llm - 2024_11
Navigation
Papers
Recent advances in large language models (LLMs) show the potential of using LLMs as evaluators for assessing the quality of text generations from LLMs. However, applying LLM evaluators naively to compare or judge between different systems can lead to unreliable results due to the intrinsic win rate estimation bias of LLM evaluators. In order to mitigate this problem, we propose two calibration methods, Bayesian Win Rate Sampling (BWRS) and Bayesian Dawid-Skene, both of which leverage Bayesian inference to more accurately infer the true win rate of generative language models. We empirically validate our methods on six datasets covering story generation, summarization, and instruction following tasks. We show that both our methods are effective in improving the accuracy of win rate estimation using LLMs as evaluators, offering a promising direction for reliable automatic text quality evaluation.
We present a comprehensive study of answer quality evaluation in Retrieval-Augmented Generation (RAG) applications using vRAG-Eval, a novel grading system that is designed to assess correctness, completeness, and honesty. We further map the grading of quality aspects aforementioned into a binary score, indicating an accept or reject decision, mirroring the intuitive "thumbs-up" or "thumbs-down" gesture commonly used in chat applications. This approach suits factual business contexts where a clear decision opinion is essential. Our assessment applies vRAG-Eval to two Large Language Models (LLMs), evaluating the quality of answers generated by a vanilla RAG application. We compare these evaluations with human expert judgments and find a substantial alignment between GPT-4's assessments and those of human experts, reaching 83% agreement on accept or reject decisions. This study highlights the potential of LLMs as reliable evaluators in closed-domain, closed-ended settings, particularly when human evaluations require significant resources.
Enterprise searches require users to have complex knowledge of queries, configurations, and metadata, rendering it difficult for them to access information as needed. Most go-to-market (GTM) platforms utilize advanced search, an interface that enables users to filter queries by various fields using categories or keywords, which, historically, however, has proven to be exceedingly cumbersome, as users are faced with seemingly hundreds of options, fields, and buttons. Consequently, querying with natural language has long been ideal, a notion further empowered by Large Language Models (LLMs). In this paper, we implement and evaluate a solution for the Zoominfo product for sellers, which prompts the LLM with natural language, producing search fields through entity extraction that are then converted into a search query. The intermediary search fields offer numerous advantages for each query, including the elimination of syntax errors, simpler ground truths, and an intuitive format for the LLM to interpret. We paired this pipeline with many advanced prompt engineering strategies, featuring an intricate system message, few-shot prompting, chain-of-thought (CoT) reasoning, and execution refinement. Furthermore, we manually created the ground truth for 500+ natural language queries, enabling the supervised fine-tuning of Llama-3-8B-Instruct and the introduction of sophisticated numerical metrics. Comprehensive experiments with closed, open source, and fine-tuned LLM models were conducted through exact, Jaccard, cosine, and semantic similarity on individual search entities to demonstrate the efficacy of our approach. Overall, the most accurate closed model had an average accuracy of 97% per query, with only one field performing under 90%, with comparable results observed from the fine-tuned models.
In digital healthcare, large language models (LLMs) have primarily been utilized to enhance question-answering capabilities and improve patient interactions. However, effective patient care necessitates LLM chains that can actively gather information by posing relevant questions. This paper presents HealthQ, a novel framework designed to evaluate the questioning capabilities of LLM healthcare chains. We implemented several LLM chains, including Retrieval-Augmented Generation (RAG), Chain of Thought (CoT), and reflective chains, and introduced an LLM judge to assess the relevance and informativeness of the generated questions. To validate HealthQ, we employed traditional Natural Language Processing (NLP) metrics such as Recall-Oriented Understudy for Gisting Evaluation (ROUGE) and Named Entity Recognition (NER)-based set comparison, and constructed two custom datasets from public medical note datasets, ChatDoctor and MTS-Dialog. Our contributions are threefold: we provide the first comprehensive study on the questioning capabilities of LLMs in healthcare conversations, develop a novel dataset generation pipeline, and propose a detailed evaluation methodology.
Machine unlearning aims to solve the problem of removing the influence of selected training examples from a learned model. Despite the increasing attention to this problem, it remains an open research question how to evaluate unlearning in large language models (LLMs), and what are the critical properties of the data to be unlearned that affect the quality and efficiency of unlearning. This work formalizes a metric to evaluate unlearning quality in generative models, and uses it to assess the trade-offs between unlearning quality and performance. We demonstrate that unlearning out-of-distribution examples requires more unlearning steps but overall presents a better trade-off overall. For in-distribution examples, however, we observe a rapid decay in performance as unlearning progresses. We further evaluate how example's memorization and difficulty affect unlearning under a classical gradient ascent-based approach.
High-performance analog and mixed-signal (AMS) circuits are mainly full-custom designed, which is time-consuming and labor-intensive. A significant portion of the effort is experience-driven, which makes the automation of AMS circuit design a formidable challenge. Large language models (LLMs) have emerged as powerful tools for Electronic Design Automation (EDA) applications, fostering advancements in the automatic design process for large-scale AMS circuits. However, the absence of high-quality datasets has led to issues such as model hallucination, which undermines the robustness of automatically generated circuit designs. To address this issue, this paper introduces AMSnet-KG, a dataset encompassing various AMS circuit schematics and netlists. We construct a knowledge graph with annotations on detailed functional and performance characteristics. Facilitated by AMSnet-KG, we propose an automated AMS circuit generation framework that utilizes the comprehensive knowledge embedded in LLMs. We first formulate a design strategy (e.g., circuit architecture using a number of circuit components) based on required specifications. Next, matched circuit components are retrieved and assembled into a complete topology, and transistor sizing is obtained through Bayesian optimization. Simulation results of the netlist are fed back to the LLM for further topology refinement, ensuring the circuit design specifications are met. We perform case studies of operational amplifier and comparator design to verify the automatic design flow from specifications to netlists with minimal human effort. The dataset used in this paper will be open-sourced upon publishing of this paper.
Large Language Models (LLMs) are currently pre-trained and fine-tuned on large cloud servers. The next frontier is LLM personalization, where a foundation model can be fine-tuned with user/task-specific data. Given the sensitive nature of such private data, it is desirable to fine-tune these models on edge devices to improve user trust. However, fine-tuning on resource-constrained edge devices presents significant challenges due to substantial memory and computational demands, as well as limited infrastructure support. We observe that inference engines (e.g., ExecuTorch) can be repurposed for fine-tuning by leveraging zeroth-order (ZO) optimization, which uses multiple forward passes to approximate gradients. However, directly applying ZO methods on edge devices is impractical due to the high computational cost of multiple model perturbations required to achieve accuracy improvements. Based on these observations, we propose a memory- and computation-efficient LLM fine-tuning method for edge devices. Our approach has three key innovations: (1) We introduce a parallelized randomized gradient estimation (P-RGE) technique that achieves high parallel efficiency by leveraging outer-loop and inner-loop parallelization. This enables multiple function queries and forward passes to be executed in parallel, reducing training time. (2) We integrate P-RGE with parameter-efficient fine-tuning methods (e.g. LoRA) to further reduce computational and memory overhead. (3) We implement a P-RGE LoRA-FA module that fully supports fine-tuning with ExecuTorch. Our approach requires no modifications to ExecuTorch's runtime code, as it can be implemented with server-side code changes only. Experiments demonstrate that P-RGE achieves substantial runtime speedups and memory savings while improving fine-tuning accuracy, paving the way for practical deployment of LLMs in real-time, on-device applications.
Large Language Models (LLMs) face significant challenges at inference time due to their high computational demands. To address this, we present Performance-Guided Knowledge Distillation (PGKD), a cost-effective and high-throughput solution for production text classification applications. PGKD utilizes teacher-student Knowledge Distillation to distill the knowledge of LLMs into smaller, task-specific models. PGKD establishes an active learning routine between the student model and the LLM; the LLM continuously generates new training data leveraging hard-negative mining, student model validation performance, and early-stopping protocols to inform the data generation. By employing a cyclical, performance-aware approach tailored for highly multi-class, sparsely annotated datasets prevalent in industrial text classification, PGKD effectively addresses training challenges and outperforms traditional BERT-base models and other knowledge distillation methods on several multi-class classification datasets. Additionally, cost and latency benchmarking reveals that models fine-tuned with PGKD are up to 130X faster and 25X less expensive than LLMs for inference on the same classification task. While PGKD is showcased for text classification tasks, its versatile framework can be extended to any LLM distillation task, including language generation, making it a powerful tool for optimizing performance across a wide range of AI applications.
Watermarking has emerged as a prominent technique for LLM-generated content detection by embedding imperceptible patterns. Despite supreme performance, its robustness against adversarial attacks remains underexplored. Previous work typically considers a grey-box attack setting, where the specific type of watermark is already known. Some even necessitates knowledge about hyperparameters of the watermarking method. Such prerequisites are unattainable in real-world scenarios. Targeting at a more realistic black-box threat model with fewer assumptions, we here propose $B^4$, a black-box scrubbing attack on watermarks. Specifically, we formulate the watermark scrubbing attack as a constrained optimization problem by capturing its objectives with two distributions, a Watermark Distribution and a Fidelity Distribution. This optimization problem can be approximately solved using two proxy distributions. Experimental results across 12 different settings demonstrate the superior performance of $B^4$ compared with other baselines.
Large language models (LLMs) have demonstrated notable potential in conducting complex tasks and are increasingly utilized in various financial applications. However, high-quality sequential financial investment decision-making remains challenging. These tasks require multiple interactions with a volatile environment for every decision, demanding sufficient intelligence to maximize returns and manage risks. Although LLMs have been used to develop agent systems that surpass human teams and yield impressive investment returns, opportunities to enhance multi-sourced information synthesis and optimize decision-making outcomes through timely experience refinement remain unexplored. Here, we introduce the FinCon, an LLM-based multi-agent framework with CONceptual verbal reinforcement tailored for diverse FINancial tasks. Inspired by effective real-world investment firm organizational structures, FinCon utilizes a manager-analyst communication hierarchy. This structure allows for synchronized cross-functional agent collaboration towards unified goals through natural language interactions and equips each agent with greater memory capacity than humans. Additionally, a risk-control component in FinCon enhances decision quality by episodically initiating a self-critiquing mechanism to update systematic investment beliefs. The conceptualized beliefs serve as verbal reinforcement for the future agent's behavior and can be selectively propagated to the appropriate node that requires knowledge updates. This feature significantly improves performance while reducing unnecessary peer-to-peer communication costs. Moreover, FinCon demonstrates strong generalization capabilities in various financial tasks, including single stock trading and portfolio management.
Text-to-SQL enables users to interact with databases through natural language, simplifying access to structured data. Although highly capable large language models (LLMs) achieve strong accuracy for complex queries, they incur unnecessary latency and dollar cost for simpler ones. In this paper, we introduce the first LLM routing approach for Text-to-SQL, which dynamically selects the most cost-effective LLM capable of generating accurate SQL for each query. We present two routing strategies (score- and classification-based) that achieve accuracy comparable to the most capable LLM while reducing costs. We design the routers for ease of training and efficient inference. In our experiments, we highlight a practical and explainable accuracy-cost trade-off on the BIRD dataset.
Large language models (LLMs) generate diverse, situated, persuasive texts from a plurality of potential perspectives, influenced heavily by their prompts and training data. As part of LLM adoption, we seek to characterize - and ideally, manage - the socio-cultural values that they express, for reasons of safety, accuracy, inclusion, and cultural fidelity. We present a validated approach to automatically (1) extracting heterogeneous latent value propositions from texts, (2) assessing resonance and conflict of values with texts, and (3) combining these operations to characterize the pluralistic value alignment of human-sourced and LLM-sourced textual data.
