llm - 2024_10
Navigation
Papers
Synthetic data augmentation via large language models (LLMs) allows researchers to leverage additional training data, thus enhancing the performance of downstream tasks, especially when real-world data is scarce. However, the generated data can deviate from the real-world data, and this misalignment can bring deficient outcomes while applying the trained model to applications. Therefore, we proposed efficient weighted-loss approaches to align synthetic data with real-world distribution by emphasizing high-quality and diversified data generated by LLMs with using merely a little real-world data. We empirically assessed the effectiveness of our method on multiple text classification tasks, and the results showed leveraging our approaches on a BERT-level model robustly outperformed standard cross-entropy and other data weighting approaches, providing potential solutions to effectively leveraging synthetic data from any suitable data generator for model training.
In this work, we address the problem of large language model (LLM) unlearning, aiming to remove unwanted data influences and associated model capabilities (e.g., copyrighted data or harmful content generation) while preserving essential model utilities, without the need for retraining from scratch. Despite the growing need for LLM unlearning, a principled optimization framework remains lacking. To this end, we revisit the state-of-the-art approach, negative preference optimization (NPO), and identify the issue of reference model bias, which could undermine NPO's effectiveness, particularly when unlearning forget data of varying difficulty. Given that, we propose a simple yet effective unlearning optimization framework, called SimNPO, showing that 'simplicity' in removing the reliance on a reference model (through the lens of simple preference optimization) benefits unlearning. We also provide deeper insights into SimNPO's advantages, supported by analysis using mixtures of Markov chains. Furthermore, we present extensive experiments validating SimNPO's superiority over existing unlearning baselines in benchmarks like TOFU and MUSE, and robustness against relearning attacks. Codes are available at https://github.com/OPTML-Group/Unlearn-Simple.
Large Language Models (LLMs) have shown promise in highly-specialized domains, however challenges are still present in aspects of accuracy and costs. These limitations restrict the usage of existing models in domain-specific tasks. While fine-tuning pre-trained models have shown promising results, this process can be computationally expensive and require massive datasets of the specialized application in hand. In this work, we bridge that gap. We have developed Phi-2-Legal and Mistral-Legal-7B, which are language models specifically designed for legal applications. These models are based on Phi-2 and Mistral-7B-v0.1, and have gone through continued pre-training with over 500 million tokens of legal texts. Our innovative approach significantly improves capabilities in legal tasks by using Large Language Models (LLMs) to convert raw training data into reading comprehension text. Our legal LLMs have demonstrated superior performance in legal benchmarks, even outperforming models trained on much larger datasets with more resources. This work emphasizes the effectiveness of continued pre-training on domain-specific texts, while using affordable LLMs for data conversion, which gives these models domain expertise while retaining general language understanding capabilities. While this work uses the legal domain as a test case, our method can be scaled and applied to any pre-training dataset, resulting in significant improvements across different tasks. These findings underscore the potential of domain-adaptive pre-training and reading comprehension for the development of highly effective domain-specific language models.
Quantifying the effect of textual interventions in social systems, such as reducing anger in social media posts to see its impact on engagement, poses significant challenges. Direct interventions on real-world systems are often infeasible, necessitating reliance on observational data. Traditional causal inference methods, typically designed for binary or discrete treatments, are inadequate for handling the complex, high-dimensional nature of textual data. This paper addresses these challenges by proposing a novel approach, CausalDANN, to estimate causal effects using text transformations facilitated by large language models (LLMs). Unlike existing methods, our approach accommodates arbitrary textual interventions and leverages text-level classifiers with domain adaptation ability to produce robust effect estimates against domain shifts, even when only the control group is observed. This flexibility in handling various text interventions is a key advancement in causal estimation for textual data, offering opportunities to better understand human behaviors and develop effective policies within social systems.
With the widespread deployment of long-context large language models (LLMs), there has been a growing demand for efficient support of high-throughput inference. However, as the key-value (KV) cache expands with the sequence length, the increasing memory footprint and the need to access it for each token generation both result in low throughput when serving long-context LLMs. While various dynamic sparse attention methods have been proposed to speed up inference while maintaining generation quality, they either fail to sufficiently reduce GPU memory consumption or introduce significant decoding latency by offloading the KV cache to the CPU. We present ShadowKV, a high-throughput long-context LLM inference system that stores the low-rank key cache and offloads the value cache to reduce the memory footprint for larger batch sizes and longer sequences. To minimize decoding latency, ShadowKV employs an accurate KV selection strategy that reconstructs minimal sparse KV pairs on-the-fly. By evaluating ShadowKV on a broad range of benchmarks, including RULER, LongBench, and Needle In A Haystack, and models like Llama-3.1-8B, Llama-3-8B-1M, GLM-4-9B-1M, Yi-9B-200K, Phi-3-Mini-128K, and Qwen2-7B-128K, we demonstrate that it can support up to 6$\times$ larger batch sizes and boost throughput by up to 3.04$\times$ on an A100 GPU without sacrificing accuracy, even surpassing the performance achievable with infinite batch size under the assumption of infinite GPU memory. The code is available at https://github.com/bytedance/ShadowKV.
Large language models (LLMs) have been increasingly employed for (interactive) decision-making, via the development of LLM-based autonomous agents. Despite their emerging successes, the performance of LLM agents in decision-making has not been fully investigated through quantitative metrics, especially in the multi-agent setting when they interact with each other, a typical scenario in real-world LLM-agent applications. To better understand the limits of LLM agents in these interactive environments, we propose to study their interactions in benchmark decision-making settings in online learning and game theory, through the performance metric of \emph{regret}. We first empirically study the {no-regret} behaviors of LLMs in canonical (non-stationary) online learning problems, as well as the emergence of equilibria when LLM agents interact through playing repeated games. We then provide some theoretical insights into the no-regret behaviors of LLM agents, under certain assumptions on the supervised pre-training and the rationality model of human decision-makers who generate the data. Notably, we also identify (simple) cases where advanced LLMs such as GPT-4 fail to be no-regret. To promote the no-regret behaviors, we propose a novel \emph{unsupervised} training loss of \emph{regret-loss}, which, in contrast to the supervised pre-training loss, does not require the labels of (optimal) actions. We then establish the statistical guarantee of generalization bound for regret-loss minimization, followed by the optimization guarantee that minimizing such a loss may automatically lead to known no-regret learning algorithms. Our further experiments demonstrate the effectiveness of our regret-loss, especially in addressing the above ``regrettable'' cases.
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.
Social relation reasoning aims to identify relation categories such as friends, spouses, and colleagues from images. While current methods adopt the paradigm of training a dedicated network end-to-end using labeled image data, they are limited in terms of generalizability and interpretability. To address these issues, we first present a simple yet well-crafted framework named {\name}, which combines the perception capability of Vision Foundation Models (VFMs) and the reasoning capability of Large Language Models (LLMs) within a modular framework, providing a strong baseline for social relation recognition. Specifically, we instruct VFMs to translate image content into a textual social story, and then utilize LLMs for text-based reasoning. {\name} introduces systematic design principles to adapt VFMs and LLMs separately and bridge their gaps. Without additional model training, it achieves competitive zero-shot results on two databases while offering interpretable answers, as LLMs can generate language-based explanations for the decisions. The manual prompt design process for LLMs at the reasoning phase is tedious and an automated prompt optimization method is desired. As we essentially convert a visual classification task into a generative task of LLMs, automatic prompt optimization encounters a unique long prompt optimization issue. To address this issue, we further propose the Greedy Segment Prompt Optimization (GSPO), which performs a greedy search by utilizing gradient information at the segment level. Experimental results show that GSPO significantly improves performance, and our method also generalizes to different image styles. The code is available at https://github.com/Mengzibin/SocialGPT.
Automated program repair has emerged as a powerful technique to mitigate the impact of software bugs on system reliability and user experience. This paper introduces RepairAgent, the first work to address the program repair challenge through an autonomous agent based on a large language model (LLM). Unlike existing deep learning-based approaches, which prompt a model with a fixed prompt or in a fixed feedback loop, our work treats the LLM as an agent capable of autonomously planning and executing actions to fix bugs by invoking suitable tools. RepairAgent freely interleaves gathering information about the bug, gathering repair ingredients, and validating fixes, while deciding which tools to invoke based on the gathered information and feedback from previous fix attempts. Key contributions that enable RepairAgent include a set of tools that are useful for program repair, a dynamically updated prompt format that allows the LLM to interact with these tools, and a finite state machine that guides the agent in invoking the tools. Our evaluation on the popular Defects4J dataset demonstrates RepairAgent's effectiveness in autonomously repairing 164 bugs, including 39 bugs not fixed by prior techniques. Interacting with the LLM imposes an average cost of 270,000 tokens per bug, which, under the current pricing of OpenAI's GPT-3.5 model, translates to 14 cents of USD per bug. To the best of our knowledge, this work is the first to present an autonomous, LLM-based agent for program repair, paving the way for future agent-based techniques in software engineering.
The discovery of causal relationships between random variables is an important yet challenging problem that has applications across many scientific domains. Differentiable causal discovery (DCD) methods are effective in uncovering causal relationships from observational data; however, these approaches often suffer from limited interpretability and face challenges in incorporating domain-specific prior knowledge. In contrast, Large Language Models (LLMs)-based causal discovery approaches have recently been shown capable of providing useful priors for causal discovery but struggle with formal causal reasoning. In this paper, we propose LLM-DCD, which uses an LLM to initialize the optimization of the maximum likelihood objective function of DCD approaches, thereby incorporating strong priors into the discovery method. To achieve this initialization, we design our objective function to depend on an explicitly defined adjacency matrix of the causal graph as its only variational parameter. Directly optimizing the explicitly defined adjacency matrix provides a more interpretable approach to causal discovery. Additionally, we demonstrate higher accuracy on key benchmarking datasets of our approach compared to state-of-the-art alternatives, and provide empirical evidence that the quality of the initialization directly impacts the quality of the final output of our DCD approach. LLM-DCD opens up new opportunities for traditional causal discovery methods like DCD to benefit from future improvements in the causal reasoning capabilities of LLMs.
Software testing is an essential part of the software development cycle to improve the code quality. Typically, a unit test consists of a test prefix and a test oracle which captures the developer's intended behaviour. A known limitation of traditional test generation techniques (e.g. Randoop and Evosuite) is that they produce test oracles that capture the actual program behaviour rather than the expected one. Recent approaches leverage Large Language Models (LLMs), trained on an enormous amount of data, to generate developer-like code and test cases. We investigate whether the LLM-generated test oracles capture the actual or expected software behaviour. We thus, conduct a controlled experiment to answer this question, by studying LLMs performance on two tasks, namely, test oracle classification and generation. The study includes developer-written and automatically generated test cases and oracles for 24 open-source Java repositories, and different well tested prompts. Our findings show that LLM-based test generation approaches are also prone on generating oracles that capture the actual program behaviour rather than the expected one. Moreover, LLMs are better at generating test oracles rather than classifying the correct ones, and can generate better test oracles when the code contains meaningful test or variable names. Finally, LLM-generated test oracles have higher fault detection potential than the Evosuite ones.
LLMs can be used in a variety of code related tasks such as translating from one programming language to another, implementing natural language requirements and code summarization. Artifacts generated by state of the art LLM technology are expected to be useful in the sense that a user will be able to use the LLM generated artifact after a small number of easy modifications. Quantifying this vague notion is challenging and it is thus hard to determine the quality of code related LLM solutions. We refer to evaluation of LLM solutions using LLM judgment as "LLM as a Judge", or LaaJ for short. In this work we introduce a methodology to generate and evaluate LaaJ implementations, utilizing an automatically generated benchmark. The purpose of the benchmark is two fold, namely, it is used both to develop and validate the LaaJs and to validate and test the LLM code related solution using the LaaJs. To that end, we developed an automated benchmark generation engine, which generates code in multiple programming languages for multiple code related tasks and which serves as the input for LaaJ evaluation. We utilize a graph representation, G, of the potential code related generations. The graph vertices are generated artifacts and edges represent possible generations, e.g., the generation of a Java program from its natural language requirements. Utilizing a chain of LLM agents and G we generate code related artifacts. Using cycles in G we formulate expectations on the generated artifacts. Taking advantage of these formulated expectations enables the development and testing of reliable LLM judgement for usefulness of the artifacts generated by the solution. Our approach enables the creation of high quality code task solutions.
