devised for situated inductive reasoning. It introduces counter-commonsense game mechanisms by modifying terrain, survival setting and task dependency while adhering to certain principles. In Mars, agents need to actively interact with their surroundings, derive useful rules and perform decision-making tasks in specific contexts. We conduct experiments on various RL-based and LLM-based methods, finding that they all struggle on this challenging situated inductive reasoning benchmark. Furthermore, we explore \textit{Induction from Reflection}, where we instruct agents to perform inductive reasoning from history trajectory. The superior performance underscores the importance of inductive reasoning in Mars. Through Mars, we aim to galvanize advancements in situated inductive reasoning and set the stage for developing the next generation of AI systems that can reason in an adaptive and context-sensitive way.
Authors: Mingan Lin, Fan Yang, Yanjun Shen, Haoze Sun, Tianpeng Li, Tao Zhang, Chenzheng Zhu, Tao Zhang, Miao Zheng, Xu Li, Yijie Zhou, Mingyang Chen, Yanzhao Qin, Youquan Li, Hao Liang, Fei Li, Yadong Li, Mang Wang, Guosheng Dong, Kun Fang, Jianhua Xu, Bin Cui, Wentao Zhang, Zenan Zhou, Weipeng Chen
Abstract: We introduce Baichuan Alignment, a detailed analysis of the alignment techniques employed in the Baichuan series of models. This represents the industry's first comprehensive account of alignment methodologies, offering valuable insights for advancing AI research. We investigate the critical components that enhance model performance during the alignment process, including optimization methods, data strategies, capability enhancements, and evaluation processes. The process spans three key stages: Prompt Augmentation System (PAS), Supervised Fine-Tuning (SFT), and Preference Alignment. The problems encountered, the solutions applied, and the improvements made are thoroughly recorded. Through comparisons across well-established benchmarks, we highlight the technological advancements enabled by Baichuan Alignment. Baichuan-Instruct is an internal model, while Qwen2-Nova-72B and Llama3-PBM-Nova-70B are instruct versions of the Qwen2-72B and Llama-3-70B base models, optimized through Baichuan Alignment. Baichuan-Instruct demonstrates significant improvements in core capabilities, with user experience gains ranging from 17% to 28%, and performs exceptionally well on specialized benchmarks. In open-source benchmark evaluations, both Qwen2-Nova-72B and Llama3-PBM-Nova-70B consistently outperform their respective official instruct versions across nearly all datasets. This report aims to clarify the key technologies behind the alignment process, fostering a deeper understanding within the community. Llama3-PBM-Nova-70B model is available at https://huggingface.co/PKU-Baichuan-MLSystemLab/Llama3-PBM-Nova-70B.
URLs: https://huggingface.co/PKU-Baichuan-MLSystemLab/Llama3-PBM-Nova-70B.
Authors: Junjie Oscar Yin, Alexander M. Rush
Abstract: Data selection can reduce the amount of training data needed to finetune LLMs; however, the efficacy of data selection scales directly with its compute. Motivated by the practical challenge of compute-constrained finetuning, we consider the setting in which both the cost of selecting data and training are budgeted for. We first formalize the problem of data selection with a cost-aware utility function, and model the data selection problem as trading off initial-selection cost for training gain. We run a comprehensive sweep of experiments across multiple tasks, varying compute budget by scaling finetuning tokens, model sizes, and data selection compute. These experiments show the validity of this model in real-world experiments. Interestingly we find that many powerful data selection methods are almost never compute-optimal, and that cheaper data selection alternatives dominate both from a theoretical and empirical perspective.