The increased use of large language models (LLMs) across a variety of real-world applications calls for automatic tools to check the factual accuracy of their outputs, as LLMs often hallucinate. This is difficult as it requires assessing the factuality of free-form open-domain responses. While there has been a lot of research on this topic, different papers use different evaluation benchmarks and measures, which makes them hard to compare and hampers future progress. To mitigate these issues, we developed OpenFactCheck, a unified framework, with three modules: (i) RESPONSEEVAL, which allows users to easily customize an automatic fact-checking system and to assess the factuality of all claims in an input document using that system, (ii) LLMEVAL, which assesses the overall factuality of an LLM, and (iii) CHECKEREVAL, a module to evaluate automatic fact-checking systems. OpenFactCheck is open-sourced (https://github.com/mbzuai-nlp/openfactcheck) and publicly released as a Python library (https://pypi.org/project/openfactcheck/) and also as a web service (http://app.openfactcheck.com). A video describing the system is available at https://youtu.be/-i9VKL0HleI.
LLMs are widely used in complex AI applications. These applications underscore the need for LLM outputs to adhere to a specific format, for their integration with other components in the systems. Typically the format rules e.g., for data serialization formats such as JSON, YAML, or Code in Programming Language are expressed as context-free grammar (CFG). Due to the hallucinations and unreliability of LLMs, instructing LLMs to adhere to specified syntax becomes an increasingly important challenge. We present SynCode, a novel framework for efficient and general syntactical decoding with LLMs, to address this challenge. SynCode ensures soundness and completeness with respect to the CFG of a formal language, effectively retaining valid tokens while filtering out invalid ones. SynCode uses an offline-constructed, efficient lookup table, the DFA mask store, derived from the DFA of the language's grammar for efficient generation. SynCode seamlessly integrates with any language defined by CFG, as evidenced by experiments focusing on generating JSON, Python, and Go outputs. Our experiments evaluating the effectiveness of SynCode for JSON generation demonstrate that SynCode eliminates all syntax errors and significantly outperforms state-of-the-art baselines. Furthermore, our results underscore how SynCode significantly reduces 96.07% of syntax errors in generated Python and Go code, showcasing its substantial impact on enhancing syntactical precision in LLM generation. Our code is available at https://github.com/uiuc-focal-lab/syncode
The development of Large Language Models (LLMs) in various languages has been advancing, but the combination of non-English languages with domain-specific contexts remains underexplored. This paper presents our findings from training and evaluating a Japanese business domain-specific LLM designed to better understand business-related documents, such as the news on current affairs, technical reports, and patents. Additionally, LLMs in this domain require regular updates to incorporate the most recent knowledge. Therefore, we also report our findings from the first experiments and evaluations involving updates to this LLM using the latest article data, which is an important problem setting that has not been addressed in previous research. From our experiments on a newly created benchmark dataset for question answering in the target domain, we found that (1) our pretrained model improves QA accuracy without losing general knowledge, and (2) a proper mixture of the latest and older texts in the training data for the update is necessary. Our pretrained model and business domain benchmark are publicly available to support further studies.
Large Language Models (LLMs) have fast become an essential tools to many conversational chatbots due to their ability to provide coherent answers for varied queries. Datasets used to train these LLMs are often a mix of generic and synthetic samples, thus lacking the verification needed to provide correct and verifiable answers for T.V. News. We collect and share a large collection of QA pairs extracted from transcripts of news recordings from various news-channels across the United States. Resultant QA pairs are then used to fine-tune an off-the-shelf LLM model. Our model surpasses base models of similar size on several open LLM benchmarks. We further integrate and propose a RAG method to improve contextualization of our answers and also point it to a verifiable news recording.
In recent years, the input context sizes of large language models (LLMs) have increased dramatically. However, existing evaluation methods have not kept pace, failing to comprehensively assess the efficiency of models in handling long contexts. To bridge this gap, we introduce the BABILong benchmark, designed to test language models' ability to reason across facts distributed in extremely long documents. BABILong includes a diverse set of 20 reasoning tasks, including fact chaining, simple induction, deduction, counting, and handling lists/sets. These tasks are challenging on their own, and even more demanding when the required facts are scattered across long natural text. Our evaluations show that popular LLMs effectively utilize only 10-20\% of the context and their performance declines sharply with increased reasoning complexity. Among alternatives to in-context reasoning, Retrieval-Augmented Generation methods achieve a modest 60\% accuracy on single-fact question answering, independent of context length. Among context extension methods, the highest performance is demonstrated by recurrent memory transformers after fine-tuning, enabling the processing of lengths up to 50 million tokens. The BABILong benchmark is extendable to any length to support the evaluation of new upcoming models with increased capabilities, and we provide splits up to 10 million token lengths.
Embeddings have become a cornerstone in the functionality of large language models (LLMs) due to their ability to transform text data into rich, dense numerical representations that capture semantic and syntactic properties. These embedding vector databases serve as the long-term memory of LLMs, enabling efficient handling of a wide range of natural language processing tasks. However, the surge in popularity of embedding vector databases in LLMs has been accompanied by significant concerns about privacy leakage. Embedding vector databases are particularly vulnerable to embedding inversion attacks, where adversaries can exploit the embeddings to reverse-engineer and extract sensitive information from the original text data. Existing defense mechanisms have shown limitations, often struggling to balance security with the performance of downstream tasks. To address these challenges, we introduce Eguard, a novel defense mechanism designed to mitigate embedding inversion attacks. Eguard employs a transformer-based projection network and text mutual information optimization to safeguard embeddings while preserving the utility of LLMs. Our approach significantly reduces privacy risks, protecting over 95% of tokens from inversion while maintaining high performance across downstream tasks consistent with original embeddings.
Retrieval-Augmented Generation (RAG) has proven to be an effective method for mitigating hallucination issues inherent in large language models (LLMs). Previous approaches typically train retrievers based on semantic similarity, lacking optimization for RAG. More recent works have proposed aligning retrievers with the preference signals of LLMs. However, these preference signals are often difficult for dense retrievers, which typically have weaker language capabilities, to understand and learn effectively. Drawing inspiration from pedagogical theories like Guided Discovery Learning, we propose a novel framework, FiGRet (Fine-grained Guidance for Retrievers), which leverages the language capabilities of LLMs to construct examples from a more granular, information-centric perspective to guide the learning of retrievers. Specifically, our method utilizes LLMs to construct easy-to-understand examples from samples where the retriever performs poorly, focusing on three learning objectives highly relevant to the RAG scenario: relevance, comprehensiveness, and purity. These examples serve as scaffolding to ultimately align the retriever with the LLM's preferences. Furthermore, we employ a dual curriculum learning strategy and leverage the reciprocal feedback between LLM and retriever to further enhance the performance of the RAG system. A series of experiments demonstrate that our proposed framework enhances the performance of RAG systems equipped with different retrievers and is applicable to various LLMs.
Hampering the interpretation of benchmark scores, evaluation data contamination has become a growing concern in the evaluation of LLMs, and an active area of research studies its effects. While evaluation data contamination is easily understood intuitively, it is surprisingly difficult to define precisely which samples should be considered contaminated and, consequently, how it impacts benchmark scores. We propose that these questions should be addressed together and that contamination metrics can be assessed based on whether models benefit from the examples they mark contaminated. We propose a novel analysis method called ConTAM, and show with a large scale survey of existing and novel n-gram based contamination metrics across 13 benchmarks and 7 models from 2 different families that ConTAM can be used to better understand evaluation data contamination and its effects. We find that contamination may have a much larger effect than reported in recent LLM releases and benefits models differently at different scales. We also find that considering only the longest contaminated substring provides a better signal than considering a union of all contaminated substrings, and that doing model and benchmark specific threshold analysis greatly increases the specificity of the results. Lastly, we investigate the impact of hyperparameter choices, finding that, among other things, both using larger values of n and disregarding matches that are infrequent in the pre-training data lead to many false negatives. With ConTAM, we provide a method to empirically ground evaluation data contamination metrics in downstream effects. With our exploration, we shed light on how evaluation data contamination can impact LLMs and provide insight into the considerations important when doing contamination analysis. We end our paper by discussing these in more detail and providing concrete suggestions for future work.
While LLM-Based agents, which use external tools to solve complex problems, have made significant progress, benchmarking their ability is challenging, thereby hindering a clear understanding of their limitations. In this paper, we propose an interactive evaluation framework, named CIBench, to comprehensively assess LLMs' ability to utilize code interpreters for data science tasks. Our evaluation framework includes an evaluation dataset and two evaluation modes. The evaluation dataset is constructed using an LLM-human cooperative approach and simulates an authentic workflow by leveraging consecutive and interactive IPython sessions. The two evaluation modes assess LLMs' ability with and without human assistance. We conduct extensive experiments to analyze the ability of 24 LLMs on CIBench and provide valuable insights for future LLMs in code interpreter utilization.
The rapid progression of multimodal large language models (MLLMs) has demonstrated superior performance on various multimodal benchmarks. However, the issue of data contamination during training creates challenges in performance evaluation and comparison. While numerous methods exist for detecting dataset contamination in large language models (LLMs), they are less effective for MLLMs due to their various modalities and multiple training phases. In this study, we introduce a multimodal data contamination detection framework, MM-Detect, designed for MLLMs. Our experimental results indicate that MM-Detect is sensitive to varying degrees of contamination and can highlight significant performance improvements due to leakage of the training set of multimodal benchmarks. Furthermore, We also explore the possibility of contamination originating from the pre-training phase of LLMs used by MLLMs and the fine-tuning phase of MLLMs, offering new insights into the stages at which contamination may be introduced.
Large Language Models (LLMs) specializing in code generation (which are also often referred to as code LLMs), e.g., StarCoder and Code Llama, play increasingly critical roles in various software development scenarios. It is also crucial for code LLMs to possess both code generation and natural language abilities for many specific applications, such as code snippet retrieval using natural language or code explanations. The intricate interaction between acquiring language and coding skills complicates the development of strong code LLMs. Furthermore, there is a lack of thorough prior studies on the LLM pretraining strategy that mixes code and natural language. In this work, we propose a pretraining strategy to enhance the integration of natural language and coding capabilities within a single LLM. Specifically, it includes two phases of training with appropriately adjusted code/language ratios. The resulting model, Crystal, demonstrates remarkable capabilities in both domains. Specifically, it has natural language and coding performance comparable to that of Llama 2 and Code Llama, respectively. Crystal exhibits better data efficiency, using 1.4 trillion tokens compared to the more than 2 trillion tokens used by Llama 2 and Code Llama. We verify our pretraining strategy by analyzing the training process and observe consistent improvements in most benchmarks. We also adopted a typical application adaptation phase with a code-centric data mixture, only to find that it did not lead to enhanced performance or training efficiency, underlining the importance of a carefully designed data recipe. To foster research within the community, we commit to open-sourcing every detail of the pretraining, including our training datasets, code, loggings and 136 checkpoints throughout the training.
Large Language Models (LLMs) are increasingly augmented with external tools and commercial services into LLM-integrated systems. While these interfaces can significantly enhance the capabilities of the models, they also introduce a new attack surface. Manipulated integrations, for example, can exploit the model and compromise sensitive data accessed through other interfaces. While previous work primarily focused on attacks targeting a model's alignment or the leakage of training data, the security of data that is only available during inference has escaped scrutiny so far. In this work, we demonstrate the vulnerabilities associated with external components and introduce a systematic approach to evaluate confidentiality risks in LLM-integrated systems. We identify two specific attack scenarios unique to these systems and formalize these into a tool-robustness framework designed to measure a model's ability to protect sensitive information. Our findings show that all examined models are highly vulnerable to confidentiality attacks, with the risk increasing significantly when models are used together with external tools.
Natural Language Generation has been rapidly developing with the advent of large language models (LLMs). While their usage has sparked significant attention from the general public, it is important for readers to be aware when a piece of text is LLM-generated. This has brought about the need for building models that enable automated LLM-generated text detection, with the aim of mitigating potential negative outcomes of such content. Existing LLM-generated detectors show competitive performances in telling apart LLM-generated and human-written text, but this performance is likely to deteriorate when paraphrased texts are considered. In this study, we devise a new data collection strategy to collect Human & LLM Paraphrase Collection (HLPC), a first-of-its-kind dataset that incorporates human-written texts and paraphrases, as well as LLM-generated texts and paraphrases. With the aim of understanding the effects of human-written paraphrases on the performance of state-of-the-art LLM-generated text detectors OpenAI RoBERTa and watermark detectors, we perform classification experiments that incorporate human-written paraphrases, watermarked and non-watermarked LLM-generated documents from GPT and OPT, and LLM-generated paraphrases from DIPPER and BART. The results show that the inclusion of human-written paraphrases has a significant impact of LLM-generated detector performance, promoting TPR@1%FPR with a possible trade-off of AUROC and accuracy.