Textual descriptions in cyber threat intelligence (CTI) reports, such as security articles and news, are rich sources of knowledge about cyber threats, crucial for organizations to stay informed about the rapidly evolving threat landscape. However, current CTI extraction methods lack flexibility and generalizability, often resulting in inaccurate and incomplete knowledge extraction. Syntax parsing relies on fixed rules and dictionaries, while model fine-tuning requires large annotated datasets, making both paradigms challenging to adapt to new threats and ontologies. To bridge the gap, we propose CTINexus, a novel framework leveraging optimized in-context learning (ICL) of large language models (LLMs) for data-efficient CTI knowledge extraction and high-quality cybersecurity knowledge graph (CSKG) construction. Unlike existing methods, CTINexus requires neither extensive data nor parameter tuning and can adapt to various ontologies with minimal annotated examples. This is achieved through (1) a carefully designed automatic prompt construction strategy with optimal demonstration retrieval for extracting a wide range of cybersecurity entities and relations; (2) a hierarchical entity alignment technique that canonicalizes the extracted knowledge and removes redundancy; (3) an ICL-enhanced long-distance relation prediction technique to further complete the CKSG with missing links. Our extensive evaluations using 150 real-world CTI reports collected from 10 platforms demonstrate that CTINexus significantly outperforms existing methods in constructing accurate and complete CSKGs, highlighting its potential to transform CTI analysis with an efficient and adaptable solution for the dynamic threat landscape.
Autoregressive (AR) Large Language Models (LLMs) have demonstrated significant success across numerous tasks. However, the AR modeling paradigm presents certain limitations; for instance, contemporary autoregressive LLMs are trained to generate one token at a time, which can result in noticeable latency. Recent advances have indicated that search and repeated sampling can enhance performance in various applications, such as theorem proving, code generation, and alignment, by utilizing greater computational resources during inference. In this study, we demonstrate that diffusion language models are capable of generating at least 32 tokens simultaneously, while exceeding the performance of AR models in text quality and on the LAMBADA natural language understanding benchmark. This outcome is achieved through a novel distillation method for discrete diffusion models, which reduces the number of inference steps by a factor of 32-64. Practically, our models, even without caching, can generate tokens at a rate that is up to 8 times faster than AR models employing KV caching, and we anticipate further improvements with the inclusion of caching. Moreover, we demonstrate the efficacy of our approach for diffusion language models with up to 860M parameters.
Large language models (LLMs) often produce errors, including factual inaccuracies, biases, and reasoning failures, collectively referred to as "hallucinations". Recent studies have demonstrated that LLMs' internal states encode information regarding the truthfulness of their outputs, and that this information can be utilized to detect errors. In this work, we show that the internal representations of LLMs encode much more information about truthfulness than previously recognized. We first discover that the truthfulness information is concentrated in specific tokens, and leveraging this property significantly enhances error detection performance. Yet, we show that such error detectors fail to generalize across datasets, implying that -- contrary to prior claims -- truthfulness encoding is not universal but rather multifaceted. Next, we show that internal representations can also be used for predicting the types of errors the model is likely to make, facilitating the development of tailored mitigation strategies. Lastly, we reveal a discrepancy between LLMs' internal encoding and external behavior: they may encode the correct answer, yet consistently generate an incorrect one. Taken together, these insights deepen our understanding of LLM errors from the model's internal perspective, which can guide future research on enhancing error analysis and mitigation.
Traffic prediction is a vital component of intelligent transportation systems. By reasoning about traffic patterns in both the spatial and temporal dimensions, accurate and interpretable predictions can be provided. A considerable challenge in traffic prediction lies in handling the diverse data distributions caused by vastly different traffic conditions occurring at different locations. LLMs have been a dominant solution due to their remarkable capacity to adapt to new datasets with very few labeled data samples, i.e., few-shot adaptability. However, existing forecasting techniques mainly focus on extracting local graph information and forming a text-like prompt, leaving LLM- based traffic prediction an open problem. This work presents a probabilistic LLM for traffic forecasting with three highlights. We propose a graph-aware LLM for traffic prediction that considers proximal traffic information. Specifically, by considering the traffic of neighboring nodes as covariates, our model outperforms the corresponding time-series LLM. Furthermore, we adopt a lightweight approach for efficient domain adaptation when facing new data distributions in few-shot fashion. The comparative experiment demonstrates the proposed method outperforms the state-of-the-art LLM-based methods and the traditional GNN- based supervised approaches. Furthermore, Strada-LLM can be easily adapted to different LLM backbones without a noticeable performance drop.
We present a simple meta quantization approach that quantizes different layers of a large language model (LLM) at different bit levels, and is independent of the underlying quantization technique. Specifically, we quantize the most important layers to higher bit precision and less important layers to lower bits. We propose two effective strategies to measure the importance of layers within LLMs: the first measures the importance of a layer based on how different its output embeddings are from the input embeddings (higher is better); the second estimates the importance of a layer using the number of layer weights that are much larger than average (smaller is better). We show that quantizing different layers at varying bits according to our importance scores results in minimal performance drop with a far more compressed model size. Finally, we present several practical key takeaways from our variable layer-wise quantization experiments: (a) LLM performance under variable quantization remains close to the original model until 25-50% of layers are moved in lower quantization using our proposed ordering but only until 5-10% if moved using no specific ordering; (b) Adding layer importance to inherently dynamic quantization techniques can further improve their performance, showing that our approach is complementary to other dynamic quantization methods; (c) Quantizing LLMs to lower bits performs substantially better than pruning unless extreme quantization (2-bit) is used; and (d) Layer-wise quantization to lower bits works better in the case of larger LLMs with more layers compared to smaller LLMs with fewer layers. Our code is publicly available at https://github.com/RazvanDu/LayerwiseQuant/.
Recent studies have demonstrated that large language models (LLMs) exhibit significant biases in evaluation tasks, particularly in preferentially rating and favoring self-generated content. However, the extent to which this bias manifests in fact-oriented tasks, especially within retrieval-augmented generation (RAG) frameworks-where keyword extraction and factual accuracy take precedence over stylistic elements-remains unclear. Our study addresses this knowledge gap by simulating two critical phases of the RAG framework. In the first phase, we access the suitability of human-authored versus model-generated passages, emulating the pointwise reranking process. The second phase involves conducting pairwise reading comprehension tests to simulate the generation process. Contrary to previous findings indicating a self-preference in rating tasks, our results reveal no significant self-preference effect in RAG frameworks. Instead, we observe that factual accuracy significantly influences LLMs' output, even in the absence of prior knowledge. Our research contributes to the ongoing discourse on LLM biases and their implications for RAG-based system, offering insights that may inform the development of more robust and unbiased LLM systems.
Machine Translation Quality Estimation (MTQE) is the task of estimating the quality of machine-translated text in real time without the need for reference translations, which is of great importance for the development of MT. After two decades of evolution, QE has yielded a wealth of results. This article provides a comprehensive overview of QE datasets, annotation methods, shared tasks, methodologies, challenges, and future research directions. It begins with an introduction to the background and significance of QE, followed by an explanation of the concepts and evaluation metrics for word-level QE, sentence-level QE, document-level QE, and explainable QE. The paper categorizes the methods developed throughout the history of QE into those based on handcrafted features, deep learning, and Large Language Models (LLMs), with a further division of deep learning-based methods into classic deep learning and those incorporating pre-trained language models (LMs). Additionally, the article details the advantages and limitations of each method and offers a straightforward comparison of different approaches. Finally, the paper discusses the current challenges in QE research and provides an outlook on future research directions.
Recently, large language models (LLMs) have been increasingly used to support various decision-making tasks, assisting humans in making informed decisions. However, when LLMs confidently provide incorrect information, it can lead humans to make suboptimal decisions. To prevent LLMs from generating incorrect information on topics they are unsure of and to improve the accuracy of generated content, prior works have proposed Retrieval Augmented Generation (RAG), where external documents are referenced to generate responses. However, traditional RAG methods focus only on retrieving documents most relevant to the input query, without specifically aiming to ensure that the human user's decisions are well-calibrated. To address this limitation, we propose a novel retrieval method called Calibrated Retrieval-Augmented Generation (CalibRAG), which ensures that decisions informed by the retrieved documents are well-calibrated. Then we empirically validate that CalibRAG improves calibration performance as well as accuracy, compared to other baselines across various datasets.
In line with the principle of honesty, there has been a growing effort to train large language models (LLMs) to generate outputs containing epistemic markers. However, evaluation in the presence of epistemic markers has been largely overlooked, raising a critical question: Could the use of epistemic markers in LLM-generated outputs lead to unintended negative consequences? To address this, we present EMBER, a benchmark designed to assess the robustness of LLM-judges to epistemic markers in both single and pairwise evaluation settings. Our findings, based on evaluations using EMBER, reveal that all tested LLM-judges, including GPT-4o, show a notable lack of robustness in the presence of epistemic markers. Specifically, we observe a negative bias toward epistemic markers, with a stronger bias against markers expressing uncertainty. This suggests that LLM-judges are influenced by the presence of these markers and do not focus solely on the correctness of the content.
Despite the impressive generative abilities of black-box large language models (LLMs), their inherent opacity hinders further advancements in capabilities such as reasoning, planning, and personalization. Existing works aim to enhance LLM capabilities via domain-specific adaptation or in-context learning, which require additional training on accessible model parameters, an infeasible option for black-box LLMs. To address this challenge, we introduce Matryoshika, a lightweight white-box LLM controller that guides a large-scale black-box LLM generator by decomposing complex tasks into a series of intermediate outputs. Specifically, we consider the black-box LLM as an environment, with Matryoshika serving as a policy to provide intermediate guidance through prompts for driving the black-box LLM. Matryoshika is trained to pivot the outputs of the black-box LLM aligning with preferences during iterative interaction, which enables controllable multi-turn generation and self-improvement in optimizing intermediate guidance. Empirical evaluations on three diverse tasks demonstrate that Matryoshika effectively enhances the capabilities of black-box LLMs in complex, long-horizon tasks, including reasoning, planning, and personalization. By leveraging this pioneering controller-generator framework to mitigate dependence on model parameters, Matryoshika provides a transparent and practical solution for improving black-box LLMs through controllable multi-turn generation using white-box LLMs.
The era of Large Language Models (LLMs) raises new demands for automatic evaluation metrics, which should be adaptable to various application scenarios while maintaining low cost and effectiveness. Traditional metrics for automatic text evaluation are often tailored to specific scenarios, while LLM-based evaluation metrics are costly, requiring fine-tuning or rely heavily on the generation capabilities of LLMs. Besides, previous LLM-based metrics ignore the fact that, within the space of LLM representations, there exist direction vectors that indicate the estimation of text quality. To this end, we introduce RepEval, a metric that leverages the projection of LLM representations for evaluation. Through simple prompt modifications, RepEval can easily transition to various tasks, requiring only minimal sample pairs for direction vector construction. Results on fourteen datasets across two evaluation tasks demonstrate the high effectiveness of our method, which exhibits a higher correlation with human judgments than previous methods, even in complex evaluation scenarios involving pair-wise selection under nuanced aspects. Our work underscores the richness of information regarding text quality embedded within LLM representations, offering insights for the development of new metrics.
Large Language Models (LLMs) have demonstrated the ability to adopt a personality and behave in a human-like manner. There is a large body of research that investigates the behavioural impacts of personality in less obvious areas such as investment attitudes or creative decision making. In this study, we investigated whether an LLM persona with a specific Big Five personality profile would perform an investment task similarly to a human with the same personality traits. We used a simulated investment task to determine if these results could be generalised into actual behaviours. In this simulated environment, our results show these personas produced meaningful behavioural differences in all assessed categories, with these behaviours generally being consistent with expectations derived from human research. We found that LLMs are able to generalise traits into expected behaviours in three areas: learning style, impulsivity and risk appetite while environmental attitudes could not be accurately represented. In addition, we showed that LLMs produce behaviour that is more reflective of human behaviour in a simulation environment compared to a survey environment.
Navigation presents a significant challenge for persons with visual impairments (PVI). While traditional aids such as white canes and guide dogs are invaluable, they fall short in delivering detailed spatial information and precise guidance to desired locations. Recent developments in large language models (LLMs) and vision-language models (VLMs) offer new avenues for enhancing assistive navigation. In this paper, we introduce Guide-LLM, an embodied LLM-based agent designed to assist PVI in navigating large indoor environments. Our approach features a novel text-based topological map that enables the LLM to plan global paths using a simplified environmental representation, focusing on straight paths and right-angle turns to facilitate navigation. Additionally, we utilize the LLM's commonsense reasoning for hazard detection and personalized path planning based on user preferences. Simulated experiments demonstrate the system's efficacy in guiding PVI, underscoring its potential as a significant advancement in assistive technology. The results highlight Guide-LLM's ability to offer efficient, adaptive, and personalized navigation assistance, pointing to promising advancements in this field.