Authors: Franklin Wang, Sumanth Hegde
Abstract: Offline paired preference optimization algorithms have become a popular approach for fine-tuning on preference data, outperforming traditional supervised fine-tuning in various tasks. However, traditional implementations often involve redundant computations, especially for tasks with long shared prompts. We introduce prefix sharing for preference tuning, a novel technique that processes chosen and rejected responses as one sequence with a shared prefix. To prevent cross-response contamination, we use a custom block-sparse attention mask. Our method achieves $1.1$-$1.5\times$ improvement in training throughput on popular DPO datasets, without any effect on convergence. When combined with sequence packing, we observe consistent $1.3$-$1.6\times$ speedups, benefiting even datasets with smaller sequence lengths. While we focus on Direct Preference Optimization (DPO), our approach is applicable to other paired preference tuning methods. By enhancing computational efficiency, our work contributes to making preference-based fine-tuning more accessible for a wider range of applications and model sizes. We open-source our code at https://github.com/frankxwang/dpo-prefix-sharing.
Authors: Shanchao Liang, Yiran Hu, Nan Jiang, Lin Tan
Abstract: Large language models (LLMs) have achieved high accuracy, i.e., more than 90 pass@1, in solving Python coding problems in HumanEval and MBPP. Thus, a natural question is, whether LLMs achieve comparable code completion performance compared to human developers? Unfortunately, one cannot answer this question using existing manual crafted or simple (e.g., single-line) code generation benchmarks, since such tasks fail to represent real-world software development tasks. In addition, existing benchmarks often use poor code correctness metrics, providing misleading conclusions. To address these challenges, we create REPOCOD, a code generation benchmark with 980 problems collected from 11 popular real-world projects, with more than 58% of them requiring file-level or repository-level context information. In addition, REPOCOD has the longest average canonical solution length (331.6 tokens) and the highest average cyclomatic complexity (9.00) compared to existing benchmarks. Each task in REPOCOD includes 313.5 developerwritten test cases on average for better correctness evaluation. In our evaluations of ten LLMs, none of the models achieve more than 30 pass@1 on REPOCOD, indicating the necessity of building stronger LLMs that can help developers in real-world software development. REPOCOD is available at https://github.com/ltasset/REPOCOD
Authors: Hazem Ibrahim, Nouar AlDahoul, Syed Mustafa Ali Abbasi, Fareed Zaffar, Talal Rahwan, Yasir Zaki
Abstract: The manner in which different racial and gender groups are portrayed in news coverage plays a large role in shaping public opinion. As such, understanding how such groups are portrayed in news media is of notable societal value, and has thus been a significant endeavour in both the computer and social sciences. Yet, the literature still lacks a longitudinal study examining both the frequency of appearance of different racial and gender groups in online news articles, as well as the context in which such groups are discussed. To fill this gap, we propose two machine learning classifiers to detect the race and age of a given subject. Next, we compile a dataset of 123,337 images and 441,321 online news articles from New York Times (NYT) and Fox News (Fox), and examine representation through two computational approaches. Firstly, we examine the frequency and prominence of appearance of racial and gender groups in images embedded in news articles, revealing that racial and gender minorities are largely under-represented, and when they do appear, they are featured less prominently compared to majority groups. Furthermore, we find that NYT largely features more images of racial minority groups compared to Fox. Secondly, we examine both the frequency and context with which racial minority groups are presented in article text. This reveals the narrow scope in which certain racial groups are covered and the frequency with which different groups are presented as victims and/or perpetrators in a given conflict. Taken together, our analysis contributes to the literature by providing two novel open-source classifiers to detect race and age from images, and shedding light on the racial and gender biases in news articles from venues on opposite ends of the American political spectrum.
Authors: Zhiqi Bu, Xiaomeng Jin, Bhanukiran Vinzamuri, Anil Ramakrishna, Kai-Wei Chang, Volkan Cevher, Mingyi Hong
Abstract: Machine unlearning has been used to remove unwanted knowledge acquired by large language models (LLMs). In this paper, we examine machine unlearning from an optimization perspective, framing it as a regularized multi-task optimization problem, where one task optimizes a forgetting objective and another optimizes the model performance. In particular, we introduce a normalized gradient difference (NGDiff) algorithm, enabling us to have better control over the trade-off between the objectives, while integrating a new, automatic learning rate scheduler. We provide a theoretical analysis and empirically demonstrate the superior performance of NGDiff among state-of-the-art unlearning methods on the TOFU and MUSE datasets while exhibiting stable training.