The wide deployment of Large Language Models (LLMs) has given rise to strong demands for optimizing their inference performance. Today's techniques serving this purpose primarily focus on reducing latency and improving throughput through algorithmic and hardware enhancements, while largely overlooking their privacy side effects, particularly in a multi-user environment. In our research, for the first time, we discovered a set of new timing side channels in LLM systems, arising from shared caches and GPU memory allocations, which can be exploited to infer both confidential system prompts and those issued by other users. These vulnerabilities echo security challenges observed in traditional computing systems, highlighting an urgent need to address potential information leakage in LLM serving infrastructures. In this paper, we report novel attack strategies designed to exploit such timing side channels inherent in LLM deployments, specifically targeting the Key-Value (KV) cache and semantic cache widely used to enhance LLM inference performance. Our approach leverages timing measurements and classification models to detect cache hits, allowing an adversary to infer private prompts with high accuracy. We also propose a token-by-token search algorithm to efficiently recover shared prompt prefixes in the caches, showing the feasibility of stealing system prompts and those produced by peer users. Our experimental studies on black-box testing of popular online LLM services demonstrate that such privacy risks are completely realistic, with significant consequences. Our findings underscore the need for robust mitigation to protect LLM systems against such emerging threats.
We use images of cars of a wide range of varieties to compose an image of an animal such as a bird or a lion for the theme of environmental protection to maximize the information about cars in a single composed image and to raise the awareness about environmental challenges. We present a novel way of image interaction with an artistically-composed photomosaic image, in which a simple operation of "click and display" is used to demonstrate the interactive switch between a tile image in a photomosaic image and the corresponding original car image, which will be automatically saved on the Desktop. We build a multimodal custom GPT named TalkMosaic by incorporating car images information and the related knowledge to ChatGPT. By uploading the original car image to TalkMosaic, we can ask questions about the given car image and get the corresponding answers efficiently and effectively such as where to buy the tire in the car image that satisfies high environmental standards. We give an in-depth analysis on how to speed up the inference of multimodal LLM using sparse attention and quantization techniques with presented probabilistic FlashAttention (PrFlashAttention) and Staircase Adaptive Quantization (SAQ) methods. The implemented prototype demonstrates the feasibility and effectiveness of the presented approach.
Although Large Language Models (LLMs) have demonstrated significant capabilities in executing complex tasks in a zero-shot manner, they are susceptible to jailbreak attacks and can be manipulated to produce harmful outputs. Recently, a growing body of research has categorized jailbreak attacks into token-level and prompt-level attacks. However, previous work primarily overlooks the diverse key factors of jailbreak attacks, with most studies concentrating on LLM vulnerabilities and lacking exploration of defense-enhanced LLMs. To address these issues, we introduced $\textbf{JailTrickBench}$ to evaluate the impact of various attack settings on LLM performance and provide a baseline for jailbreak attacks, encouraging the adoption of a standardized evaluation framework. Specifically, we evaluate the eight key factors of implementing jailbreak attacks on LLMs from both target-level and attack-level perspectives. We further conduct seven representative jailbreak attacks on six defense methods across two widely used datasets, encompassing approximately 354 experiments with about 55,000 GPU hours on A800-80G. Our experimental results highlight the need for standardized benchmarking to evaluate these attacks on defense-enhanced LLMs. Our code is available at https://github.com/usail-hkust/JailTrickBench.
One useful application of NLP models is to support people in reading complex text from unfamiliar domains (e.g., scientific articles). Simplifying the entire text makes it understandable but sometimes removes important details. On the contrary, helping adult readers understand difficult concepts in context can enhance their vocabulary and knowledge. In a preliminary human study, we first identify that lack of context and unfamiliarity with difficult concepts is a major reason for adult readers' difficulty with domain-specific text. We then introduce "targeted concept simplification," a simplification task for rewriting text to help readers comprehend text containing unfamiliar concepts. We also introduce WikiDomains, a new dataset of 22k definitions from 13 academic domains paired with a difficult concept within each definition. We benchmark the performance of open-source and commercial LLMs and a simple dictionary baseline on this task across human judgments of ease of understanding and meaning preservation. Interestingly, our human judges preferred explanations about the difficult concept more than simplification of the concept phrase. Further, no single model achieved superior performance across all quality dimensions, and automated metrics also show low correlations with human evaluations of concept simplification ($\sim0.2$), opening up rich avenues for research on personalized human reading comprehension support.
State-of-the-art results in large language models (LLMs) often rely on scale, which becomes computationally expensive. This has sparked a research agenda to reduce these models' parameter counts and computational costs without significantly impacting their performance. Our study focuses on transformer-based LLMs, specifically targeting the computationally intensive feedforward networks (FFNs), which are less studied than attention blocks. We consider three structured linear parameterizations of the FFN using efficient low-rank and block-diagonal matrices. In contrast to many previous works that examined these approximations, our study i) explores these structures from a training-from-scratch perspective, ii) scales up to 1.3B parameters, and iii) is conducted within recent Transformer-based LLMs rather than convolutional architectures. We demonstrate that these structures can lead to actual computational gains in various scenarios, including online decoding when using a pre-merge technique. Additionally, we propose a novel training regime, called \textit{self-guided training}, aimed at improving the poor training dynamics that these approximations exhibit when used from initialization. Interestingly, the scaling performance of structured matrices is explored, revealing steeper curves in scaling training FLOPs, along with a favorable scaling trend in the overtraining regime. Specifically, we show that wide and structured networks can utilize training FLOPs more efficiently, with fewer parameters and lower loss than dense models at their optimal trade-off. Our code is available at \url{https://github.com/CLAIRE-Labo/StructuredFFN/tree/main}.
This paper introduces a novel model compression approach through dynamic layer-specific pruning in Large Language Models (LLMs), enhancing the traditional methodology established by SliceGPT. By transitioning from constant to dynamic slicing, our method leverages the newly proposed Layer Redundancy (LR) score, which assesses how much change each layer changes its input by measuring the cosine similarity of the input to the output of the layer. We use this score to prune parts of individual layers based on redundancy in such a way that the average pruned percentage for all layers is a fixed value. We conducted extensive experiments using models like Llama3-8B and Mistral-7B on multiple datasets, evaluating different slicing bases and percentages to determine optimal configurations that balance efficiency and performance. Our findings show that our dynamic slicing approach not only maintains but, in many cases, enhances model performance compared to the baseline established by constant slicing methods. For instance, in several settings, we see performance improvements of up to 5% over the SliceGPT baseline. Additionally, a perplexity decrease by as much as 7% was observed across multiple benchmarks, validating the effectiveness of our method. The code, model weights, and datasets are open-sourced at https://github.com/RazvanDu/DynamicSlicing.
Feature engineering is crucial for optimizing machine learning model performance, particularly in tabular data classification tasks. Leveraging advancements in natural language processing, this study presents a systematic approach to enrich tabular datasets with features derived from large language model embeddings. Through a comprehensive ablation study on diverse datasets, we assess the impact of RoBERTa and GPT-2 embeddings on ensemble classifiers, including Random Forest, XGBoost, and CatBoost. Results indicate that integrating embeddings with traditional numerical and categorical features often enhances predictive performance, especially on datasets with class imbalance or limited features and samples, such as UCI Adult, Heart Disease, Titanic, and Pima Indian Diabetes, with improvements particularly notable in XGBoost and CatBoost classifiers. Additionally, feature importance analysis reveals that LLM-derived features frequently rank among the most impactful for the predictions. This study provides a structured approach to embedding-based feature enrichment and illustrates its benefits in ensemble learning for tabular data.
Individuals with fine motor impairments, such as those caused by conditions like Parkinson's disease, cerebral palsy, or dyspraxia, face significant challenges in interacting with traditional computer interfaces. Historically, scripted automation has offered some assistance, but these solutions are often too rigid and task-specific, failing to adapt to the diverse needs of users. The advent of Large Language Models (LLMs) promised a more flexible approach, capable of interpreting natural language commands to navigate complex user interfaces. However, current LLMs often misinterpret user intent and have no fallback measures when user instructions do not directly align with the specific wording used in the Document Object Model (DOM). This research presents Dexterity Assist (DexAssist), a dual-LLM system designed to improve the reliability of automated user interface control. Both LLMs work iteratively to ensure successful task execution: the Navigator LLM generates actions based on user input, while the Support LLM assesses the success of these actions and provides continuous feedback based on the DOM content. Our framework displays an increase of ~36 percentage points in overall accuracy within the first iteration of the Support LLM, highlighting its effectiveness in resolving errors in real-time. The main contributions of this paper are the design of a novel dual LLM-based accessibility system, its implementation, and its initial evaluation using 3 e-commerce websites. We conclude by underscoring the potential to build on this framework by optimizing computation time and fine-tuning.
This paper introduces distribution-based prediction, a novel approach to using Large Language Models (LLMs) as predictive tools by interpreting output token probabilities as distributions representing the models' learned representation of the world. This distribution-based nature offers an alternative perspective for analyzing algorithmic fidelity, complementing the approach used in silicon sampling. We demonstrate the use of distribution-based prediction in the context of recent United States presidential election, showing that this method can be used to determine task specific bias, prompt noise, and algorithmic fidelity. This approach has significant implications for assessing the reliability and increasing transparency of LLM-based predictions across various domains.
Automated knowledge extraction from scientific literature can potentially accelerate materials discovery. We have investigated an approach for extracting synthesis protocols for reticular materials from scientific literature using large language models (LLMs). To that end, we introduce a Knowledge Extraction Pipeline (KEP) that automatizes LLM-assisted paragraph classification and information extraction. By applying prompt engineering with in-context learning (ICL) to a set of open-source LLMs, we demonstrate that LLMs can retrieve chemical information from PDF documents, without the need for fine-tuning or training and at a reduced risk of hallucination. By comparing the performance of five open-source families of LLMs in both paragraph classification and information extraction tasks, we observe excellent model performance even if only few example paragraphs are included in the ICL prompts. The results show the potential of the KEP approach for reducing human annotations and data curation efforts in automated scientific knowledge extraction.
Large Language Models (LLMs) have demonstrated remarkable skills across various design domains, including UI generation. However, current LLMs for UI generation tend to offer generic solutions that lack a deep understanding of task context and user preferences in specific scenarios. We present \textit{CrowdGenUI}, a framework that enhances LLM-driven UI generation with a crowdsourced user preference library. This approach addresses the limitations of existing methods by guiding LLM reasoning with user preferences, enabling the generation of UI widgets that align more closely with user needs and task-specific requirements. Using image editing as a test domain, we built this library from 50 users, capturing 720 user preferences, which include the predictability, efficiency, and explorability of multiple UI widgets. In a user study with 72 additional participants, our framework outperformed standard LLM-generated widgets in meeting user preferences and task requirements. We discuss these findings to inform future opportunities for designing user-centered and customizable UIs by comprehensively analyzing the extendability of the proposed framework and crowdsourced library.
Solving Partial Differential Equations (PDEs) is ubiquitous in science and engineering. Computational complexity and difficulty in writing numerical solvers has motivated the development of machine learning techniques to generate solutions quickly. Many existing methods are purely data driven, relying solely on numerical solution fields, rather than known system information such as boundary conditions and governing equations. However, the recent rise in popularity of Large Language Models (LLMs) has enabled easy integration of text in multimodal machine learning models. In this work, we use pretrained LLMs to integrate various amounts known system information into PDE learning. Our multimodal approach significantly outperforms our baseline model, FactFormer, in both next-step prediction and autoregressive rollout performance on the 2D Heat, Burgers, Navier-Stokes, and Shallow Water equations. Further analysis shows that pretrained LLMs provide highly structured latent space that is consistent with the amount of system information provided through text.