Owing to the impressive general intelligence of large language models (LLMs), there has been a growing trend to integrate them into recommender systems to gain a more profound insight into human interests and intentions. Existing LLMs-based recommender systems primarily leverage item attributes and user interaction histories in textual format, improving the single task like rating prediction or explainable recommendation. Nevertheless, these approaches overlook the crucial contribution of traditional collaborative signals in discerning users' profound intentions and disregard the interrelatedness among tasks. To address these limitations, we introduce a novel framework known as CKF, specifically developed to boost multi-task recommendations via personalized collaborative knowledge fusion into LLMs. Specifically, our method synergizes traditional collaborative filtering models to produce collaborative embeddings, subsequently employing the meta-network to construct personalized mapping bridges tailored for each user. Upon mapped, the embeddings are incorporated into meticulously designed prompt templates and then fed into an advanced LLM to represent user interests. To investigate the intrinsic relationship among diverse recommendation tasks, we develop Multi-Lora, a new parameter-efficient approach for multi-task optimization, adept at distinctly segregating task-shared and task-specific information. This method forges a connection between LLMs and recommendation scenarios, while simultaneously enriching the supervisory signal through mutual knowledge transfer among various tasks. Extensive experiments and in-depth robustness analyses across four common recommendation tasks on four large public data sets substantiate the effectiveness and superiority of our framework.
Sentiment analysis (SA) aims to identify the sentiment expressed in a text, such as a product review. Given a review and the sentiment associated with it, this work formulates SA as a combination of two tasks: (1) a causal discovery task that distinguishes whether a review "primes" the sentiment (Causal Hypothesis C1), or the sentiment "primes" the review (Causal Hypothesis C2); and (2) the traditional prediction task to model the sentiment using the review as input. Using the peak-end rule in psychology, we classify a sample as C1 if its overall sentiment score approximates an average of all the sentence-level sentiments in the review, and C2 if the overall sentiment score approximates an average of the peak and end sentiments. For the prediction task, we use the discovered causal mechanisms behind the samples to improve LLM performance by proposing causal prompts that give the models an inductive bias of the underlying causal graph, leading to substantial improvements by up to 32.13 F1 points on zero-shot five-class SA. Our code is at https://github.com/cogito233/causal-sa
Physically assistive robots present an opportunity to significantly increase the well-being and independence of individuals with motor impairments or other forms of disability who are unable to complete activities of daily living (ADLs). Speech interfaces, especially ones that utilize Large Language Models (LLMs), can enable individuals to effectively and naturally communicate high-level commands and nuanced preferences to robots. In this work, we demonstrate an LLM-based speech interface for a commercially available assistive feeding robot. Our system is based on an iteratively designed framework, from the paper "VoicePilot: Harnessing LLMs as Speech Interfaces for Physically Assistive Robots," that incorporates human-centric elements for integrating LLMs as interfaces for robots. It has been evaluated through a user study with 11 older adults at an independent living facility. Videos are located on our project website: https://sites.google.com/andrew.cmu.edu/voicepilot/.
Our interest is in constructing interactive systems involving a human-expert interacting with a machine learning engine on data analysis tasks. This is of relevance when addressing complex problems arising in areas of science, the environment, medicine and so on, which are not immediately amenable to the usual methods of statistical or mathematical modelling. In such situations, it is possible that harnessing human expertise and creativity to modern machine-learning capabilities of identifying patterns by constructing new internal representations of the data may provide some insight to possible solutions. In this paper, we examine the implementation of an abstract protocol developed for interaction between agents, each capable of constructing predictions and explanations. The \PXP protocol, described in [12] is motivated by the notion of ''two-way intelligibility'' and is specified using a pair of communicating finite-state machines. While the formalisation allows the authors to prove several properties about the protocol, no implementation was presented. Here, we address this shortcoming for the case in which one of the agents acts as a ''generator'' using a large language model (LLM) and the other is an agent that acts as a ''tester'' using either a human-expert, or a proxy for a human-expert (for example, a database compiled using human-expertise). We believe these use-cases will be a widely applicable form of interaction for problems of the kind mentioned above. We present an algorithmic description of general-purpose implementation, and conduct preliminary experiments on its use in two different areas (radiology and drug-discovery). The experimental results provide early evidence in support of the protocol's capability of capturing one- and two-way intelligibility in human-LLM in the manner proposed in [12].
Aligning human preference and value is an important requirement for contemporary foundation models. State-of-the-art techniques such as Reinforcement Learning from Human Feedback (RLHF) often consist of two stages: 1) supervised fine-tuning (SFT), where the model is fine-tuned by learning from human demonstration data; 2) Preference learning, where preference data is used to learn a reward model, which is in turn used by a reinforcement learning (RL) step to fine-tune the model. Such reward model serves as a proxy to human preference, and it is critical to guide the RL step towards improving the model quality. In this work, we argue that the SFT stage significantly benefits from learning a reward model as well. Instead of using the human demonstration data directly via supervised learning, we propose to leverage an Inverse Reinforcement Learning (IRL) technique to simultaneously build an reward model and a policy model. This approach leads to new SFT algorithms that are not only efficient to implement, but are robust to the presence of low-quality supervised learning data. Moreover, we discover a connection between the proposed IRL based approach, and a recent line of works called Self-Play Fine-tune (SPIN). Theoretically, we show that the proposed algorithms converge to the stationary solutions of the IRL problem. Empirically, we align 1B and 7B models using proposed methods and evaluate them on a reward benchmark model and the HuggingFace Open LLM Leaderboard. The proposed methods show significant performance improvement over existing SFT approaches. Our results indicate that it is beneficial to leverage reward learning throughout the entire alignment process.
This paper introduces a novel approach using Large Language Models (LLMs) integrated into an agent framework for flexible and effective personal mobility generation. LLMs overcome the limitations of previous models by effectively processing semantic data and offering versatility in modeling various tasks. Our approach addresses three research questions: aligning LLMs with real-world urban mobility data, developing reliable activity generation strategies, and exploring LLM applications in urban mobility. The key technical contribution is a novel LLM agent framework that accounts for individual activity patterns and motivations, including a self-consistency approach to align LLMs with real-world activity data and a retrieval-augmented strategy for interpretable activity generation. We evaluate our LLM agent framework and compare it with state-of-the-art personal mobility generation approaches, demonstrating the effectiveness of our approach and its potential applications in urban mobility. Overall, this study marks the pioneering work of designing an LLM agent framework for activity generation based on real-world human activity data, offering a promising tool for urban mobility analysis.
Large Language Models (LLMs) have increasingly been utilized in social simulations, where they are often guided by carefully crafted instructions to stably exhibit human-like behaviors during simulations. Nevertheless, we doubt the necessity of shaping agents' behaviors for accurate social simulations. Instead, this paper emphasizes the importance of spontaneous phenomena, wherein agents deeply engage in contexts and make adaptive decisions without explicit directions. We explored spontaneous cooperation across three competitive scenarios and successfully simulated the gradual emergence of cooperation, findings that align closely with human behavioral data. This approach not only aids the computational social science community in bridging the gap between simulations and real-world dynamics but also offers the AI community a novel method to assess LLMs' capability of deliberate reasoning.
The retrieval-augmented generation (RAG) approach is used to reduce the confabulation of large language models (LLMs) for question answering by retrieving and providing additional context coming from external knowledge sources (e.g., by adding the context to the prompt). However, injecting incorrect information can mislead the LLM to generate an incorrect answer. In this paper, we evaluate the effectiveness and robustness of four LLMs against misinformation - Gemma 2, GPT-4o-mini, Llama~3.1, and Mixtral - in answering biomedical questions. We assess the answer accuracy on yes-no and free-form questions in three scenarios: vanilla LLM answers (no context is provided), "perfect" augmented generation (correct context is provided), and prompt-injection attacks (incorrect context is provided). Our results show that Llama 3.1 (70B parameters) achieves the highest accuracy in both vanilla (0.651) and "perfect" RAG (0.802) scenarios. However, the accuracy gap between the models almost disappears with "perfect" RAG, suggesting its potential to mitigate the LLM's size-related effectiveness differences. We further evaluate the ability of the LLMs to generate malicious context on one hand and the LLM's robustness against prompt-injection attacks on the other hand, using metrics such as attack success rate (ASR), accuracy under attack, and accuracy drop. As adversaries, we use the same four LLMs (Gemma 2, GPT-4o-mini, Llama 3.1, and Mixtral) to generate incorrect context that is injected in the target model's prompt. Interestingly, Llama is shown to be the most effective adversary, causing accuracy drops of up to 0.48 for vanilla answers and 0.63 for "perfect" RAG across target models. Our analysis reveals that robustness rankings vary depending on the evaluation measure, highlighting the complexity of assessing LLM resilience to adversarial attacks.
Recent advancements in Large Language Models (LLMs) have shown remarkable performance across a wide range of tasks. Despite this, the auto-regressive nature of LLM decoding, which generates only a single token per forward propagation, fails to fully exploit the parallel computational power of GPUs, leading to considerable latency. To address this, we introduce a novel speculative decoding method named FIRP which generates multiple tokens instead of one at each decoding step. We achieve this by predicting the intermediate hidden states of future tokens (tokens have not been decoded yet) and then using these pseudo hidden states to decode future tokens, specifically, these pseudo hidden states are predicted with simple linear transformation in intermediate layers of LLMs. Once predicted, they participate in the computation of all the following layers, thereby assimilating richer semantic information. As the layers go deeper, the semantic gap between pseudo and real hidden states is narrowed and it becomes feasible to decode future tokens with high accuracy. To validate the effectiveness of FIRP, we conduct extensive experiments, showing a speedup ratio of 1.9x-3x in several models and datasets, analytical experiments also prove our motivations.
The advent of Large Language Models (LLMs) has revolutionized text generation, producing outputs that closely mimic human writing. This blurring of lines between machine- and human-written text presents new challenges in distinguishing one from the other a task further complicated by the frequent updates and closed nature of leading proprietary LLMs. Traditional logits-based detection methods leverage surrogate models for identifying LLM-generated content when the exact logits are unavailable from black-box LLMs. However, these methods grapple with the misalignment between the distributions of the surrogate and the often undisclosed target models, leading to performance degradation, particularly with the introduction of new, closed-source models. Furthermore, while current methodologies are generally effective when the source model is identified, they falter in scenarios where the model version remains unknown, or the test set comprises outputs from various source models. To address these limitations, we present Distribution-Aligned LLMs Detection (DALD), an innovative framework that redefines the state-of-the-art performance in black-box text detection even without logits from source LLMs. DALD is designed to align the surrogate model's distribution with that of unknown target LLMs, ensuring enhanced detection capability and resilience against rapid model iterations with minimal training investment. By leveraging corpus samples from publicly accessible outputs of advanced models such as ChatGPT, GPT-4 and Claude-3, DALD fine-tunes surrogate models to synchronize with unknown source model distributions effectively.
We introduce TorchOpera, a compound AI system for enhancing the safety and quality of prompts and responses for Large Language Models. TorchOpera ensures that all user prompts are safe, contextually grounded, and effectively processed, while enhancing LLM responses to be relevant and high quality. TorchOpera utilizes the vector database for contextual grounding, rule-based wrappers for flexible modifications, and specialized mechanisms for detecting and adjusting unsafe or incorrect content. We also provide a view of the compound AI system to reduce the computational cost. Extensive experiments show that TorchOpera ensures the safety, reliability, and applicability of LLMs in real-world settings while maintaining the efficiency of LLM responses.
Speech Emotion Recognition (SER) focuses on identifying emotional states from spoken language. The 2024 IEEE SLT-GenSEC Challenge on Post Automatic Speech Recognition (ASR) Emotion Recognition tasks participants to explore the capabilities of large language models (LLMs) for emotion recognition using only text data. We propose a novel approach that first refines all available transcriptions to ensure data reliability. We then segment each complete conversation into smaller dialogues and use these dialogues as context to predict the emotion of the target utterance within the dialogue. Finally, we investigated different context lengths and prompting techniques to improve prediction accuracy. Our best submission exceeded the baseline by 20% in unweighted accuracy, achieving the best performance in the challenge. All our experiments' codes, prediction results, and log files are publicly available.
Scientific innovation is pivotal for humanity, and harnessing large language models (LLMs) to generate research ideas could transform discovery. However, existing LLMs often produce simplistic and repetitive suggestions due to their limited ability in acquiring external knowledge for innovation. To address this problem, we introduce an enhanced planning and search methodology designed to boost the creative potential of LLM-based systems. Our approach involves an iterative process to purposely plan the retrieval of external knowledge, progressively enriching the idea generation with broader and deeper insights. Validation through automated and human assessments indicates that our framework substantially elevates the quality of generated ideas, particularly in novelty and diversity. The number of unique novel ideas produced by our framework is 3.4 times higher than without it. Moreover, our method outperforms the current state-of-the-art, generating at least 2.5 times more top-rated ideas based on 170 seed papers in a Swiss Tournament evaluation.