Many human interactions, such as political debates, are carried out in group settings, where there are arbitrarily many participants, each with different views and agendas. To explore such complex social settings, we present SAUCE: a customizable Python platform, allowing researchers to plug-and-play various LLMs participating in discussions on any topic chosen by the user. Our platform takes care of instantiating the models, scheduling their responses, managing the discussion history, and producing a comprehensive output log, all customizable through configuration files, requiring little to no coding skills. A novel feature of SAUCE is our asynchronous communication feature, where models decide when to speak in addition to what to say, thus modeling an important facet of human communication. We show SAUCE's attractiveness in two initial experiments, and invite the community to use it in simulating various group simulations.
This work analyzes the use of large language models (LLMs) for detecting domain generation algorithms (DGAs). We perform a detailed evaluation of two important techniques: In-Context Learning (ICL) and Supervised Fine-Tuning (SFT), showing how they can improve detection. SFT increases performance by using domain-specific data, whereas ICL helps the detection model to quickly adapt to new threats without requiring much retraining. We use Meta's Llama3 8B model, on a custom dataset with 68 malware families and normal domains, covering several hard-to-detect schemes, including recent word-based DGAs. Results proved that LLM-based methods can achieve competitive results in DGA detection. In particular, the SFT-based LLM DGA detector outperforms state-of-the-art models using attention layers, achieving 94% accuracy with a 4% false positive rate (FPR) and excelling at detecting word-based DGA domains.
We study the emergence of agency from scratch by using Large Language Model (LLM)-based agents. In previous studies of LLM-based agents, each agent's characteristics, including personality and memory, have traditionally been predefined. We focused on how individuality, such as behavior, personality, and memory, can be differentiated from an undifferentiated state. The present LLM agents engage in cooperative communication within a group simulation, exchanging context-based messages in natural language. By analyzing this multi-agent simulation, we report valuable new insights into how social norms, cooperation, and personality traits can emerge spontaneously. This paper demonstrates that autonomously interacting LLM-powered agents generate hallucinations and hashtags to sustain communication, which, in turn, increases the diversity of words within their interactions. Each agent's emotions shift through communication, and as they form communities, the personalities of the agents emerge and evolve accordingly. This computational modeling approach and its findings will provide a new method for analyzing collective artificial intelligence.
Large Language Models (LLMs) are trained on vast amounts of data, most of which is automatically scraped from the internet. This data includes encyclopedic documents that harbor a vast amount of general knowledge (e.g., Wikipedia) but also potentially overlap with benchmark datasets used for evaluating LLMs. Consequently, evaluating models on test splits that might have leaked into the training set is prone to misleading conclusions. To foster sound evaluation of language models, we introduce a new test dataset named RepLiQA, suited for question-answering and topic retrieval tasks. RepLiQA is a collection of five splits of test sets, four of which have not been released to the internet or exposed to LLM APIs prior to this publication. Each sample in RepLiQA comprises (1) a reference document crafted by a human annotator and depicting an imaginary scenario (e.g., a news article) absent from the internet; (2) a question about the document's topic; (3) a ground-truth answer derived directly from the information in the document; and (4) the paragraph extracted from the reference document containing the answer. As such, accurate answers can only be generated if a model can find relevant content within the provided document. We run a large-scale benchmark comprising several state-of-the-art LLMs to uncover differences in performance across models of various types and sizes in a context-conditional language modeling setting. Released splits of RepLiQA can be found here: https://huggingface.co/datasets/ServiceNow/repliqa.
We introduce MPLSandbox, an out-of-the-box multi-programming language sandbox designed to provide unified and comprehensive feedback from compiler and analysis tools for Large Language Models (LLMs). It can automatically identify the programming language of the code, compiling and executing it within an isolated sub-sandbox to ensure safety and stability. In addition, MPLSandbox also integrates both traditional and LLM-based code analysis tools, providing a comprehensive analysis of generated code. MPLSandbox can be effortlessly integrated into the training and deployment of LLMs to improve the quality and correctness of their generated code. It also helps researchers streamline their workflows for various LLM-based code-related tasks, reducing the development cost. To validate the effectiveness of MPLSandbox, we integrate it into training and deployment approaches, and also employ it to optimize workflows for a wide range of real-world code-related tasks. Our goal is to enhance researcher productivity on LLM-based code-related tasks by simplifying and automating workflows through delegation to MPLSandbox.
Static Application Security Testing(SAST) tools are crucial for early bug detection and code quality but often generate false positives that slow development. Automating false positive mitigation is thus essential for advancing SAST tools. Past efforts use static/dynamic analysis or machine learning. The advent of Large Language Models, adept at understanding natural language and code, offers promising ways to improve the accuracy and usability of SAST tools. However, existing LLM-based methods need improvement in two key areas: first, extracted code snippets related to warnings are often cluttered with irrelevant control and data flows, reducing precision; second, critical code contexts are often missing, leading to incomplete representations that can mislead LLMs and cause inaccurate assessments. To ensure the use of precise and complete code context, thereby avoiding misguidance and enabling LLMs to reach accurate conclusions, we propose LLM4FPM. One of its core components is eCPG-Slicer, which builds an extended code property graph and extracts line-level, precise code context. Moreover, LLM4FPM incorporates FARF algorithm, which builds a file reference graph and then efficiently detects all files related to a warning in linear time, enabling eCPG-Slicer to gather complete code context across these files. We evaluate LLM4FPM on Juliet dataset, where it comprehensively outperforms the baseline, achieving an F1 score above 99% across various CWEs. LLM4FPM leverages a free, open-source model, avoiding costly alternatives and reducing inspection costs by up to $2758 per run on Juliet, with an average inspection time of 4.7 seconds per warning. Our work emphasizes the critical impact of precise and complete code context and highlights the potential of combining program analysis with LLMs, improving the quality and efficiency of software development.
We introduce a novel reinforcement learning framework of LLM agents named AGILE (AGent that Interacts and Learns from Environments) designed to perform complex conversational tasks with users, leveraging LLMs, memory, tools, and interactions with experts. The agent possesses capabilities beyond conversation, including reflection, tool usage, and expert consultation. We formulate the construction of such an LLM agent as a reinforcement learning (RL) problem, in which the LLM serves as the policy model. We fine-tune the LLM using labeled data of actions and the PPO algorithm. We focus on question answering and release a dataset for agents called ProductQA, comprising challenging questions in online shopping. Our extensive experiments on ProductQA, MedMCQA and HotPotQA show that AGILE agents based on 7B and 13B LLMs trained with PPO can outperform GPT-4 agents. Our ablation study highlights the indispensability of memory, tools, consultation, reflection, and reinforcement learning in achieving the agent's strong performance. Datasets and code are available at https://github.com/bytarnish/AGILE.
Large language models (LLMs) inherit biases from their training data and alignment processes, influencing their responses in subtle ways. While many studies have examined these biases, little work has explored their robustness during interactions. In this paper, we introduce a novel approach where two instances of an LLM engage in self-debate, arguing opposing viewpoints to persuade a neutral version of the model. Through this, we evaluate how firmly biases hold and whether models are susceptible to reinforcing misinformation or shifting to harmful viewpoints. Our experiments span multiple LLMs of varying sizes, origins, and languages, providing deeper insights into bias persistence and flexibility across linguistic and cultural contexts.
Recently, there has been a growing trend of utilizing Large Language Model (LLM) to evaluate the quality of other LLMs. Many studies have employed proprietary close-sourced models, especially GPT-4, as the evaluator. Alternatively, other works have fine-tuned judge models based on open-source LLMs as the evaluator. While the fine-tuned judge models are claimed to achieve comparable evaluation capability with GPT-4, in this work, we conduct an empirical study of judge models. Our findings indicate that although the fine-tuned judge models achieve high performance on in-domain test sets, even surpassing GPT-4, they underperform GPT-4 across several dimensions, including generalizability, fairness, aspect-specific evaluation, and scalability. We also reveal that the fine-tuned judge model inherently operates as a task-specific classifier, consequently imposing the limitations. Finally, we introduce a integrated method, leveraging GPT-4 to compensate for the limitations and improve the fine-tuned judges. Experiment results show our method achieves accuracy on par with GPT-4 with only 50% of the API expense.
Scaling large language models (LLMs) demands extensive data and computing resources, which are traditionally constrained to data centers by the high-bandwidth requirements of distributed training. Low-bandwidth methods like federated learning (FL) could enable collaborative training of larger models across weakly-connected GPUs if they can effectively be used for pre-training. To achieve this, we introduce Photon, the first complete system for federated end-to-end LLM training, leveraging cross-silo FL for global-scale training with minimal communication overheads. Using Photon, we train the first federated family of decoder-only LLMs from scratch. We show that: (1) Photon can train model sizes up to 7B in a federated fashion while reaching an even better perplexity than centralized pre-training; (2) Photon model training time decreases with available compute, achieving a similar compute-time trade-off to centralized; and (3) Photon outperforms the wall-time of baseline distributed training methods by 35% via communicating 64x-512xless. Our proposal is robust to data heterogeneity and converges twice as fast as previous methods like DiLoCo. This surprising data efficiency stems from a unique approach combining small client batch sizes with extremely high learning rates, enabled by federated averaging's robustness to hyperparameters. Photon thus represents the first economical system for global internet-wide LLM pre-training.
Color-word associations play a fundamental role in human cognition and design applications. Large Language Models (LLMs) have become widely available and demonstrated intelligent behaviors in various benchmarks with natural conversation skills. However, their ability to replicate human color-word associations remains understudied. We compared multiple generations of LLMs (from GPT-3 to GPT-4o) against human color-word associations using data collected from over 10,000 Japanese participants, involving 17 colors and words from eight categories in Japanese. Our findings reveal a clear progression in LLM performance across generations, with GPT-4o achieving the highest accuracy in predicting the best voted word for each color and category. However, the highest median performance was approximately 50% even for GPT-4o with visual inputs (chance level is 10%), and the performance levels varied significantly across word categories and colors, indicating a failure to fully replicate human color-word associations. On the other hand, color discrimination ability estimated from our color-word association data showed that LLMs demonstrated high correlation with human color discrimination patterns, similarly to previous studies. Our study highlights both the advancements in LLM capabilities and their persistent limitations, suggesting differences in semantic memory structures between humans and LLMs in representing color-word associations.
With the development of large language models (LLMs), the ability to handle longer contexts has become a key capability for Web applications such as cross-document understanding and LLM-powered search systems. However, this progress faces two major challenges: performance degradation due to sequence lengths out-of-distribution, and excessively long inference times caused by the quadratic computational complexity of attention. These issues hinder the application of LLMs in long-context scenarios. In this paper, we propose Dynamic Token-Level KV Cache Selection (TokenSelect), a model-agnostic, training-free method for efficient and accurate long-context inference. TokenSelect builds upon the observation of non-contiguous attention sparsity, using Query-Key dot products to measure per-head KV Cache criticality at token-level. By per-head soft voting mechanism, TokenSelect selectively involves a small number of critical KV cache tokens in the attention calculation without sacrificing accuracy. To further accelerate TokenSelect, we designed the Selection Cache based on observations of consecutive Query similarity and implemented efficient dot product kernel, significantly reducing the overhead of token selection. A comprehensive evaluation of TokenSelect demonstrates up to 23.84x speedup in attention computation and up to 2.28x acceleration in end-to-end latency, while providing superior performance compared to state-of-the-art long-context inference methods.
Recent work in database query optimization has used complex machine learning strategies, such as customized reinforcement learning schemes. Surprisingly, we show that LLM embeddings of query text contain useful semantic information for query optimization. Specifically, we show that a simple binary classifier deciding between alternative query plans, trained only on a small number of labeled embedded query vectors, can outperform existing heuristic systems. Although we only present some preliminary results, an LLM-powered query optimizer could provide significant benefits, both in terms of performance and simplicity.
Adapting Large Language Models (LLMs) for recommendation requires careful consideration of the decoding process, given the inherent differences between generating items and natural language. Existing approaches often directly apply LLMs' original decoding methods. However, we find these methods encounter significant challenges: 1) amplification bias -- where standard length normalization inflates scores for items containing tokens with generation probabilities close to 1 (termed ghost tokens), and 2) homogeneity issue -- generating multiple similar or repetitive items for a user. To tackle these challenges, we introduce a new decoding approach named Debiasing-Diversifying Decoding (D3). D3 disables length normalization for ghost tokens to alleviate amplification bias, and it incorporates a text-free assistant model to encourage tokens less frequently generated by LLMs for counteracting recommendation homogeneity. Extensive experiments on real-world datasets demonstrate the method's effectiveness in enhancing accuracy and diversity. The code is available at https://github.com/SAI990323/DecodingMatters.
Code translation aims to convert a program from one programming language (PL) to another. This long-standing software engineering task is crucial for modernizing legacy systems, ensuring cross-platform compatibility, enhancing performance, and more. However, automating this process remains challenging due to many syntactic and semantic differences between PLs. Recent studies show that even advanced techniques such as large language models (LLMs), especially open-source LLMs, still struggle with the task. Currently, code LLMs are trained with source code from multiple programming languages, thus presenting multilingual capabilities. In this paper, we investigate whether such multilingual capabilities can be harnessed to enhance code translation. To achieve this goal, we introduce InterTrans, an LLM-based automated code translation approach that, in contrast to existing approaches, leverages intermediate translations across PLs to bridge the syntactic and semantic gaps between source and target PLs. InterTrans contains two stages. It first utilizes a novel Tree of Code Translation (ToCT) algorithm to plan transitive intermediate translation sequences between a given source and target PL, then validates them in a specific order. We evaluate InterTrans with three open LLMs on three benchmarks (i.e., CodeNet, HumanEval-X, and TransCoder) involving six PLs. Results show an absolute improvement between 18.3% to 43.3% in Computation Accuracy (CA) for InterTrans over Direct Translation with 10 attempts. The best-performing variant of InterTrans (with Magicoder LLM) achieved an average CA of 87.3%-95.4% on three benchmarks.
Large Language Models (LLMs) have demonstrated unparalleled effectiveness in various NLP tasks, and integrating LLMs with automatic speech recognition (ASR) is becoming a mainstream paradigm. Building upon this momentum, our research delves into an in-depth examination of this paradigm on a large open-source Chinese dataset. Specifically, our research aims to evaluate the impact of various configurations of speech encoders, LLMs, and projector modules in the context of the speech foundation encoder-LLM ASR paradigm. Furthermore, we introduce a three-stage training approach, expressly developed to enhance the model's ability to align auditory and textual information. The implementation of this approach, alongside the strategic integration of ASR components, enabled us to achieve the SOTA performance on the AISHELL-1, Test_Net, and Test_Meeting test sets. Our analysis presents an empirical foundation for future research in LLM-based ASR systems and offers insights into optimizing performance using Chinese datasets. We will publicly release all scripts used for data preparation, training, inference, and scoring, as well as pre-trained models and training logs to promote reproducible research.
We investigate the patterns of incorrect answers produced by large language models (LLMs) during evaluation. These errors exhibit highly non-intuitive behaviors unique to each model. By analyzing these patterns, we measure the similarities between LLMs and construct a taxonomy that categorizes them based on their error correlations. Our findings reveal that the incorrect responses are not randomly distributed but systematically correlated across models, providing new insights into the underlying structures and relationships among LLMs.
As large language models (LLMs) become increasingly powerful, traditional evaluation metrics tend to saturate, making it challenging to distinguish between models based on their performance. We propose a general method to transform existing LLM evaluations into a series of progressively more difficult tasks. These enhanced evaluations emphasize reasoning capabilities and can reveal relative performance differences that are not apparent in the original assessments. To demonstrate the effectiveness of our approach, we create a new multiple-choice test corpus, extend it into a family of evaluations, and assess a collection of LLMs. Our results offer insights into the comparative reasoning abilities of these models, particularly highlighting distinctions between OpenAI's o1-preview and Google's gemini-pro-1.5-002.
Recent advancements in automatic code generation using large language models (LLMs) have brought us closer to fully automated secure software development. However, existing approaches often rely on a single agent for code generation, which struggles to produce secure, vulnerability-free code. Traditional program synthesis with LLMs has primarily focused on functional correctness, often neglecting critical dynamic security implications that happen during runtime. To address these challenges, we propose AutoSafeCoder, a multi-agent framework that leverages LLM-driven agents for code generation, vulnerability analysis, and security enhancement through continuous collaboration. The framework consists of three agents: a Coding Agent responsible for code generation, a Static Analyzer Agent identifying vulnerabilities, and a Fuzzing Agent performing dynamic testing using a mutation-based fuzzing approach to detect runtime errors. Our contribution focuses on ensuring the safety of multi-agent code generation by integrating dynamic and static testing in an iterative process during code generation by LLM that improves security. Experiments using the SecurityEval dataset demonstrate a 13% reduction in code vulnerabilities compared to baseline LLMs, with no compromise in functionality.
During software evolution, it is advocated that test code should co-evolve with production code. In real development scenarios, test updating may lag behind production code changing, which may cause compilation failure or bring other troubles. Existing techniques based on pre-trained language models can be directly adopted to repair obsolete tests caused by such unsynchronized code changes, especially syntactic-related ones. However, the lack of task-oriented contextual information affects the repair accuracy on large-scale projects. Starting from an obsolete test, the key challenging task is precisely identifying and constructing Test-Repair-Oriented Contexts (TROCtxs) from the whole repository within a limited token size. In this paper, we propose SYNTER (SYNtactic-breaking-changes-induced TEst Repair), a novel approach based on LLMs to automatically repair obsolete test cases via precise and concise TROCtxs construction. Inspired by developers' programming practices, we design three types of TROCtx: class context, usage context, and environment context. Given an obsolete test case to repair, SYNTER firstly collects the related code information for each type of TROCtx through static analysis techniques automatically. Then, it generates reranking queries to identify the most relevant TROCtxs, which will be taken as the repair-required key contexts and be input to the large language model for the final test repair. To evaluate the effectiveness of SYNTER, we construct a benchmark dataset that contains a set of obsolete tests caused by syntactic breaking changes. The experimental results show that SYNTER outperforms baseline approaches both on textual- and intent-matching metrics. With the augmentation of constructed TROCtxs, hallucinations are reduced by 57.1%.
The swift progress and widespread acceptance of artificial intelligence (AI) systems highlight a pressing requirement to comprehend both the capabilities and potential risks associated with AI. Given the linguistic complexity, cultural richness, and underrepresented status of Arabic in AI research, there is a pressing need to focus on Large Language Models (LLMs) performance and safety for Arabic-related tasks. Despite some progress in their development, there is a lack of comprehensive trustworthiness evaluation benchmarks, which presents a major challenge in accurately assessing and improving the safety of LLMs when prompted in Arabic. In this paper, we introduce AraTrust, the first comprehensive trustworthiness benchmark for LLMs in Arabic. AraTrust comprises 522 human-written multiple-choice questions addressing diverse dimensions related to truthfulness, ethics, safety, physical health, mental health, unfairness, illegal activities, privacy, and offensive language. We evaluated a set of LLMs against our benchmark to assess their trustworthiness. GPT-4 was the most trustworthy LLM, while open-source models, particularly AceGPT 7B and Jais 13B, struggled to achieve a score of 60% in our benchmark.
The safety of Large Language Models (LLMs) has gained increasing attention in recent years, but there still lacks a comprehensive approach for detecting safety issues within LLMs' responses in an aligned, customizable and explainable manner. In this paper, we propose ShieldLM, an LLM-based safety detector, which aligns with common safety standards, supports customizable detection rules, and provides explanations for its decisions. To train ShieldLM, we compile a large bilingual dataset comprising 14,387 query-response pairs, annotating the safety of responses based on various safety standards. Through extensive experiments, we demonstrate that ShieldLM surpasses strong baselines across four test sets, showcasing remarkable customizability and explainability. Besides performing well on standard detection datasets, ShieldLM has also been shown to be effective as a safety evaluator for advanced LLMs. ShieldLM is released at \url{https://github.com/thu-coai/ShieldLM} to support accurate and explainable safety detection under various safety standards.
Computer science has historically presented barriers for non-native English speaking (NNES) students, often due to language and terminology challenges. With the rise of large language models (LLMs), there is potential to leverage this technology to support NNES students more effectively. Recent implementations of LLMs as tutors in classrooms have shown promising results. In this study, we deployed an LLM tutor in an accelerated introductory computing course to evaluate its effectiveness specifically for NNES students. Key insights for LLM tutor use are as follows: NNES students signed up for the LLM tutor at a similar rate to native English speakers (NES); NNES students used the system at a lower rate than NES students -- to a small effect; NNES students asked significantly more questions in languages other than English compared to NES students, with many of the questions being multilingual by incorporating English programming keywords. Results for views of the LLM tutor are as follows: both NNES and NES students appreciated the LLM tutor for its accessibility, conversational style, and the guardrails put in place to guide users to answers rather than directly providing solutions; NNES students highlighted its approachability as they did not need to communicate in perfect English; NNES students rated help-seeking preferences of online resources higher than NES students; Many NNES students were unfamiliar with computing terminology in their native languages. These results suggest that LLM tutors can be a valuable resource for NNES students in computing, providing tailored support that enhances their learning experience and overcomes language barriers.
Multi-agent strategies have emerged as a promising approach to enhance the reasoning abilities of Large Language Models (LLMs) by assigning specialized roles in the problem-solving process. Concurrently, Tree of Thoughts (ToT) methods have shown potential in improving reasoning for complex question-answering tasks by exploring diverse reasoning paths. A critical limitation in multi-agent reasoning is the 'Reasoner' agent's shallow exploration of reasoning paths. While ToT strategies could help mitigate this problem, they may generate flawed reasoning branches, which could harm the trustworthiness of the final answer. To leverage the strengths of both multi-agent reasoning and ToT strategies, we introduce a novel approach combining ToT-based Reasoner agents with a Thought Validator agent. Multiple Reasoner agents operate in parallel, employing ToT to explore diverse reasoning paths. The Thought Validator then scrutinizes these paths, considering a Reasoner's conclusion only if its reasoning is valid. This method enables a more robust voting strategy by discarding faulty reasoning paths, enhancing the system's ability to tackle tasks requiring systematic and trustworthy reasoning. Our method demonstrates superior performance compared to existing techniques when evaluated on the GSM8K dataset, outperforming the standard ToT strategy by an average 5.6% across four LLMs. The code and related content can be found in: https://github.com/SecureAIAutonomyLab/MA-ToT
We introduce GamePlot, an LLM-powered assistant that supports game designers in crafting immersive narratives for turn-based games, and allows them to test these games through a collaborative game play and refine the plot throughout the process. Our user study with 14 game designers shows high levels of both satisfaction with the generated game plots and sense of ownership over the narratives, but also reconfirms that LLM are limited in their ability to generate complex and truly innovative content. We also show that diverse user populations have different expectations from AI assistants, and encourage researchers to study how tailoring assistants to diverse user groups could potentially lead to increased job satisfaction and greater creativity and innovation over time.
Large Language Models (LLMs) have shown impressive capabilities in code generation for popular programming languages. However, their performance on Low-Resource Programming Languages (LRPLs) and Domain-Specific Languages (DSLs) remains a significant challenge, affecting millions of developers-3.5 million users in Rust alone-who cannot fully utilize LLM capabilities. LRPLs and DSLs encounter unique obstacles, including data scarcity and, for DSLs, specialized syntax that is poorly represented in general-purpose datasets. Addressing these challenges is crucial, as LRPLs and DSLs enhance development efficiency in specialized domains, such as finance and science. While several surveys discuss LLMs in software engineering, none focus specifically on the challenges and opportunities associated with LRPLs and DSLs. Our survey fills this gap by systematically reviewing the current state, methodologies, and challenges in leveraging LLMs for code generation in these languages. We filtered 111 papers from over 27,000 published studies between 2020 and 2024 to evaluate the capabilities and limitations of LLMs in LRPLs and DSLs. We report the LLMs used, benchmarks, and metrics for evaluation, strategies for enhancing performance, and methods for dataset collection and curation. We identified four main evaluation techniques and several metrics for assessing code generation in LRPLs and DSLs. Our analysis categorizes improvement methods into six groups and summarizes novel architectures proposed by researchers. Despite various techniques and metrics, a standard approach and benchmark dataset for evaluating code generation in LRPLs and DSLs are lacking. This survey serves as a resource for researchers and practitioners at the intersection of LLMs, software engineering, and specialized programming languages, laying the groundwork for future advancements in code generation for LRPLs and DSLs.