Text detoxification, a variant of style transfer tasks, finds useful applications in online social media. This work presents a fine-tuning method that only uses non-parallel data to turn large language models (LLM) into a detoxification rewritter. We model the fine-tuning process as a Stackelberg game between an LLM (leader) and a toxicity screener (follower), which is a binary style classifier (toxic or non-toxic). The LLM aims to align its preference according to the screener and generate paraphases passing the screening. The primary challenge of non-parallel data fine-tuning is incomplete preference. In the case of unsuccessful paraphrases, the classifier cannot establish a preference between the input and paraphrase, as they belong to the same toxic style. Hence, preference-alignment fine-tuning methods, such as direct preference optimization (DPO), no longer apply. To address the challenge of incomplete preference, we propose Stackelberg response optimization (SRO), adapted from DPO, to enable the LLM to learn from the follower's response. The gist is that SRO decreases the likelihood of generating the paraphrase if it fails the follower's screening while performing DPO on the pair of the toxic input and its paraphrase when the latter passes the screening. Experiments indicate that the SRO-fine-tunned LLM achieves satisfying performance comparable to state-of-the-art models regarding style accuracy, content similarity, and fluency. The overall detoxification performance surpasses other computing methods and matches the human reference. Additional empirical evidence suggests that SRO is sensitive to the screener's feedback, and a slight perturbation leads to a significant performance drop. We release the code and LLM models at \url{https://github.com/XXXinhong/Detoxification_LLM}.
The potential of using Large Language Models (LLMs) themselves to evaluate LLM outputs offers a promising method for assessing model performance across various contexts. Previous research indicates that LLM-as-a-judge exhibits a strong correlation with human judges in the context of general instruction following. However, for instructions that require specialized knowledge, the validity of using LLMs as judges remains uncertain. In our study, we applied a mixed-methods approach, conducting pairwise comparisons in which both subject matter experts (SMEs) and LLMs evaluated outputs from domain-specific tasks. We focused on two distinct fields: dietetics, with registered dietitian experts, and mental health, with clinical psychologist experts. Our results showed that SMEs agreed with LLM judges 68% of the time in the dietetics domain and 64% in mental health when evaluating overall preference. Additionally, the results indicated variations in SME-LLM agreement across domain-specific aspect questions. Our findings emphasize the importance of keeping human experts in the evaluation process, as LLMs alone may not provide the depth of understanding required for complex, knowledge specific tasks. We also explore the implications of LLM evaluations across different domains and discuss how these insights can inform the design of evaluation workflows that ensure better alignment between human experts and LLMs in interactive systems.
Evaluating Large Language Models (LLMs) on reasoning benchmarks demonstrates their ability to solve compositional questions. However, little is known of whether these models engage in genuine logical reasoning or simply rely on implicit cues to generate answers. In this paper, we investigate the transitive reasoning capabilities of two distinct LLM architectures, LLaMA 2 and Flan-T5, by manipulating facts within two compositional datasets: QASC and Bamboogle. We controlled for potential cues that might influence the models' performance, including (a) word/phrase overlaps across sections of test input; (b) models' inherent knowledge during pre-training or fine-tuning; and (c) Named Entities. Our findings reveal that while both models leverage (a), Flan-T5 shows more resilience to experiments (b and c), having less variance than LLaMA 2. This suggests that models may develop an understanding of transitivity through fine-tuning on knowingly relevant datasets, a hypothesis we leave to future work.
This study explores the integration of large language models (LLMs) into classic inflation nowcasting frameworks, particularly in light of high inflation volatility periods such as the COVID-19 pandemic. We propose InflaBERT, a BERT-based LLM fine-tuned to predict inflation-related sentiment in news. We use this model to produce NEWS, an index capturing the monthly sentiment of the news regarding inflation. Incorporating our expectation index into the Cleveland Fed's model, which is only based on macroeconomic autoregressive processes, shows a marginal improvement in nowcast accuracy during the pandemic. This highlights the potential of combining sentiment analysis with traditional economic indicators, suggesting further research to refine these methodologies for better real-time inflation monitoring. The source code is available at https://github.com/paultltc/InflaBERT.
Many uncontrollable disease outbreaks of the past exposed several vulnerabilities in the healthcare systems worldwide. While advancements in technology assisted in the rapid creation of the vaccinations, there needs to be a pressing focus on the prevention and prediction of such massive outbreaks. Early detection and intervention of an outbreak can drastically reduce its impact on public health while also making the healthcare system more resilient. The complexity of disease transmission dynamics, influence of various directly and indirectly related factors and limitations of traditional approaches are the main bottlenecks in taking preventive actions. Specifically, this paper implements deep learning algorithms and LLM's to predict the severity of infectious disease outbreaks. Utilizing the historic data of several diseases that have spread in India and the climatic data spanning the past decade, the insights from our research aim to assist in creating a robust predictive system for any outbreaks in the future.
Task planning in language agents is emerging as an important research topic alongside the development of large language models (LLMs). It aims to break down complex user requests in natural language into solvable sub-tasks, thereby fulfilling the original requests. In this context, the sub-tasks can be naturally viewed as a graph, where the nodes represent the sub-tasks, and the edges denote the dependencies among them. Consequently, task planning is a decision-making problem that involves selecting a connected path or subgraph within the corresponding graph and invoking it. In this paper, we explore graph learning-based methods for task planning, a direction that is orthogonal to the prevalent focus on prompt design. Our interest in graph learning stems from a theoretical discovery: the biases of attention and auto-regressive loss impede LLMs' ability to effectively navigate decision-making on graphs, which is adeptly addressed by graph neural networks (GNNs). This theoretical insight led us to integrate GNNs with LLMs to enhance overall performance. Extensive experiments demonstrate that GNN-based methods surpass existing solutions even without training, and minimal training can further enhance their performance. The performance gain increases with a larger task graph size.
LLM watermarking, which embeds imperceptible yet algorithmically detectable signals in model outputs to identify LLM-generated text, has become crucial in mitigating the potential misuse of large language models. However, the abundance of LLM watermarking algorithms, their intricate mechanisms, and the complex evaluation procedures and perspectives pose challenges for researchers and the community to easily experiment with, understand, and assess the latest advancements. To address these issues, we introduce MarkLLM, an open-source toolkit for LLM watermarking. MarkLLM offers a unified and extensible framework for implementing LLM watermarking algorithms, while providing user-friendly interfaces to ensure ease of access. Furthermore, it enhances understanding by supporting automatic visualization of the underlying mechanisms of these algorithms. For evaluation, MarkLLM offers a comprehensive suite of 12 tools spanning three perspectives, along with two types of automated evaluation pipelines. Through MarkLLM, we aim to support researchers while improving the comprehension and involvement of the general public in LLM watermarking technology, fostering consensus and driving further advancements in research and application. Our code is available at https://github.com/THU-BPM/MarkLLM.
With the proliferation of LLM-integrated applications such as GPT-s, millions are deployed, offering valuable services through proprietary instruction prompts. These systems, however, are prone to prompt extraction attacks through meticulously designed queries. To help mitigate this problem, we introduce the Raccoon benchmark which comprehensively evaluates a model's susceptibility to prompt extraction attacks. Our novel evaluation method assesses models under both defenseless and defended scenarios, employing a dual approach to evaluate the effectiveness of existing defenses and the resilience of the models. The benchmark encompasses 14 categories of prompt extraction attacks, with additional compounded attacks that closely mimic the strategies of potential attackers, alongside a diverse collection of defense templates. This array is, to our knowledge, the most extensive compilation of prompt theft attacks and defense mechanisms to date. Our findings highlight universal susceptibility to prompt theft in the absence of defenses, with OpenAI models demonstrating notable resilience when protected. This paper aims to establish a more systematic benchmark for assessing LLM robustness against prompt extraction attacks, offering insights into their causes and potential countermeasures. Resources of Raccoon are publicly available at https://github.com/M0gician/RaccoonBench.
To better understand existing LLMs, we may examine the human mental (cognitive/psychological) architecture, and its components and structures. Based on psychological, philosophical, and cognitive science literatures, it is argued that, within the human mental architecture, existing LLMs correspond well with implicit mental processes (intuition, instinct, and so on). However, beyond such implicit processes, explicit processes (with better symbolic capabilities) are also present within the human mental architecture, judging from psychological, philosophical, and cognitive science literatures. Various theoretical and empirical issues and questions in this regard are explored. Furthermore, it is argued that existing dual-process computational cognitive architectures (models of the human cognitive/psychological architecture) provide usable frameworks for fundamentally enhancing LLMs by introducing dual processes (both implicit and explicit) and, in the meantime, can also be enhanced by LLMs. The results are synergistic combinations (in several different senses simultaneously).
Cross-lingual summarization (CLS) aims to generate a summary for the source text in a different target language. Currently, instruction-tuned large language models (LLMs) excel at various English tasks. However, unlike languages such as English, Chinese or Spanish, for those relatively low-resource languages with limited usage or data, recent studies have shown that LLMs' performance on CLS tasks remains unsatisfactory even with few-shot settings. This raises the question: Are LLMs capable of handling cross-lingual summarization tasks for low-resource languages? To resolve this question, we fully explore the potential of large language models on cross-lingual summarization task for low-resource languages through our four-step zero-shot method: Summarization, Improvement, Translation and Refinement (SITR) with correspondingly designed prompts. We test our proposed method with multiple LLMs on two well-known cross-lingual summarization datasets with various low-resource target languages. The results show that: i) GPT-3.5 and GPT-4 significantly and consistently outperform other baselines when using our zero-shot SITR methods. ii) By employing our proposed method, we unlock the potential of LLMs, enabling them to effectively handle cross-lingual summarization tasks for relatively low-resource languages.
Text classification involves categorizing a given text, such as determining its sentiment or identifying harmful content. With the advancement of large language models (LLMs), these models have become highly effective at performing text classification tasks. However, they still show vulnerabilities to variations in text formatting. Recent research demonstrates that modifying input formats, such as vertically aligning words for encoder-based models, can substantially lower accuracy in text classification tasks. While easily understood by humans, these inputs can significantly mislead models, posing a potential risk of bypassing detection in real-world scenarios involving harmful or sensitive information. With the expanding application of LLMs, a crucial question arises: Do decoder-based LLMs exhibit similar vulnerabilities to vertically formatted text input? In this paper, we investigate the impact of vertical text input on the performance of various LLMs across multiple text classification datasets and analyze the underlying causes. Our findings are as follows: (i) Vertical text input significantly degrades the accuracy of LLMs in text classification tasks. (ii) Chain of Thought (CoT) reasoning does not help LLMs recognize vertical input or mitigate its vulnerability, but few-shot learning with careful analysis does. (iii) We explore the underlying cause of the vulnerability by analyzing the inherent issues in tokenization and attention matrices.
Recent research has shown that Large Language Models (LLMs) are vulnerable to automated jailbreak attacks, where adversarial suffixes crafted by algorithms appended to harmful queries bypass safety alignment and trigger unintended responses. Current methods for generating these suffixes are computationally expensive and have low Attack Success Rates (ASR), especially against well-aligned models like Llama2 and Llama3. To overcome these limitations, we introduce ADV-LLM, an iterative self-tuning process that crafts adversarial LLMs with enhanced jailbreak ability. Our framework significantly reduces the computational cost of generating adversarial suffixes while achieving nearly 100\% ASR on various open-source LLMs. Moreover, it exhibits strong attack transferability to closed-source models, achieving 99% ASR on GPT-3.5 and 49% ASR on GPT-4, despite being optimized solely on Llama3. Beyond improving jailbreak ability, ADV-LLM provides valuable insights for future safety alignment research through its ability to generate large datasets for studying LLM safety.
Large language models (LLMs) could be valuable personal AI agents across various domains, provided they can precisely follow user instructions. However, recent studies have shown significant limitations in LLMs' instruction-following capabilities, raising concerns about their reliability in high-stakes applications. Accurately estimating LLMs' uncertainty in adhering to instructions is critical to mitigating deployment risks. We present, to our knowledge, the first systematic evaluation of the uncertainty estimation abilities of LLMs in the context of instruction-following. Our study identifies key challenges with existing instruction-following benchmarks, where multiple factors are entangled with uncertainty stems from instruction-following, complicating the isolation and comparison across methods and models. To address these issues, we introduce a controlled evaluation setup with two benchmark versions of data, enabling a comprehensive comparison of uncertainty estimation methods under various conditions. Our findings show that existing uncertainty methods struggle, particularly when models make subtle errors in instruction following. While internal model states provide some improvement, they remain inadequate in more complex scenarios. The insights from our controlled evaluation setups provide a crucial understanding of LLMs' limitations and potential for uncertainty estimation in instruction-following tasks, paving the way for more trustworthy AI agents.