An unintended consequence of the vast pretraining of Large Language Models (LLMs) is the verbatim memorization of fragments of their training data, which may contain sensitive or copyrighted information. In recent years, unlearning has emerged as a solution to effectively remove sensitive knowledge from models after training. Yet, recent work has shown that supposedly deleted information can still be extracted by malicious actors through various attacks. Still, current attacks retrieve sets of possible candidate generations and are unable to pinpoint the output that contains the actual target information. We propose activation steering as a method for exact information retrieval from unlearned LLMs. We introduce a novel approach to generating steering vectors, named Anonymized Activation Steering. Additionally, we develop a simple word frequency method to pinpoint the correct answer among a set of candidates when retrieving unlearned information. Our evaluation across multiple unlearning techniques and datasets demonstrates that activation steering successfully recovers general knowledge (e.g., widely known fictional characters) while revealing limitations in retrieving specific information (e.g., details about non-public individuals). Overall, our results demonstrate that exact information retrieval from unlearned models is possible, highlighting a severe vulnerability of current unlearning techniques.
Credibility signals represent a wide range of heuristics typically used by journalists and fact-checkers to assess the veracity of online content. Automating the extraction of credibility signals presents significant challenges due to the necessity of training high-accuracy, signal-specific extractors, coupled with the lack of sufficiently large annotated datasets. This paper introduces Pastel (Prompted weAk Supervision wiTh crEdibility signaLs), a weakly supervised approach that leverages large language models (LLMs) to extract credibility signals from web content, and subsequently combines them to predict the veracity of content without relying on human supervision. We validate our approach using four article-level misinformation detection datasets, demonstrating that Pastel outperforms zero-shot veracity detection by 38.3% and achieves 86.7% of the performance of the state-of-the-art system trained with human supervision. Moreover, in cross-domain settings where training and testing datasets originate from different domains, Pastel significantly outperforms the state-of-the-art supervised model by 63%. We further study the association between credibility signals and veracity, and perform an ablation study showing the impact of each signal on model performance. Our findings reveal that 12 out of the 19 proposed signals exhibit strong associations with veracity across all datasets, while some signals show domain-specific strengths.
Recent breakthroughs in large language models (LLMs) have generated both interest and concern about their potential adoption as accessible information sources or communication tools across different domains. In public health -- where stakes are high and impacts extend across populations -- adopting LLMs poses unique challenges that require thorough evaluation. However, structured approaches for assessing potential risks in public health remain under-explored. To address this gap, we conducted focus groups with health professionals and health issue experiencers to unpack their concerns, situated across three distinct and critical public health issues that demand high-quality information: vaccines, opioid use disorder, and intimate partner violence. We synthesize participants' perspectives into a risk taxonomy, distinguishing and contextualizing the potential harms LLMs may introduce when positioned alongside traditional health communication. This taxonomy highlights four dimensions of risk in individual behaviors, human-centered care, information ecosystem, and technology accountability. For each dimension, we discuss specific risks and example reflection questions to help practitioners adopt a risk-reflexive approach. This work offers a shared vocabulary and reflection tool for experts in both computing and public health to collaboratively anticipate, evaluate, and mitigate risks in deciding when to employ LLM capabilities (or not) and how to mitigate harm when they are used.
In the era of data-driven decision-making, accurate table-level representations and efficient table recommendation systems are becoming increasingly crucial for improving table management, discovery, and analysis. However, existing approaches to tabular data representation often face limitations, primarily due to their focus on cell-level tasks and the lack of high-quality training data. To address these challenges, we first formulate a clear definition of table similarity in the context of data transformation activities within data-driven enterprises. This definition serves as the foundation for synthetic data generation, which require a well-defined data generation process. Building on this, we propose a novel synthetic data generation pipeline that harnesses the code generation and data manipulation capabilities of Large Language Models (LLMs) to create a large-scale synthetic dataset tailored for table-level representation learning. Through manual validation and performance comparisons on the table recommendation task, we demonstrate that the synthetic data generated by our pipeline aligns with our proposed definition of table similarity and significantly enhances table representations, leading to improved recommendation performance.
Recently, there has been increasing interest in using Large Language Models (LLMs) to construct complex multi-agent systems to perform tasks such as compiling literature reviews, drafting consumer reports, and planning vacations. Many tools and libraries exist for helping create such systems, however none support recursive multi-agent systems -- where the models themselves flexibly decide when to delegate tasks and how to organize their delegation structure. In this work, we introduce ReDel: a toolkit for recursive multi-agent systems that supports custom tool-use, delegation schemes, event-based logging, and interactive replay in an easy-to-use web interface. We show that, using ReDel, we are able to easily identify potential areas of improvements through the visualization and debugging tools. Our code, documentation, and PyPI package are open-source and free to use under the MIT license at https://github.com/zhudotexe/redel.
Multilingual LLMs have achieved remarkable benchmark performance, but we find they continue to underperform on non-Latin script languages across contemporary LLM families. This discrepancy arises from the fact that LLMs are pretrained with orthographic scripts, which are dominated by Latin characters that obscure their shared phonology with non-Latin scripts. We propose leveraging phonemic transcriptions as complementary signals to induce script-invariant representations. Our study demonstrates that integrating phonemic signals improves performance across both non-Latin and Latin languages, with a particularly significant impact on closing the performance gap between the two. Through detailed experiments, we show that phonemic and orthographic scripts retrieve distinct examples for in-context learning (ICL). This motivates our proposed Mixed-ICL retrieval strategy, where further aggregation leads to our significant performance improvements for both Latin script languages (up to 12.6%) and non-Latin script languages (up to 15.1%) compared to randomized ICL retrieval.
In this paper, we present a dynamic semantic clustering approach inspired by the Chinese Restaurant Process, aimed at addressing uncertainty in the inference of Large Language Models (LLMs). We quantify uncertainty of an LLM on a given query by calculating entropy of the generated semantic clusters. Further, we propose leveraging the (negative) likelihood of these clusters as the (non)conformity score within Conformal Prediction framework, allowing the model to predict a set of responses instead of a single output, thereby accounting for uncertainty in its predictions. We demonstrate the effectiveness of our uncertainty quantification (UQ) technique on two well known question answering benchmarks, COQA and TriviaQA, utilizing two LLMs, Llama2 and Mistral. Our approach achieves SOTA performance in UQ, as assessed by metrics such as AUROC, AUARC, and AURAC. The proposed conformal predictor is also shown to produce smaller prediction sets while maintaining the same probabilistic guarantee of including the correct response, in comparison to existing SOTA conformal prediction baseline.
Despite the popularity of large language model (LLM) quantization for inference acceleration, significant uncertainty remains regarding the accuracy-performance trade-offs associated with various quantization formats. We present a comprehensive empirical study of quantized accuracy, evaluating popular quantization formats (FP8, INT8, INT4) across academic benchmarks and real-world tasks, on the entire Llama-3.1 model family. Additionally, our study examines the difference in text generated by quantized models versus their uncompressed counterparts. Beyond benchmarks, we also present a couple of quantization improvements which allowed us to obtain state-of-the-art accuracy recovery results. Our investigation, encompassing over 500,000 individual evaluations, yields several key findings: (1) FP8 weight and activation quantization (W8A8-FP) is lossless across all model scales, (2) INT8 weight and activation quantization (W8A8-INT), when properly tuned, incurs surprisingly low 1-3% accuracy degradation, and (3) INT4 weight-only quantization (W4A16-INT) is competitive with 8-bit integer weight and activation quantization. To address the question of the "best" format for a given deployment environment, we conduct inference performance analysis using the popular open-source vLLM framework on various GPU architectures. We find that W4A16 offers the best cost-efficiency for synchronous deployments, and for asynchronous deployment on mid-tier GPUs. At the same time, W8A8 formats excel in asynchronous "continuous batching" deployment of mid- and large-size models on high-end GPUs. Our results provide a set of practical guidelines for deploying quantized LLMs across scales and performance requirements.
Large Language Models (LLMs) have demonstrated impressive capabilities across diverse Natural Language Processing (NLP) tasks, including language understanding, reasoning, and generation. However, general-domain LLMs often struggle with financial tasks due to the technical and specialized nature of financial texts. This study investigates the efficacy of instruction fine-tuning smaller-scale LLMs, including Mistral-7B, Llama3-8B, and Phi3-mini, to enhance their performance in financial text classification tasks. We fine-tuned both instruction-tuned and base models across four financial classification tasks, achieving significant improvements in task-specific performance. Furthermore, we evaluated the zero-shot capabilities of these fine-tuned models on three unseen complex financial tasks, including argument classification, deal completeness classification, and causal classification. Our results indicate while base model fine-tuning led to greater degradation, instruction-tuned models maintained more robust performance. To address this degradation, we employed model merging techniques, integrating single-task domain-specific fine-tuned models with the base model. Using this merging method resulted in significant enhancements in zero-shot performance, even exceeding the original model's accuracy on certain datasets. Our findings underscore the effectiveness of instruction fine-tuning and model merging for adapting LLMs to specialized financial text classification tasks.
Customer Relationship Management (CRM) systems are vital for modern enterprises, providing a foundation for managing customer interactions and data. Integrating AI agents into CRM systems can automate routine processes and enhance personalized service. However, deploying and evaluating these agents is challenging due to the lack of realistic benchmarks that reflect the complexity of real-world CRM tasks. To address this issue, we introduce CRMArena, a novel benchmark designed to evaluate AI agents on realistic tasks grounded in professional work environments. Following guidance from CRM experts and industry best practices, we designed CRMArena with nine customer service tasks distributed across three personas: service agent, analyst, and manager. The benchmark includes 16 commonly used industrial objects (e.g., account, order, knowledge article, case) with high interconnectivity, along with latent variables (e.g., complaint habits, policy violations) to simulate realistic data distributions. Experimental results reveal that state-of-the-art LLM agents succeed in less than 40% of the tasks with ReAct prompting, and less than 55% even with function-calling abilities. Our findings highlight the need for enhanced agent capabilities in function-calling and rule-following to be deployed in real-world work environments. CRMArena is an open challenge to the community: systems that can reliably complete tasks showcase direct business value in a popular work environment.
Large language models (LLMs) exhibit remarkable capabilities on not just language tasks, but also various tasks that are not linguistic in nature, such as logical reasoning and social inference. In the human brain, neuroscience has identified a core language system that selectively and causally supports language processing. We here ask whether similar specialization for language emerges in LLMs. We identify language-selective units within 18 popular LLMs, using the same localization approach that is used in neuroscience. We then establish the causal role of these units by demonstrating that ablating LLM language-selective units -- but not random units -- leads to drastic deficits in language tasks. Correspondingly, language-selective LLM units are more aligned to brain recordings from the human language system than random units. Finally, we investigate whether our localization method extends to other cognitive domains: while we find specialized networks in some LLMs for reasoning and social capabilities, there are substantial differences among models. These findings provide functional and causal evidence for specialization in large language models, and highlight parallels with the functional organization in the brain.
The development of large language models (LLMs) has been instrumental in advancing state-of-the-art natural language processing applications. Training LLMs with billions of parameters and trillions of tokens require sophisticated distributed systems that enable composing and comparing several state-of-the-art techniques in order to efficiently scale across thousands of accelerators. However, existing solutions are complex, scattered across multiple libraries/repositories, lack interoperability, and are cumbersome to maintain. Thus, curating and empirically comparing training recipes require non-trivial engineering effort. This paper introduces TorchTitan, an open-source, PyTorch-native distributed training system that unifies state-of-the-art techniques, streamlining integration and reducing overhead. TorchTitan enables 3D parallelism in a modular manner with elastic scaling, providing comprehensive logging, checkpointing, and debugging tools for production-ready training. It also incorporates hardware-software co-designed solutions, leveraging features like Float8 training and SymmetricMemory. As a flexible test bed, TorchTitan facilitates custom recipe curation and comparison, allowing us to develop optimized training recipes for Llama 3.1 and provide guidance on selecting techniques for maximum efficiency based on our experiences. We thoroughly assess TorchTitan on the Llama 3.1 family of LLMs, spanning 8 billion to 405 billion parameters, and showcase its exceptional performance, modular composability, and elastic scalability. By stacking training optimizations, we demonstrate accelerations of 65.08% with 1D parallelism at the 128-GPU scale (Llama 3.1 8B), an additional 12.59% with 2D parallelism at the 256-GPU scale (Llama 3.1 70B), and an additional 30% with 3D parallelism at the 512-GPU scale (Llama 3.1 405B) on NVIDIA H100 GPUs over optimized baselines.