Personalization has emerged as a critical research area in modern intelligent systems, focusing on mining users' behavioral history and adapting to their preferences for delivering tailored experiences. Despite the remarkable few-shot capabilities exhibited by black-box large language models (LLMs), the inherent opacity of their model parameters presents significant challenges in aligning the generated output with individual expectations. Existing solutions have primarily focused on prompt design to incorporate user-specific profiles and behaviors; however, such approaches often struggle to generalize effectively due to their inability to capture shared knowledge among all users. To address these challenges, we propose HYDRA, a model factorization framework that captures both user-specific behavior patterns from historical data and shared general knowledge among all users to deliver personalized generation. In order to capture user-specific behavior patterns, we first train a reranker to prioritize the most useful information from top-retrieved relevant historical records. By combining the prioritized history with the corresponding query, we train an adapter to align the output with individual user-specific preferences, eliminating the reliance on access to inherent model parameters of black-box LLMs. Both the reranker and the adapter can be decomposed into a base model with multiple user-specific heads, resembling a hydra. The base model maintains shared knowledge across users, while the multiple personal heads capture user-specific preferences. Experimental results demonstrate that HYDRA outperforms existing state-of-the-art prompt-based methods by an average relative improvement of 9.01% across five diverse personalization tasks in the LaMP benchmark. Our implementation is available at https://github.com/night-chen/HYDRA.
AI-based systems, including Large Language Models (LLMs), impact millions by supporting diverse tasks but face issues like misinformation, bias, and misuse. Ethical AI development is crucial as new technologies and concerns emerge, but objective, practical ethical guidance remains debated. This study examines LLMs in developing ethical AI systems, assessing how trustworthiness-enhancing techniques affect ethical AI output generation. Using the Design Science Research (DSR) method, we identify techniques for LLM trustworthiness: multi-agents, distinct roles, structured communication, and multiple rounds of debate. We design the multi-agent prototype LLM-BMAS, where agents engage in structured discussions on real-world ethical AI issues from the AI Incident Database. The prototype's performance is evaluated through thematic analysis, hierarchical clustering, ablation studies, and source code execution. Our system generates around 2,000 lines per run, compared to only 80 lines in the ablation study. Discussions reveal terms like bias detection, transparency, accountability, user consent, GDPR compliance, fairness evaluation, and EU AI Act compliance, showing LLM-BMAS's ability to generate thorough source code and documentation addressing often-overlooked ethical AI issues. However, practical challenges in source code integration and dependency management may limit smooth system adoption by practitioners. This study aims to shed light on enhancing trustworthiness in LLMs to support practitioners in developing ethical AI-based systems.
We provide a sober look at the application of Multimodal Large Language Models (MLLMs) in autonomous driving, challenging common assumptions about their ability to interpret dynamic driving scenarios. Despite advances in models like GPT-4o, their performance in complex driving environments remains largely unexplored. Our experimental study assesses various MLLMs as world models using in-car camera perspectives and reveals that while these models excel at interpreting individual images, they struggle to synthesize coherent narratives across frames, leading to considerable inaccuracies in understanding (i) ego vehicle dynamics, (ii) interactions with other road actors, (iii) trajectory planning, and (iv) open-set scene reasoning. We introduce the Eval-LLM-Drive dataset and DriveSim simulator to enhance our evaluation, highlighting gaps in current MLLM capabilities and the need for improved models in dynamic real-world environments.
Large Language Models (LLMs) have recently shown great promise in planning and reasoning applications. These tasks demand robust systems, which arguably require a causal understanding of the environment. While LLMs can acquire and reflect common sense causal knowledge from their pretraining data, this information is often incomplete, incorrect, or inapplicable to a specific environment. In contrast, causal representation learning (CRL) focuses on identifying the underlying causal structure within a given environment. We propose a framework that integrates CRLs with LLMs to enable causally-aware reasoning and planning. This framework learns a causal world model, with causal variables linked to natural language expressions. This mapping provides LLMs with a flexible interface to process and generate descriptions of actions and states in text form. Effectively, the causal world model acts as a simulator that the LLM can query and interact with. We evaluate the framework on causal inference and planning tasks across temporal scales and environmental complexities. Our experiments demonstrate the effectiveness of the approach, with the causally-aware method outperforming LLM-based reasoners, especially for longer planning horizons.
LLMs are seeing growing use for applications which require large context windows, and with these large context windows KV cache activations surface as the dominant contributor to memory consumption during inference. Quantization is a promising approach for compressing KV cache activations; however, existing solutions fail to represent activations accurately in sub-4-bit precision. Our work, KVQuant, facilitates low precision KV cache quantization by incorporating several novel methods: (i) Per-Channel Key Quantization, where we adjust the dimension along which we quantize the Key activations to better match the distribution; (ii) Pre-RoPE Key Quantization, where we quantize Key activations before the rotary positional embedding to mitigate its impact on quantization; (iii) Non-Uniform KV Cache Quantization, where we derive per-layer sensitivity-weighted non-uniform datatypes that better represent the distributions; and (iv) Per-Vector Dense-and-Sparse Quantization, where we isolate outliers separately for each vector to minimize skews in quantization ranges. By applying our method to the LLaMA, Llama-2, Llama-3, and Mistral models, we achieve < 0.1 perplexity degradation with 3-bit quantization on both Wikitext-2 and C4, outperforming existing approaches. Our method enables serving LLaMA-7B with a context length of up to 1 million on a single A100-80GB GPU and up to 10 million on an 8-GPU system. We develop custom CUDA kernels for KVQuant, showing that we can achieve up to ~1.7x speedups, compared to baseline fp16 matrix-vector multiplications, for the LLaMA-7B model.
Large Language Models (LLMs) have the potential to automate reward engineering by leveraging their broad domain knowledge across various tasks. However, they often need many iterations of trial-and-error to generate effective reward functions. This process is costly because evaluating every sampled reward function requires completing the full policy optimization process for each function. In this paper, we introduce an LLM-driven reward generation framework that is able to produce state-of-the-art policies on the challenging Bi-DexHands benchmark with 20x fewer reward function samples than the prior state-of-the-art work. Our key insight is that we reduce the problem of generating task-specific rewards to the problem of coarsely estimating task progress. Our two-step solution leverages the task domain knowledge and the code synthesis abilities of LLMs to author progress functions that estimate task progress from a given state. Then, we use this notion of progress to discretize states, and generate count-based intrinsic rewards using the low-dimensional state space. We show that the combination of LLM-generated progress functions and count-based intrinsic rewards is essential for our performance gains, while alternatives such as generic hash-based counts or using progress directly as a reward function fall short.
This study explores the ability of Large Language Model (LLM) agents to detect and correct hallucinations in AI-generated content. A primary agent was tasked with creating a blog about a fictional Danish artist named Flipfloppidy, which was then reviewed by another agent for factual inaccuracies. Most LLMs hallucinated the existence of this artist. Across 4,900 test runs involving various combinations of primary and reviewing agents, advanced AI models such as Llama3-70b and GPT-4 variants demonstrated near-perfect accuracy in identifying hallucinations and successfully revised outputs in 85% to 100% of cases following feedback. These findings underscore the potential of advanced AI models to significantly enhance the accuracy and reliability of generated content, providing a promising approach to improving AI workflow orchestration.
Fine-tuning Large Language Models (LLMs) has become a crucial technique for adapting pre-trained models to downstream tasks. However, the enormous size of LLMs poses significant challenges in terms of computational complexity and resource requirements. Low-Rank Adaptation (LoRA) has emerged as a promising solution. However, there exists a gap between the practical performance of low-rank adaptations and its theoretical optimum. In this work, we propose eXtreme Gradient Boosting LoRA (XGBLoRA), a novel framework that bridges this gap by leveraging the power of ensemble learning. Inspired by gradient boosting, XGBLoRA iteratively learns and merges a sequence of LoRA adaptations to refine model predictions. It achieves better performance than the standard LoRA, while enjoying the computational efficiency of rank-1 adaptations. We provide theoretical analysis to show the convergence and optimality of our approach, and conduct extensive experiments on a range of natural language processing tasks. The results demonstrate that XGBLoRA consistently outperforms standard LoRA and achieves performance comparable to full fine-tuning with significantly fewer trainable parameters. This work advances parameter-efficient fine-tuning for LLMs, and offers a promising solution for adapting LLMs to downstream tasks while optimizing performance and efficiency.
Generating diverse and effective clarifying questions is crucial for improving query understanding and retrieval performance in open-domain conversational search (CS) systems. We propose AGENT-CQ (Automatic GENeration, and evaluaTion of Clarifying Questions), an end-to-end LLM-based framework addressing the challenges of scalability and adaptability faced by existing methods that rely on manual curation or template-based approaches. AGENT-CQ consists of two stages: a generation stage employing LLM prompting strategies to generate clarifying questions, and an evaluation stage (CrowdLLM) that simulates human crowdsourcing judgments using multiple LLM instances to assess generated questions and answers based on comprehensive quality metrics. Extensive experiments on the ClariQ dataset demonstrate CrowdLLM's effectiveness in evaluating question and answer quality. Human evaluation and CrowdLLM show that the AGENT-CQ - generation stage, consistently outperforms baselines in various aspects of question and answer quality. In retrieval-based evaluation, LLM-generated questions significantly enhance retrieval effectiveness for both BM25 and cross-encoder models compared to human-generated questions.
Home robots performing personalized tasks must adeptly balance user preferences with environmental affordances. We focus on organization tasks within constrained spaces, such as arranging items into a refrigerator, where preferences for placement collide with physical limitations. The robot must infer user preferences based on a small set of demonstrations, which is easier for users to provide than extensively defining all their requirements. While recent works use Large Language Models (LLMs) to learn preferences from user demonstrations, they encounter two fundamental challenges. First, there is inherent ambiguity in interpreting user actions, as multiple preferences can often explain a single observed behavior. Second, not all user preferences are practically feasible due to geometric constraints in the environment. To address these challenges, we introduce APRICOT, a novel approach that merges LLM-based Bayesian active preference learning with constraint-aware task planning. APRICOT refines its generated preferences by actively querying the user and dynamically adapts its plan to respect environmental constraints. We evaluate APRICOT on a dataset of diverse organization tasks and demonstrate its effectiveness in real-world scenarios, showing significant improvements in both preference satisfaction and plan feasibility. The project website is at https://portal-cornell.github.io/apricot/
We present CoqPilot, a VS Code extension designed to help automate writing of Coq proofs. The plugin collects the parts of proofs marked with the admit tactic in a Coq file, i.e., proof holes, and combines LLMs along with non-machine-learning methods to generate proof candidates for the holes. Then, CoqPilot checks if each proof candidate solves the given subgoal and, if successful, replaces the hole with it. The focus of CoqPilot is twofold. Firstly, we want to allow users to seamlessly combine multiple Coq generation approaches and provide a zero-setup experience for our tool. Secondly, we want to deliver a platform for LLM-based experiments on Coq proof generation. We developed a benchmarking system for Coq generation methods, available in the plugin, and conducted an experiment using it, showcasing the framework's possibilities. Demo of CoqPilot is available at: https://youtu.be/oB1Lx-So9Lo. Code at: https://github.com/JetBrains-Research/coqpilot
Large language models (LLMs) have shown remarkable capabilities, but still struggle with processing extensive contexts, limiting their ability to maintain coherence and accuracy over long sequences. In contrast, the human brain excels at organising and retrieving episodic experiences across vast temporal scales, spanning a lifetime. In this work, we introduce EM-LLM, a novel approach that integrates key aspects of human episodic memory and event cognition into LLMs with no fine-tuning, enabling them to handle practically infinite context lengths while maintaining computational efficiency. EM-LLM organises sequences of tokens into coherent episodic events using a combination of Bayesian surprise and graph-theoretic boundary refinement in an online fashion. When needed, these events are retrieved through a two-stage memory process, combining similarity-based and temporally contiguous retrieval for efficient and human-like access to relevant information. Experiments on the LongBench and InfiniteBench benchmarks demonstrate EM-LLM's superior performance, consistently outperforming the state-of-the-art retrieval model InfLLM across various baseline LLMs. In addition, EM-LLM outperforms its popular counterpart, RAG, in a wide range of tasks, while requiring similar resources. Notably, EM-LLM's performance even surpasses full-context models in most tasks, while successfully performing retrieval across 10 million tokens - a scale computationally infeasible for such models. Finally, our analysis reveals strong correlations between EM-LLM's event segmentation and human-perceived events, suggesting a bridge between this artificial system and its biological counterpart, thereby offering a novel computational framework for exploring human memory mechanisms.