The advent of Large Language Models (LLMs) have shown promise in various creative domains, including culinary arts. However, many LLMs still struggle to deliver the desired level of culinary creativity, especially when tasked with adapting recipes to meet specific cultural requirements. This study focuses on cuisine transfer-applying elements of one cuisine to another-to assess LLMs' culinary creativity. We employ a diverse set of LLMs to generate and evaluate culturally adapted recipes, comparing their evaluations against LLM and human judgments. We introduce the ASH (authenticity, sensitivity, harmony) benchmark to evaluate LLMs' recipe generation abilities in the cuisine transfer task, assessing their cultural accuracy and creativity in the culinary domain. Our findings reveal crucial insights into both generative and evaluative capabilities of LLMs in the culinary domain, highlighting strengths and limitations in understanding and applying cultural nuances in recipe creation. The code and dataset used in this project will be openly available in \url{http://github.com/dmis-lab/CulinaryASH}.
Quantization leverages lower-precision weights to reduce the memory usage of large language models (LLMs) and is a key technique for enabling their deployment on commodity hardware. While LLM quantization's impact on utility has been extensively explored, this work for the first time studies its adverse effects from a security perspective. We reveal that widely used quantization methods can be exploited to produce a harmful quantized LLM, even though the full-precision counterpart appears benign, potentially tricking users into deploying the malicious quantized model. We demonstrate this threat using a three-staged attack framework: (i) first, we obtain a malicious LLM through fine-tuning on an adversarial task; (ii) next, we quantize the malicious model and calculate constraints that characterize all full-precision models that map to the same quantized model; (iii) finally, using projected gradient descent, we tune out the poisoned behavior from the full-precision model while ensuring that its weights satisfy the constraints computed in step (ii). This procedure results in an LLM that exhibits benign behavior in full precision but when quantized, it follows the adversarial behavior injected in step (i). We experimentally demonstrate the feasibility and severity of such an attack across three diverse scenarios: vulnerable code generation, content injection, and over-refusal attack. In practice, the adversary could host the resulting full-precision model on an LLM community hub such as Hugging Face, exposing millions of users to the threat of deploying its malicious quantized version on their devices.
This paper proposes a novel approach to evaluate Counter Narrative (CN) generation using a Large Language Model (LLM) as an evaluator. We show that traditional automatic metrics correlate poorly with human judgements and fail to capture the nuanced relationship between generated CNs and human perception. To alleviate this, we introduce a model ranking pipeline based on pairwise comparisons of generated CNs from different models, organized in a tournament-style format. The proposed evaluation method achieves a high correlation with human preference, with a $\rho$ score of 0.88. As an additional contribution, we leverage LLMs as zero-shot CN generators and provide a comparative analysis of chat, instruct, and base models, exploring their respective strengths and limitations. Through meticulous evaluation, including fine-tuning experiments, we elucidate the differences in performance and responsiveness to domain-specific data. We conclude that chat-aligned models in zero-shot are the best option for carrying out the task, provided they do not refuse to generate an answer due to security concerns.
Video Temporal Grounding (VTG) is a crucial capability for video understanding models and plays a vital role in downstream tasks such as video browsing and editing. To effectively handle various tasks simultaneously and enable zero-shot prediction, there is a growing trend in employing video LLMs for VTG tasks. However, current video LLM-based methods rely exclusively on natural language generation, lacking the ability to model the clear structure inherent in videos, which restricts their effectiveness in tackling VTG tasks. To address this issue, this paper first formally introduces causal event modeling framework, which represents videos as sequences of events, and predict the current event using previous events, video inputs, and textural instructions. Each event consists of three components: timestamps, salient scores, and textual captions. We then propose a novel task-interleaved video LLM called TRACE to effectively implement the causal event modeling framework in practice. The TRACE processes visual frames, timestamps, salient scores, and text as distinct tasks, employing various encoders and decoding heads for each. Task tokens are arranged in an interleaved sequence according to the causal event modeling framework's formulation. Extensive experiments on various VTG tasks and datasets demonstrate the superior performance of TRACE compared to state-of-the-art video LLMs. Our model and code are available at \url{https://github.com/gyxxyg/TRACE}.
As the development and application of Large Language Models (LLMs) continue to advance rapidly, enhancing their trustworthiness and aligning them with human preferences has become a critical area of research. Traditional methods rely heavily on extensive data for Reinforcement Learning from Human Feedback (RLHF), but representation engineering offers a new, training-free approach. This technique leverages semantic features to control the representation of LLM's intermediate hidden states, enabling the model to meet specific requirements such as increased honesty or heightened safety awareness. However, a significant challenge arises when attempting to fulfill multiple requirements simultaneously. It proves difficult to encode various semantic contents, like honesty and safety, into a singular semantic feature, restricting its practicality. In this work, we address this issue through ``Sparse Activation Control''. By delving into the intrinsic mechanisms of LLMs, we manage to identify and pinpoint components that are closely related to specific tasks within the model, i.e., attention heads. These heads display sparse characteristics that allow for near-independent control over different tasks. Our experiments, conducted on the open-source Llama series models, have yielded encouraging results. The models were able to align with human preferences on issues of safety, factuality, and bias concurrently.
Assertion-based verification (ABV) is a critical method to ensure logic designs comply with their architectural specifications. ABV requires assertions, which are generally converted from specifications through human interpretation by verification engineers. Existing methods for generating assertions from specification documents are limited to sentences extracted by engineers, discouraging their practical applications. In this work, we present AssertLLM, an automatic assertion generation framework that processes complete specification documents. AssertLLM can generate assertions from both natural language and waveform diagrams in specification files. It first converts unstructured specification sentences and waveforms into structured descriptions using natural language templates. Then, a customized Large Language Model (LLM) generates the final assertions based on these descriptions. Our evaluation demonstrates that AssertLLM can generate more accurate and higher-quality assertions compared to GPT-4o and GPT-3.5.
Assertion-based verification (ABV) is a critical method for ensuring design circuits comply with their architectural specifications, which are typically described in natural language. This process often requires human interpretation by verification engineers to convert these specifications into functional verification assertions. Existing methods for generating assertions from natural language specifications are limited to sentences extracted by engineers, discouraging its practical application. In this work, we present AssertLLM, an automatic assertion generation framework that processes complete specification files. AssertLLM breaks down the complex task into three phases, incorporating three customized Large Language Models (LLMs) for extracting structural specifications, mapping signal definitions, and generating assertions. Our evaluation of AssertLLM on a full design, encompassing 23 I/O signals, demonstrates that 89\% of the generated assertions are both syntactically and functionally accurate.
Retrieval-Augmented Generation (RAG) mitigates the issue of hallucination in Large Language Models (LLMs) by integrating information retrieval techniques. However, in the tourism domain, since the query is usually brief and the content in the database is diverse, existing RAG may contain a significant amount of irrelevant or contradictory information contents after retrieval. To address this challenge, we propose the QCG-Rerank model. This model first performs an initial retrieval to obtain candidate chunks and then enhances semantics by extracting critical information to expand the original query. Next, we utilize the expanded query and candidate chunks to calculate similarity scores as the initial transition probability and construct the chunks graph. Subsequently, We iteratively compute the transition probabilities based on an initial estimate until convergence. The chunks with the highest score are selected and input into the LLMs to generate responses. We evaluate the model on Cultour, IIRC, StrategyQA, HotpotQA, SQuAD, and MuSiQue datasets. The experimental results demonstrate the effectiveness and superiority of the QCG-Rerank method.
Current speech-based LLMs are predominantly trained on extensive ASR and TTS datasets, excelling in tasks related to these domains. However, their ability to handle direct speech-to-speech conversations remains notably constrained. These models often rely on an ASR-to-TTS chain-of-thought pipeline, converting speech into text for processing before generating audio responses, which introduces latency and loses audio features. We propose a method that implicitly internalizes ASR chain of thought into a speech LLM, enhancing its native speech understanding capabilities. Our approach reduces latency and improves the model's native understanding of speech, paving the way for more efficient and natural real-time audio interactions. We also release a large-scale synthetic conversational dataset to facilitate further research.
Large language models (LLMs) now exhibit near human-level performance in various tasks, but their performance drops drastically after a handful of high-resource languages due to the imbalance in pre-training data. Inspired by the human process of second language acquisition, particularly code-switching (the practice of language alternation in a conversation), we propose code-switching curriculum learning (CSCL) to enhance cross-lingual transfer for LLMs. CSCL mimics the stages of human language learning by progressively training models with a curriculum consisting of 1) token-level code-switching, 2) sentence-level code-switching, and 3) monolingual corpora. Using Qwen 2 as our underlying model, we demonstrate the efficacy of the CSCL in improving language transfer to Korean, achieving significant performance gains compared to monolingual continual pre-training methods. Ablation studies reveal that both token- and sentence-level code-switching significantly enhance cross-lingual transfer and that curriculum learning amplifies these effects. We also extend our findings into various languages, including Japanese (high-resource) and Indonesian (low-resource), and using two additional models (Gemma 2 and Phi 3.5). We further show that CSCL mitigates spurious correlations between language resources and safety alignment, presenting a robust, efficient framework for more equitable language transfer in LLMs. We observe that CSCL is effective for low-resource settings where high-quality, monolingual corpora for language transfer are hardly available.
Mathematical theorem proving is an important testbed for large language models' deep and abstract reasoning capability. This paper focuses on improving LLMs' ability to write proofs in formal languages that permit automated proof verification/evaluation. Most previous results provide human-written lemmas to the theorem prover, which is an arguably oversimplified setting that does not sufficiently test the provers' planning and decomposition capabilities. Instead, we work in a more natural setup where the lemmas that are directly relevant to the theorem are not given to the theorem prover at test time. We design an RL-based training algorithm that encourages the model to decompose a theorem into lemmas, prove the lemmas, and then prove the theorem by using the lemmas. Our reward mechanism is inspired by how mathematicians train themselves: even if a theorem is too challenging to be proved by the current model, a positive reward is still given to the model for any correct and novel lemmas that are proposed and proved in this process. During training, our model proposes and proves lemmas that are not in the training dataset. In fact, these newly-proposed correct lemmas consist of 37.7% of the training replay buffer when we train on the dataset extracted from Archive of Formal Proofs (AFP). The model trained by our RL algorithm outperforms that trained by supervised finetuning, improving the pass rate from 40.8% to 45.5% on AFP test set, and from 36.5% to 39.5% on an out-of-distribution test set.
This paper investigates supervised fine-tuning of large language models (LLMs) to improve their pedagogical alignment in computing education, addressing concerns that LLMs may hinder learning outcomes. The project utilised a proprietary dataset of 2,500 high quality question/answer pairs from programming course forums, and explores two research questions: the suitability of university course forums in contributing to fine-tuning datasets, and how supervised fine-tuning can improve LLMs' alignment with educational principles such as constructivism. Initial findings suggest benefits in pedagogical alignment of LLMs, with deeper evaluations required.
ToolPlanner: A Tool Augmented LLM for Multi Granularity Instructions with Path Planning and Feedback
Recently, tool-augmented LLMs have gained increasing attention. Given an instruction, tool-augmented LLMs can interact with various external tools in multiple rounds and provide a final answer. However, previous LLMs were trained on overly detailed instructions, which included API names or parameters, while real users would not explicitly mention these API details. This leads to a gap between trained LLMs and real-world scenarios. In addition, most works ignore whether the interaction process follows the instruction. To address these issues, we constructed a training dataset called MGToolBench, which contains statement and category-level instructions to better reflect real-world scenarios. In addition, we propose ToolPlanner, a two-stage reinforcement learning framework that utilizes path planning and two feedback mechanisms to enhance the LLM's task completion and instruction-following capabilities. Experimental results show that ToolPlanner significantly improves the Match Rate, Pass Rate and Win Rate by 26.8%, 20.2%, and 5.6% compared to the SOTA model. Human evaluation verifies that the multi-granularity instructions can better align with users' usage habits. Our data and code will be released upon acceptance.