The Federal Funds rate in the United States plays a significant role in both domestic and international financial markets. However, research has predominantly focused on the effects of adjustments to the Federal Funds rate rather than on the decision-making process itself. Recent advancements in large language models(LLMs) offer a potential method for reconstructing the original FOMC meetings, which are responsible for setting the Federal Funds rate. In this paper, we propose a five-stage FOMC meeting simulation framework, MiniFed, which employs LLM agents to simulate real-world FOMC meeting members and optimize the FOMC structure. This framework effectively revitalizes the FOMC meeting process and facilitates projections of the Federal Funds rate. Experimental results demonstrate that our proposed MiniFed framework achieves both high accuracy in Federal Funds rate projections and behavioral alignment with the agents' real-world counterparts. Given that few studies have focused on employing LLM agents to simulate large-scale real-world conferences, our work can serve as a benchmark for future developments.
Large language models (LLMs) can store a significant amount of factual knowledge in their parameters. However, their parametric knowledge may conflict with the information provided in the context -- this phenomenon, known as \emph{context-memory knowledge conflicts}, can lead to undesirable model behaviour, such as reliance on outdated or incorrect information. Analysing the internal activations of LLMs, we find that they can internally register the signals of knowledge conflict at mid-layers. Such signals allow us to detect whether a knowledge conflict occurs and use \emph{inference-time} intervention strategies to resolve it. In this work, we propose \textsc{SpARE}, a \emph{training-free} representation engineering method that uses pre-trained sparse auto-encoders (SAEs) to control the knowledge selection behaviour of LLMs. \textsc{SpARE} identifies the functional features that control the knowledge selection behaviours and applies them to edit the internal activations of LLMs at inference time. Our experimental results show that \textsc{SpARE} can effectively control the usage of either knowledge source to resolve knowledge conflict in open-domain question-answering tasks, surpassing existing representation engineering methods ($+10\%$) as well as contrastive decoding methods ($+15\%$).
Retrieval-augmented generation (RAG) is a key means to effectively enhance large language models (LLMs) in many knowledge-based tasks. However, existing RAG methods struggle with knowledge-intensive reasoning tasks, because useful information required to these tasks are badly scattered. This characteristic makes it difficult for existing RAG methods to accurately identify key information and perform global reasoning with such noisy augmentation. In this paper, motivated by the cognitive theories that humans convert raw information into various structured knowledge when tackling knowledge-intensive reasoning, we proposes a new framework, StructRAG, which can identify the optimal structure type for the task at hand, reconstruct original documents into this structured format, and infer answers based on the resulting structure. Extensive experiments across various knowledge-intensive tasks show that StructRAG achieves state-of-the-art performance, particularly excelling in challenging scenarios, demonstrating its potential as an effective solution for enhancing LLMs in complex real-world applications.
Large language models (LLMs) have achieved a degree of success in generating coherent and contextually relevant text, yet they remain prone to a significant challenge known as hallucination: producing information that is not substantiated by the input or external knowledge. Previous efforts to mitigate hallucinations have focused on techniques such as fine-tuning models on high-quality datasets, incorporating fact-checking mechanisms, and developing adversarial training methods. While these approaches have shown some promise, they often address the issue at the level of individual model outputs, leaving unexplored the effects of inter-model interactions on hallucination. This study investigates the phenomenon of hallucination in LLMs through a novel experimental framework where multiple instances of GPT-4o-Mini models engage in a debate-like interaction prompted with questions from the TruthfulQA dataset. One model is deliberately instructed to generate plausible but false answers while the other models are asked to respond truthfully. The experiment is designed to assess whether the introduction of misinformation by one model can challenge the truthful majority to better justify their reasoning, improving performance on the TruthfulQA benchmark. The findings suggest that inter-model interactions can offer valuable insights into improving the accuracy and robustness of LLM outputs, complementing existing mitigation strategies.
Large language models (LLMs) are being used in data science code generation tasks, but they often struggle with complex sequential tasks, leading to logical errors. Their application to geospatial data processing is particularly challenging due to difficulties in incorporating complex data structures and spatial constraints, effectively utilizing diverse function calls, and the tendency to hallucinate less-used geospatial libraries. To tackle these problems, we introduce GeoAgent, a new interactive framework designed to help LLMs handle geospatial data processing more effectively. GeoAgent pioneers the integration of a code interpreter, static analysis, and Retrieval-Augmented Generation (RAG) techniques within a Monte Carlo Tree Search (MCTS) algorithm, offering a novel approach to geospatial data processing. In addition, we contribute a new benchmark specifically designed to evaluate the LLM-based approach in geospatial tasks. This benchmark leverages a variety of Python libraries and includes both single-turn and multi-turn tasks such as data acquisition, data analysis, and visualization. By offering a comprehensive evaluation among diverse geospatial contexts, this benchmark sets a new standard for developing LLM-based approaches in geospatial data analysis tasks. Our findings suggest that relying solely on knowledge of LLM is insufficient for accurate geospatial task programming, which requires coherent multi-step processes and multiple function calls. Compared to the baseline LLMs, the proposed GeoAgent has demonstrated superior performance, yielding notable improvements in function calls and task completion. In addition, these results offer valuable insights for the future development of LLM agents in automatic geospatial data analysis task programming.
An exposition of the mathematics underpinning the neural network architecture of a GPT-3-style LLM.
Direct Preference Optimization (DPO) has emerged as a more computationally efficient alternative to Reinforcement Learning from Human Feedback (RLHF) with Proximal Policy Optimization (PPO), eliminating the need for reward models and online sampling. Despite these benefits, DPO and its variants remain sensitive to hyper-parameters and prone to instability, particularly on mathematical datasets. We argue that these issues arise from the unidirectional likelihood-derivative negative feedback inherent in the log-likelihood loss function. To address this, we propose a novel LLM alignment loss that establishes a stable Bidirectional Negative Feedback (BNF) during optimization. Our proposed BNF loss eliminates the need for pairwise contrastive losses and does not require any extra tunable hyper-parameters or pairwise preference data, streamlining the alignment pipeline to be as simple as supervised fine-tuning. We conduct extensive experiments across two challenging QA benchmarks and four reasoning benchmarks. The experimental results show that BNF achieves comparable performance to the best methods on QA benchmarks, while its performance decrease on the four reasoning benchmarks is significantly lower compared to the best methods, thus striking a better balance between value alignment and reasoning ability. In addition, we further validate the performance of BNF on non-pairwise datasets, and conduct in-depth analysis of log-likelihood and logit shifts across different preference optimization methods.
Large Language Models (LLMs), trained on a large amount of corpus, have demonstrated remarkable abilities. However, it may not be sufficient to directly apply open-source LLMs like Llama to certain real-world scenarios, since most of them are trained for \emph{general} purposes. Thus, the demands for customizing publicly available LLMs emerge, but are currently under-studied. In this work, we consider customizing pre-trained LLMs with new human preferences. Specifically, the LLM should not only meet the new preference but also preserve its original capabilities after customization. Drawing inspiration from the observation that human preference can be expressed as a reward model, we propose to cast LLM customization as optimizing the sum of two reward functions, one of which (denoted as $r_1$) was used to pre-train the LLM while the other (denoted as $r_2$) characterizes the new human preference. The obstacle here is that both reward functions are unknown, making the application of modern reinforcement learning methods infeasible. Thanks to the residual Q-learning framework, we can restore the customized LLM with the pre-trained LLM and the \emph{residual Q-function} without the reward function $r_1$. Moreover, we find that for a fixed pre-trained LLM, the reward function $r_2$ can be derived from the residual Q-function, enabling us to directly learn the residual Q-function from the new human preference data upon the Bradley-Terry model. We name our method Q-Adapter as it introduces an adapter module to approximate the residual Q-function for customizing the pre-trained LLM towards the new preference. Experiments based on the Llama-3.1 model on the DSP dataset and HH-RLHF dataset illustrate the superior effectiveness of Q-Adapter on both retaining existing knowledge and learning new preferences. Code is available at \url{https://github.com/mansicer/Q-Adapter}.
The growing use of large language model (LLM)-based chatbots has raised concerns about fairness. Fairness issues in LLMs can lead to severe consequences, such as bias amplification, discrimination, and harm to marginalized communities. While existing fairness benchmarks mainly focus on single-turn dialogues, multi-turn scenarios, which in fact better reflect real-world conversations, present greater challenges due to conversational complexity and potential bias accumulation. In this paper, we propose a comprehensive fairness benchmark for LLMs in multi-turn dialogue scenarios, \textbf{FairMT-Bench}. Specifically, we formulate a task taxonomy targeting LLM fairness capabilities across three stages: context understanding, user interaction, and instruction trade-offs, with each stage comprising two tasks. To ensure coverage of diverse bias types and attributes, we draw from existing fairness datasets and employ our template to construct a multi-turn dialogue dataset, \texttt{FairMT-10K}. For evaluation, GPT-4 is applied, alongside bias classifiers including Llama-Guard-3 and human validation to ensure robustness. Experiments and analyses on \texttt{FairMT-10K} reveal that in multi-turn dialogue scenarios, current LLMs are more likely to generate biased responses, and there is significant variation in performance across different tasks and models. Based on this, we curate a challenging dataset, \texttt{FairMT-1K}, and test 15 current state-of-the-art (SOTA) LLMs on this dataset. The results show the current state of fairness in LLMs and showcase the utility of this novel approach for assessing fairness in more realistic multi-turn dialogue contexts, calling for future work to focus on LLM fairness improvement and the adoption of \texttt{FairMT-1K} in such efforts.
Recent studies have identified one aggravating factor of LLM hallucinations as the knowledge inconsistency between pre-training and fine-tuning, where unfamiliar fine-tuning data mislead the LLM to fabricate plausible but wrong outputs. In this paper, we propose a novel fine-tuning strategy called Prereq-Tune to address this knowledge inconsistency and reduce hallucinations. Fundamentally, Prereq-Tune disentangles the learning of skills and knowledge, so the model learns only the task skills without being impacted by the knowledge inconsistency. To achieve this, Prereq-Tune introduces an additional prerequisite learning stage to learn the necessary knowledge for SFT, allowing subsequent SFT to focus only on task skills. Prereq-Tune can also be combined with fictitious synthetic data to enhance the grounding of LLM outputs to their internal knowledge. Experiments show that Prereq-Tune outperforms existing baselines in improving LLM's factuality across short QA and long-form generation tasks. It also opens new possibilities for knowledge-controlled generation in LLMs. Our code is available at https://github.com/UCSB-NLP-Chang/Prereq_tune.git.
With the rise of AI-generated content spewed at scale from large language models (LLMs), genuine concerns about the spread of fake news have intensified. The perceived ability of LLMs to produce convincing fake news at scale poses new challenges for both human and automated fake news detection systems. To address this gap, this work presents the findings from a university-level competition which aimed to explore how LLMs can be used by humans to create fake news, and to assess the ability of human annotators and AI models to detect it. A total of 110 participants used LLMs to create 252 unique fake news stories, and 84 annotators participated in the detection tasks. Our findings indicate that LLMs are ~68% more effective at detecting real news than humans. However, for fake news detection, the performance of LLMs and humans remains comparable (~60% accuracy). Additionally, we examine the impact of visual elements (e.g., pictures) in news on the accuracy of detecting fake news stories. Finally, we also examine various strategies used by fake news creators to enhance the credibility of their AI-generated content. This work highlights the increasing complexity of detecting AI-generated fake news, particularly in collaborative human-AI settings.
In the field of automated programming, large language models (LLMs) have demonstrated foundational generative capabilities when given detailed task descriptions. However, their current functionalities are primarily limited to function-level development, restricting their effectiveness in complex project environments and specific application scenarios, such as complicated image-processing tasks. This paper presents a multi-agent framework that utilises a hybrid set of LLMs, including GPT-4o and locally deployed open-source models, which collaboratively complete auto-programming tasks. Each agent plays a distinct role in the software development cycle, collectively forming a virtual organisation that works together to produce software products. By establishing a tree-structured thought distribution and development mechanism across project, module, and function levels, this framework offers a cost-effective and efficient solution for code generation. We evaluated our approach using benchmark datasets, and the experimental results demonstrate that VisionCoder significantly outperforms existing methods in image processing auto-programming tasks.
This research introduces a novel methodology for assigning quantifiable, controllable and psychometrically validated personalities to Large Language Models-Based Agents (Agents) using the Big Five personality framework. It seeks to overcome the constraints of human subject studies, proposing Agents as an accessible tool for social science inquiry. Through a series of four studies, this research demonstrates the feasibility of assigning psychometrically valid personality traits to Agents, enabling them to replicate complex human-like behaviors. The first study establishes an understanding of personality constructs and personality tests within the semantic space of an LLM. Two subsequent studies -- using empirical and simulated data -- illustrate the process of creating Agents and validate the results by showing strong correspondence between human and Agent answers to personality tests. The final study further corroborates this correspondence by using Agents to replicate known human correlations between personality traits and decision-making behaviors in scenarios involving risk-taking and ethical dilemmas, thereby validating the effectiveness of the psychometric approach to design Agents and its applicability to social and behavioral research.
Recent advances in large language models (LLMs) have shown promise for scalable educational applications, but their use in dialog-based tutoring systems remains challenging due to the need for effective pedagogical strategies and the high costs associated with expert-curated datasets. Our study explores the use of smaller, more affordable LLMs for one-on-one tutoring in the context of solving reading comprehension problems. We developed a synthetic tutoring dialog dataset, evaluated by human teachers, and fine-tuned a smaller LLM using this dataset. Furthermore, we conducted an interactive experiment comparing the performance of the fine-tuned model with a larger model in real-world tutoring scenarios. Our results show that the fine-tuned model performs on par with the larger model but at a lower cost, demonstrating a viable, cost-effective approach for implementing LLM-based tutoring systems in educational settings.
The advent of large language models (LLMs) has revolutionized the field of text generation, producing outputs that closely mimic human-like writing. Although academic and industrial institutions have developed detectors to prevent the malicious usage of LLM-generated texts, other research has doubt about the robustness of these systems. To stress test these detectors, we introduce a proxy-attack strategy that effortlessly compromises LLMs, causing them to produce outputs that align with human-written text and mislead detection systems. Our method attacks the source model by leveraging a reinforcement learning (RL) fine-tuned humanized small language model (SLM) in the decoding phase. Through an in-depth analysis, we demonstrate that our attack strategy is capable of generating responses that are indistinguishable to detectors, preventing them from differentiating between machine-generated and human-written text. We conduct systematic evaluations on extensive datasets using proxy-attacked open-source models, including Llama2-13B, Llama3-70B, and Mixtral-8*7B in both white- and black-box settings. Our findings show that the proxy-attack strategy effectively deceives the leading detectors, resulting in an average AUROC drop of 70.4% across multiple datasets, with a maximum drop of 90.3% on a single dataset. Furthermore, in cross-discipline scenarios, our strategy also bypasses these detectors, leading to a significant relative decrease of up to 90.9%, while in cross-language scenario, the drop reaches 91.3%. Despite our proxy-attack strategy successfully bypassing the detectors with such significant relative drops, we find that the generation quality of the attacked models remains preserved, even within a modest utility budget, when compared to the text produced by the original, unattacked source model.
Narratives are widely recognized as a powerful tool for structuring information and facilitating comprehension of complex ideas in various domains such as science communication. This paper investigates whether incorporating narrative elements can assist Large Language Models (LLMs) in solving complex problems more effectively. We propose a novel approach, Story of Thought (SoT), integrating narrative structures into prompting techniques for problem-solving. This approach involves constructing narratives around problem statements and creating a framework to identify and organize relevant information. Our experiments show that using various LLMs with SoT consistently surpasses using them with other techniques on physics, chemistry, math, and biology questions in both the GPQA and JEEBench datasets. The narrative-based information curation process in SoT enhances problem comprehension by contextualizing critical in-domain information and highlighting causal relationships within the problem space.
Aligning Large Language Models (LLM) to address subjectivity and nuanced preference levels requires adequate flexibility and control, which can be a resource-intensive and time-consuming procedure. Existing training-time alignment methods require full re-training when a change is needed and inference-time ones typically require access to the reward model at each inference step. To address these limitations, we introduce inference-time model alignment method that learns encoded representations of preference dimensions, called \textit{Alignment Vectors} (AV). These representations are computed by subtraction of the base model from the aligned model as in model editing enabling dynamically adjusting the model behavior during inference through simple linear operations. Even though the preference dimensions can span various granularity levels, here we focus on three gradual response levels across three specialized domains: medical, legal, and financial, exemplifying its practical potential. This new alignment paradigm introduces adjustable preference knobs during inference, allowing users to tailor their LLM outputs while reducing the inference cost by half compared to the prompt engineering approach. Additionally, we find that AVs are transferable across different fine-tuning stages of the same model, demonstrating their flexibility. AVs also facilitate multidomain, diverse preference alignment, making the process 12x faster than the retraining approach.
This work investigates the in-context learning abilities of pretrained large language models (LLMs) when instructed to translate text from a low-resource language into a high-resource language as part of an automated machine translation pipeline. We conduct a set of experiments translating Southern Quechua to Spanish and examine the informativity of various types of context retrieved from a constrained database of digitized pedagogical materials (dictionaries and grammar lessons) and parallel corpora. Using both automatic and human evaluation of model output, we conduct ablation studies that manipulate (1) context type (morpheme translations, grammar descriptions, and corpus examples), (2) retrieval methods (automated vs. manual), and (3) model type. Our results suggest that even relatively small LLMs are capable of utilizing prompt context for zero-shot low-resource translation when provided a minimally sufficient amount of relevant linguistic information. However, the variable effects of context type, retrieval method, model type, and language-specific factors highlight the limitations of using even the best LLMs as translation systems for the majority of the world's 7,000+ languages and their speakers.
The proliferation of large language models (LLMs) has led to the adoption of Mixture-of-Experts (MoE) architectures that dynamically leverage specialized subnetworks for improved efficiency and performance. Despite their benefits, MoE models face significant challenges during inference, including inefficient memory management and suboptimal batching, due to misaligned design choices between the model architecture and the system policies. Furthermore, the conventional approach of training MoEs from scratch is increasingly prohibitive in terms of cost. In this paper, we propose a novel framework Read-ME that transforms pre-trained dense LLMs into smaller MoE models (in contrast to "upcycling" generalist MoEs), avoiding the high costs of ground-up training. Our approach employs activation sparsity to extract experts. To compose experts, we examine the widely-adopted layer-wise router design and show its redundancy, and thus we introduce the pre-gating router decoupled from the MoE backbone that facilitates system-friendly pre-computing and lookahead scheduling, enhancing expert-aware batching and caching. Our codesign therefore addresses critical gaps on both the algorithmic and system fronts, establishing a scalable and efficient alternative for LLM inference in resource-constrained settings. Read-ME outperforms other popular open-source dense models of similar scales, achieving improvements of up to 10.1% on MMLU, and improving mean end-to-end latency up to 6.1%. Codes are available at: https://github.com/VITA-Group/READ-ME.
This project aims to investigate a novel sequence generation method inspired by the AlphaGo paradigm, adapting it for use with large language models (LLMs). The proposed approach involves creating search trees of different possible completions and evaluating these completions based on model confidence. By considering various paths in the search tree and scoring them according to the model's confidence in each completion, we can generate diverse and high-quality sequences. This research explores the implementation of this paradigm by using confidence as a proxy for response quality akin to beam search \citep{vijayakumar2016diverse}. The primary goal of this paper is to outline the paradigm and demonstrate its potential, rather than focusing on achieving perfect results. The paper will outline the reasons why we believe this paradigm has the potential to improve LLMs in the following manners: 1) increase output quality, 2) decrease errors, 3) eliminate or reduce the compound error problems, 4) generate diverse and creative completions, 5) allow for iterative problem-solving, and 6) self-training. We expect this approach to yield a set of diverse and coherent sequences, offering insights into balancing exploration and exploitation in sequence generation. Potential applications include creative text generation tasks, such as storytelling and content creation, as well as other natural language processing domains, like machine translation and automated summarization. The goal is that the model will be far more effective as it will be able to consider many possible variations allowing it to find the ideal completion. This research aims to contribute to the understanding of effective search strategies in sequence generation and their impact on generating high-quality, varied textual outputs.
Large language models (LLMs) have revolutionized natural language processing, albeit at the cost of immense memory and computation requirements. Post-training quantization (PTQ) is becoming the de facto method to reduce the memory footprint and improve the inference throughput of LLMs. In this work, we aim to push the upper limit of LLM PTQ by optimizing the weight rounding parameters with the block reconstruction technique, a predominant method in previous vision models. We propose TesseraQ, a new state-of-the-art PTQ technique, to quantize the weights of LLMs to ultra-low bits. To effectively optimize the rounding in LLMs and stabilize the reconstruction process, we introduce progressive adaptive rounding. This approach iteratively transits the soft rounding variables to hard variables during the reconstruction process. Additionally, we optimize the dequantization scale parameters to fully leverage the block reconstruction technique. We demonstrate that TesseraQ can be seamlessly integrated with existing scaling or clipping-based PTQ algorithms such as AWQ and OmniQuant, significantly enhancing their performance and establishing a new state-of-the-art. For instance, when compared to AWQ, TesseraQ improves the wikitext2 perplexity from 14.65 to 6.82 and average downstream accuracy from 50.52 to 59.27 with 2-bit weight-only quantization of LLaMA-2-7B. Across a range of quantization schemes, including W2A16, W3A16, W3A3, and W4A4, TesseraQ consistently exhibits superior performance.
With the growing spread of misinformation online, research has increasingly focused on detecting and tracking fake news. However, an overlooked issue is that fake news does not naturally exist in social networks -- it often originates from distorted facts or deliberate fabrication by malicious actors. Understanding how true news gradually evolves into fake news is critical for early detection and prevention, reducing its spread and impact. Hence, in this paper, we take the first step toward simulating and revealing this evolution, proposing a Fake News evolUtion Simulation framEwork (FUSE) based on large language models (LLMs). Specifically, we employ LLM as agents to represent individuals in a simulated social network. We define four types of agents commonly observed in daily interactions: spreaders, who propagate information; commentators, who provide opinions and interpretations; verifiers, who check the accuracy of information; and bystanders, who passively observe without engaging. For simulated environments, we model various social network structures, such as high-clustering networks and scale-free networks, to mirror real-world network dynamics. Each day, the agents engage in belief exchanges, reflect on their thought processes, and reintroduce the news accordingly. Given the lack of prior work in this area, we developed a FUSE-EVAL evaluation framework to measure the deviation from true news during the fake news evolution process. The results show that FUSE successfully captures the underlying patterns of how true news transforms into fake news and accurately reproduces previously discovered instances of fake news, aligning closely with human evaluations. Moreover, our work provides insights into the fact that combating fake news should not be delayed until it has fully evolved; instead, prevention in advance is key to achieving better outcomes.
Target volume contouring for radiation therapy is considered significantly more challenging than the normal organ segmentation tasks as it necessitates the utilization of both image and text-based clinical information. Inspired by the recent advancement of large language models (LLMs) that can facilitate the integration of the textural information and images, here we present a novel LLM-driven multimodal AI, namely LLMSeg, that utilizes the clinical text information and is applicable to the challenging task of target volume contouring for radiation therapy, and validate it within the context of breast cancer radiation therapy target volume contouring. Using external validation and data-insufficient environments, which attributes highly conducive to real-world applications, we demonstrate that the proposed model exhibits markedly improved performance compared to conventional unimodal AI models, particularly exhibiting robust generalization performance and data efficiency. To our best knowledge, this is the first LLM-driven multimodal AI model that integrates the clinical text information into target volume delineation for radiation oncology.
Large language models (LLMs) have demonstrated great performance across various benchmarks, showing potential as general-purpose task solvers. However, as LLMs are typically trained on vast amounts of data, a significant concern in their evaluation is data contamination, where overlap between training data and evaluation datasets inflates performance assessments. While multiple approaches have been developed to identify data contamination, these approaches rely on specific assumptions that may not hold universally across different settings. To bridge this gap, we systematically review 47 papers on data contamination detection, categorize the underlying assumptions, and assess whether they have been rigorously validated. We identify and analyze eight categories of assumptions and test three of them as case studies. Our analysis reveals that when classifying instances used for pretraining LLMs, detection approaches based on these three assumptions perform close to random guessing, suggesting that current LLMs learn data distributions rather than memorizing individual instances. Overall, this work underscores the importance of approaches clearly stating their underlying assumptions and testing their validity across various scenarios.
Large Language Models (LLMs) demonstrate strong proficiency in generating code for high-resource programming languages (HRPLs) like Python but struggle significantly with low-resource programming languages (LRPLs) such as Racket or D. This performance gap deepens the digital divide, preventing developers using LRPLs from benefiting equally from LLM advancements and reinforcing disparities in innovation within underrepresented programming communities. While generating additional training data for LRPLs is promising, it faces two key challenges: manual annotation is labor-intensive and costly, and LLM-generated LRPL code is often of subpar quality. The underlying cause of this issue is the gap between natural language to programming language gap (NL-PL Gap), which is especially pronounced in LRPLs due to limited aligned data. In this work, we introduce a novel approach called Bridge-Coder, which leverages LLMs' intrinsic capabilities to enhance the performance on LRPLs. Our method consists of two key stages. Bridge Generation, where we create high-quality dataset by utilizing LLMs' general knowledge understanding, proficiency in HRPLs, and in-context learning abilities. Then, we apply the Bridged Alignment, which progressively improves the alignment between NL instructions and LRPLs. Experimental results across multiple LRPLs show that Bridge-Coder significantly enhances model performance, demonstrating the effectiveness and generalization of our approach. Furthermore, we offer a detailed analysis of the key components of our method, providing valuable insights for future work aimed at addressing the challenges associated with LRPLs.
Social biases can manifest in language agency. While several studies approached agency-related bias in human-written language, very limited research has investigated such biases in Large Language Model (LLM)-generated content. In addition, previous works often rely on string-matching techniques to identify agentic and communal words within texts, which fall short of accurately classifying language agency. We introduce the novel Language Agency Bias Evaluation (LABE) benchmark, which comprehensively evaluates biases in LLMs by analyzing agency levels attributed to different demographic groups in model generations. LABE leverages 5,400 template-based prompts, an accurate agency classifier, and corresponding bias metrics to test for gender, racial, and intersectional language agency biases in LLMs on 3 text generation tasks: biographies, professor reviews, and reference letters. We also contribute the Language Agency Classification (LAC) dataset, consisting of 3,724 agentic and communal sentences. Using LABE, we unveil language agency social biases in 3 recent LLMs: ChatGPT, Llama3, and Mistral. We observe that: (1) LLM generations tend to demonstrate greater gender bias than human-written texts; (2) Models demonstrate remarkably higher levels of intersectional bias than the other bias aspects. Those who are at the intersection of gender and racial minority groups--such as Black females--are consistently described by texts with lower levels of agency, aligning with real-world social inequalities; (3) Among the 3 LLMs investigated, Llama3 demonstrates the greatest overall bias; (4) Not only does prompt-based mitigation fail to resolve language agency bias in LLMs, but it frequently leads to the exacerbation of biases in generated texts.
Consider the math problem: "Lily received 3 cookies from her best friend yesterday and ate 5 for breakfast. Today, her friend gave her 3 more cookies. How many cookies does Lily have now?" Many large language models (LLMs) in previous research approach this problem by calculating the answer "1" using the equation "3 - 5 + 3." However, from a human perspective, we recognize the inherent flaw in this problem: Lily cannot eat 5 cookies if she initially only had 3. This discrepancy prompts a key question: Are current LLMs merely Blind Solver that apply mathematical operations without deeper reasoning, or can they function as Logical Thinker capable of identifying logical inconsistencies? To explore this question, we propose a benchmark dataset, FaultyMath, which includes faulty math problems of rich diversity: i) multiple mathematical categories, e.g., algebra, geometry, number theory, etc., ii) varying levels of difficulty, and iii) different origins of faultiness -- ranging from violations of common sense and ambiguous statements to mathematical contradictions and more. We evaluate a broad spectrum of LLMs, including open-source, closed-source, and math-specialized models, using FaultyMath across three dimensions: (i) How accurately can the models detect faulty math problems without being explicitly prompted to do so? (ii) When provided with hints -- either correct or misleading -- about the validity of the problems, to what extent do LLMs adapt to become reliable Logical Thinker? (iii) How trustworthy are the explanations generated by LLMs when they recognize a math problem as flawed? Through extensive experimentation and detailed analysis, our results demonstrate that existing LLMs largely function as Blind Solver and fall short of the reasoning capabilities required to perform as Logical Thinker.
The advancement of large language models (LLMs) has predominantly focused on high-resource languages, leaving low-resource languages, such as those in the Finno-Ugric family, significantly underrepresented. This paper addresses this gap by focusing on V\~oro, Livonian, and Komi. We cover almost the entire cycle of LLM creation, from data collection to instruction tuning and evaluation. Our contributions include developing multilingual base and instruction-tuned models; creating evaluation benchmarks, including the smugri-MT-bench multi-turn conversational benchmark; and conducting human evaluation. We intend for this work to promote linguistic diversity, ensuring that lesser-resourced languages can benefit from advancements in NLP.
NLP benchmarks rely on standardized datasets for training and evaluating models and are crucial for advancing the field. Traditionally, expert annotations ensure high-quality labels; however, the cost of expert annotation does not scale well with the growing demand for larger datasets required by modern models. While crowd-sourcing provides a more scalable solution, it often comes at the expense of annotation precision and consistency. Recent advancements in large language models (LLMs) offer new opportunities to enhance the annotation process, particularly for detecting label errors in existing datasets. In this work, we consider the recent approach of LLM-as-a-judge, leveraging an ensemble of LLMs to flag potentially mislabeled examples. Through a case study of four datasets from the TRUE benchmark, covering different tasks and domains, we empirically analyze the labeling quality of existing datasets, and compare expert, crowd-sourced, and our LLM-based annotations in terms of agreement, label quality, and efficiency, demonstrating the strengths and limitations of each annotation method. Our findings reveal a substantial number of label errors, which, when corrected, induce a significant upward shift in reported model performance. This suggests that many of the LLMs so-called mistakes are due to label errors rather than genuine model failures. Additionally, we discuss the implications of mislabeled data and propose methods to mitigate them in training to improve model performance.
In this paper, we propose a model-agnostic cost-effective approach to developing bilingual base large language models (LLMs) to support English and any target language. The method includes vocabulary expansion, initialization of new embeddings, model training and evaluation. We performed our experiments with three languages, each using a non-Latin script - Ukrainian, Arabic, and Georgian. Our approach demonstrates improved language performance while reducing computational costs. It mitigates the disproportionate penalization of underrepresented languages, promoting fairness and minimizing adverse phenomena such as code-switching and broken grammar. Additionally, we introduce new metrics to evaluate language quality, revealing that vocabulary size significantly impacts the quality of generated text.
Solving complex chart Q&A tasks requires advanced visual reasoning abilities in multimodal large language models (MLLMs). Recent studies highlight that these abilities consist of two main parts: recognizing key information from visual inputs and conducting reasoning over it. Thus, a promising approach to enhance MLLMs is to construct relevant training data focusing on the two aspects. However, collecting and annotating complex charts and questions is costly and time-consuming, and ensuring the quality of annotated answers remains a challenge. In this paper, we propose Code-as-Intermediary Translation (CIT), a cost-effective, efficient and easily scalable data synthesis method for distilling visual reasoning abilities from LLMs to MLLMs. The code serves as an intermediary that translates visual chart representations into textual representations, enabling LLMs to understand cross-modal information. Specifically, we employ text-based synthesizing techniques to construct chart-plotting code and produce ReachQA, a dataset containing 3k reasoning-intensive charts and 20k Q&A pairs to enhance both recognition and reasoning abilities. Experiments show that when fine-tuned with our data, models not only perform well on chart-related benchmarks, but also demonstrate improved multimodal reasoning abilities on general mathematical benchmarks like MathVista. The code and dataset are publicly available at https://github.com/hewei2001/ReachQA.
A primary challenge in large language model (LLM) development is their onerous pre-training cost. Typically, such pre-training involves optimizing a self-supervised objective (such as next-token prediction) over a large corpus. This paper explores a promising paradigm to improve LLM pre-training efficiency and quality by suitably leveraging a small language model (SLM). In particular, this paradigm relies on an SLM to both (1) provide soft labels as additional training supervision, and (2) select a small subset of valuable ("informative" and "hard") training examples. Put together, this enables an effective transfer of the SLM's predictive distribution to the LLM, while prioritizing specific regions of the training data distribution. Empirically, this leads to reduced LLM training time compared to standard training, while improving the overall quality. Theoretically, we develop a statistical framework to systematically study the utility of SLMs in enabling efficient training of high-quality LLMs. In particular, our framework characterizes how the SLM's seemingly low-quality supervision can enhance the training of a much more capable LLM. Furthermore, it also highlights the need for an adaptive utilization of such supervision, by striking a balance between the bias and variance introduced by the SLM-provided soft labels. We corroborate our theoretical framework by improving the pre-training of an LLM with 2.8B parameters by utilizing a smaller LM with 1.5B parameters on the Pile dataset.
Large Language Models (LLMs) have demonstrated great potential in robotic applications by providing essential general knowledge. Mobile robots rely on map comprehension for tasks like localization and navigation. In this paper, we explore enabling LLMs to comprehend the topology and hierarchy of Area Graph, a text-based hierarchical, topometric semantic map representation utilizing polygons to demark areas such as rooms or buildings. Our experiments demonstrate that with the right map representation, LLMs can effectively comprehend Area Graph's topology and hierarchy. After straightforward fine-tuning, the LLaMA2 models exceeded ChatGPT-3.5 in mastering these aspects. Our dataset, dataset generation code, fine-tuned LoRA adapters can be accessed at https://github.com/xiefujing/LLM-osmAG-Comprehension.
Advancements in distributed training and efficient attention mechanisms have significantly expanded the context window sizes of large language models (LLMs). However, recent work reveals that the effective context lengths of open-source LLMs often fall short, typically not exceeding half of their training lengths. In this work, we attribute this limitation to the left-skewed frequency distribution of relative positions formed in LLMs pretraining and post-training stages, which impedes their ability to effectively gather distant information. To address this challenge, we introduce ShifTed Rotray position embeddING (STRING). STRING shifts well-trained positions to overwrite the original ineffective positions during inference, enhancing performance within their existing training lengths. Experimental results show that without additional training, STRING dramatically improves the performance of the latest large-scale models, such as Llama3.1 70B and Qwen2 72B, by over 10 points on popular long-context benchmarks RULER and InfiniteBench, establishing new state-of-the-art results for open-source LLMs. Compared to commercial models, Llama 3.1 70B with \method even achieves better performance than GPT-4-128K and clearly surpasses Claude 2 and Kimi-chat.
In this paper, we propose a novel framework that leverages large language models (LLMs) for predicting missing values in time-varying graph signals by exploiting spatial and temporal smoothness. We leverage the power of LLM to achieve a message-passing scheme. For each missing node, its neighbors and previous estimates are fed into and processed by LLM to infer the missing observations. Tested on the task of the online prediction of wind-speed graph signals, our model outperforms online graph filtering algorithms in terms of accuracy, demonstrating the potential of LLMs in effectively addressing partially observed signals in graphs.
Recent research has focused on literary machine translation (MT) as a new challenge in MT. However, the evaluation of literary MT remains an open problem. We contribute to this ongoing discussion by introducing LITEVAL-CORPUS, a paragraph-level parallel corpus comprising multiple verified human translations and outputs from 9 MT systems, which totals over 2k paragraphs and includes 13k annotated sentences across four language pairs, costing 4.5k Euro. This corpus enables us to (i) examine the consistency and adequacy of multiple annotation schemes, (ii) compare evaluations by students and professionals, and (iii) assess the effectiveness of LLM-based metrics. We find that Multidimensional Quality Metrics (MQM), as the de facto standard in non-literary human MT evaluation, is inadequate for literary translation: While Best-Worst Scaling (BWS) with students and Scalar Quality Metric (SQM) with professional translators prefer human translations at rates of ~82% and ~94%, respectively, MQM with student annotators prefers human professional translations over the translations of the best-performing LLMs in only ~42% of cases. While automatic metrics generally show a moderate correlation with human MQM and SQM, they struggle to accurately identify human translations, with rates of at most ~20%. Our overall evaluation indicates that human professional translations consistently outperform LLM translations, where even the most recent LLMs tend to produce more literal and less diverse translations compared to human translations. However, newer LLMs such as GPT-4o perform substantially better than older ones.
The availability of high-quality data is one of the most important factors in improving the reasoning capability of LLMs. Existing works have demonstrated the effectiveness of creating more instruction data from seed questions or knowledge bases. Recent research indicates that continually scaling up data synthesis from strong models (e.g., GPT-4) can further elicit reasoning performance. Though promising, the open-sourced community still lacks high-quality data at scale and scalable data synthesis methods with affordable costs. To address this, we introduce ScaleQuest, a scalable and novel data synthesis method that utilizes "small-size" (e.g., 7B) open-source models to generate questions from scratch without the need for seed data with complex augmentation constraints. With the efficient ScaleQuest, we automatically constructed a mathematical reasoning dataset consisting of 1 million problem-solution pairs, which are more effective than existing open-sourced datasets. It can universally increase the performance of mainstream open-source models (i.e., Mistral, Llama3, DeepSeekMath, and Qwen2-Math) by achieving 29.2% to 46.4% gains on MATH. Notably, simply fine-tuning the Qwen2-Math-7B-Base model with our dataset can even surpass Qwen2-Math-7B-Instruct, a strong and well-aligned model on closed-source data, and proprietary models such as GPT-4-Turbo and Claude-3.5 Sonnet.