Mathematical optimization is fundamental to decision-making across diverse domains, from operations research to healthcare. Yet, translating real-world problems into optimization models remains a formidable challenge, often demanding specialized expertise. This paper formally introduces the concept of $\textbf{autoformulation}$ -- an automated approach to creating optimization models from natural language descriptions for commercial solvers. We identify the three core challenges of autoformulation: (1) defining the vast, problem-dependent hypothesis space, (2) efficiently searching this space under uncertainty, and (3) evaluating formulation correctness (ensuring a formulation accurately represents the problem). To address these challenges, we introduce a novel method leveraging $\textit{Large Language Models}$ (LLMs) within a $\textit{Monte-Carlo Tree Search}$ framework. This approach systematically explores the space of possible formulations by exploiting the hierarchical nature of optimization modeling. LLMs serve two key roles: as dynamic formulation hypothesis generators and as evaluators of formulation correctness. To enhance search efficiency, we introduce a pruning technique to remove trivially equivalent formulations. Empirical evaluations across benchmarks containing linear and mixed-integer programming problems demonstrate our method's superior performance. Additionally, we observe significant efficiency gains from employing LLMs for correctness evaluation and from our pruning techniques.
Large language models (LLMs) have demonstrated the potential to mimic human social intelligence. However, most studies focus on simplistic and static self-report or performance-based tests, which limits the depth and validity of the analysis. In this paper, we developed a novel framework, InterIntent, to assess LLMs' social intelligence by mapping their ability to understand and manage intentions in a game setting. We focus on four dimensions of social intelligence: situational awareness, self-regulation, self-awareness, and theory of mind. Each dimension is linked to a specific game task: intention selection, intention following, intention summarization, and intention guessing. Our findings indicate that while LLMs exhibit high proficiency in selecting intentions, achieving an accuracy of 88%, their ability to infer the intentions of others is significantly weaker, trailing human performance by 20%. Additionally, game performance correlates with intention understanding, highlighting the importance of the four components towards success in this game. These findings underline the crucial role of intention understanding in evaluating LLMs' social intelligence and highlight the potential of using social deduction games as a complex testbed to enhance LLM evaluation. InterIntent contributes a structured approach to bridging the evaluation gap in social intelligence within multiplayer games.
Knowledge graphs (KGs) are increasingly utilized for data integration, representation, and visualization. While KG population is critical, it is often costly, especially when data must be extracted from unstructured text in natural language, which presents challenges, such as ambiguity and complex interpretations. Large Language Models (LLMs) offer promising capabilities for such tasks, excelling in natural language understanding and content generation. However, their tendency to ``hallucinate'' can produce inaccurate outputs. Despite these limitations, LLMs offer rapid and scalable processing of natural language data, and with prompt engineering and fine-tuning, they can approximate human-level performance in extracting and structuring data for KGs. This study investigates LLM effectiveness for the KG population, focusing on the Enslaved.org Hub Ontology. In this paper, we report that compared to the ground truth, LLM's can extract ~90% of triples, when provided a modular ontology as guidance in the prompts.
Large Language Models are commonly used in retrieval-augmented applications to execute user instructions based on data from external sources. For example, modern search engines use LLMs to answer queries based on relevant search results; email plugins summarize emails by processing their content through an LLM. However, the potentially untrusted provenance of these data sources can lead to prompt injection attacks, where the LLM is manipulated by natural language instructions embedded in the external data, causing it to deviate from the user's original instruction(s). We define this deviation as task drift. Task drift is a significant concern as it allows attackers to exfiltrate data or influence the LLM's output for other users. We study LLM activations as a solution to detect task drift, showing that activation deltas - the difference in activations before and after processing external data - are strongly correlated with this phenomenon. Through two probing methods, we demonstrate that a simple linear classifier can detect drift with near-perfect ROC AUC on an out-of-distribution test set. We evaluate these methods by making minimal assumptions about how user's tasks, system prompts, and attacks can be phrased. We observe that this approach generalizes surprisingly well to unseen task domains, such as prompt injections, jailbreaks, and malicious instructions, without being trained on any of these attacks. Interestingly, the fact that this solution does not require any modifications to the LLM (e.g., fine-tuning), as well as its compatibility with existing meta-prompting solutions, makes it cost-efficient and easy to deploy. To encourage further research on activation-based task inspection, decoding, and interpretability, we release our large-scale TaskTracker toolkit, featuring a dataset of over 500K instances, representations from six SoTA language models, and inspection tools.
Large language models (LLMs) are trained on data assumed to include natural language pragmatics, but do they actually behave like pragmatic speakers? We attempt to answer this question using the Rational Speech Act (RSA) framework, which models pragmatic reasoning in human communication. Using the paradigm of a reference game constructed from the TUNA corpus, we score candidate referential utterances in both a state-of-the-art LLM (Llama3-8B-Instruct) and in the RSA model, comparing and contrasting these scores. Given that RSA requires defining alternative utterances and a truth-conditional meaning function, we explore such comparison for different choices of each of these requirements. We find that while scores from the LLM have some positive correlation with those from RSA, there isn't sufficient evidence to claim that it behaves like a pragmatic speaker. This initial study paves way for further targeted efforts exploring different models and settings, including human-subject evaluation, to see if LLMs truly can, or be made to, behave like pragmatic speakers.
The emergence of Large Language Model(LLM) technologies has led to a rapidly growing demand for compute resources in models. In response, the enterprises are building large-scale multi-tenant GPU clusters with 10k or even ore GPUs. In contrast to the rapidly growing cluster size, the bandwidth of clusters has also been increasing to meet communication demands, with 800 Gbps optical modules already in practical use and 1.6 Tbps modules on the horizon. However, designing clusters that simultaneously meet the requirements of large scale and high bandwidth is challenging due to the limited capacity of electrical switch chips. Unlike electrical switch chips, the single-port bandwidth of MEMS-OCS is solely determined by the optical module, making it straightforward to achieve both bandwidth and scability requirement. In this paper, we propose an opto-electronic hybrid architecture called \textbf{LumosCore}. We address the issues of L2 protocols incompatibility potential network contention and algorithm time complexity through physical topology and logical topology design. Additionally, we design a polynomial-time complexity link reconfiguration algorithm to reconfigure MEMS-OCS with minimal time overhead. We validate the feasibility of the proposed scheme in a cluster consisting of 128 NPUs, and through simulation based on real traces, we demonstrate the superiority of \textbf{LumosCore} over traditional architectures.
Recent advancements in Large Language Models (LLMs) have showcased striking results on existing logical reasoning benchmarks, with some models even surpassing human performance. However, the true depth of their competencies and robustness in reasoning tasks remains an open question. To this end, in this paper, we focus on two popular reasoning tasks: arithmetic reasoning and code generation. Particularly, we introduce (i) a general ontology of perturbations for math and coding questions, (ii) a semi-automatic method to apply these perturbations, and (iii) two datasets, GSMORE and HUMANEVAL-CORE, respectively, of perturbed math and coding problems to probe LLM capabilities in numeric reasoning and coding tasks. Through comprehensive evaluations of both closed-source and open-source LLMs, we show a significant performance drop across all the models against the perturbed questions, suggesting that the current LLMs lack robust problem solving skills and structured reasoning abilities in many areas, as defined by our ontology. We open-source the datasets and source codes at: https://github.com/declare-lab/LLM-ReasoningTest.
While `jailbreaks' have been central to research on the safety and reliability of LLMs (large language models), the underlying mechanisms behind these attacks are not well understood. Some prior works have used linear methods to analyze jailbreak prompts or model refusal. Here, however, we compare linear and nonlinear methods to study the features in prompts that contribute to successful jailbreaks. We do this by probing for jailbreak success based only on the portions of the latent representations corresponding to prompt tokens. First, we introduce a dataset of 10,800 jailbreak attempts from 35 attack methods. We then show that different jailbreaking methods work via different nonlinear features in prompts. Specifically, we find that while probes can distinguish between successful and unsuccessful jailbreaking prompts with a high degree of accuracy, they often transfer poorly to held-out attack methods. We also show that nonlinear probes can be used to mechanistically jailbreak the LLM by guiding the design of adversarial latent perturbations. These mechanistic jailbreaks are able to jailbreak Gemma-7B-IT more reliably than 34 of the 35 techniques that it was trained on. Ultimately, our results suggest that jailbreaks cannot be thoroughly understood in terms of universal or linear prompt features alone.
Aligning large language models (LLMs) with human intent is critical for enhancing their performance across a variety of tasks. Standard alignment techniques, such as Direct Preference Optimization (DPO), often rely on the binary Bradley-Terry (BT) model, which can struggle to capture the complexities of human preferences -- particularly in the presence of noisy or inconsistent labels and frequent ties. To address these limitations, we introduce the Tie-rank Oriented Bradley-Terry model (TOBT), an extension of the BT model that explicitly incorporates ties, enabling more nuanced preference representation. Building on this, we propose Tie-rank Oriented Direct Preference Optimization (TODO), a novel alignment algorithm that leverages TOBT's ternary ranking system to improve preference alignment. In evaluations on Mistral-7B and Llama 3-8B models, TODO consistently outperforms DPO in modeling preferences across both in-distribution and out-of-distribution datasets. Additional assessments using MT Bench and benchmarks such as Piqa, ARC-c, and MMLU further demonstrate TODO's superior alignment performance. Notably, TODO also shows strong results in binary preference alignment, highlighting its versatility and potential for broader integration into LLM alignment. The implementation details can be found in https://github.com/XXares/TODO.
Reinforcement Learning from Human Feedback (RLHF) has been proven to be an effective method for preference alignment of large language models (LLMs) and is widely used in the post-training process of LLMs. However, RLHF struggles with handling multiple competing preferences. This leads to a decrease in the alignment of LLMs with human preferences. To address this issue, we propose Preference Mixture of LoRAs (PMoL) from the perspective of model architecture, which can adapt to any number of preferences to mix. PMoL combines Mixture of Experts (MoE) and Low Rank Adaptor (LoRA). This architecture is innovatively applied to the research of preference alignment and has achieved significant performance improvement. The expert group soft loss is used to enable MoE with the ability to mix preferences. Through comprehensive evaluation by the reward model and GPT-4o, the experiment results show that PMoL has superior preference mixing capabilities compared to baseline methods. PMoL achieves better preference alignment with lower training costs.
Text summarization is a well-established task within the natural language processing (NLP) community. However, the focus on controllable summarization tailored to user requirements is gaining traction only recently. While several efforts explore controllability in text summarization, the investigation of Multi-Attribute Controllable Summarization (MACS) remains limited. This work addresses this gap by examining the MACS task through the lens of large language models (LLMs), using various learning paradigms, particularly low-rank adapters. We experiment with different popular adapter fine-tuning strategies to assess the effectiveness of the resulting models in retaining cues and patterns associated with multiple controllable attributes. Additionally, we propose and evaluate a novel hierarchical adapter fusion technique to integrate learnings from two distinct controllable attributes. Subsquently, we present our findings, discuss the challenges encountered, and suggest potential avenues for advancing the MACS task.
LLM agents have the potential to revolutionize defensive cyber operations, but their offensive capabilities are not yet fully understood. To prepare for emerging threats, model developers and governments are evaluating the cyber capabilities of foundation models. However, these assessments often lack transparency and a comprehensive focus on offensive capabilities. In response, we introduce the Catastrophic Cyber Capabilities Benchmark (3CB), a novel framework designed to rigorously assess the real-world offensive capabilities of LLM agents. Our evaluation of modern LLMs on 3CB reveals that frontier models, such as GPT-4o and Claude 3.5 Sonnet, can perform offensive tasks such as reconnaissance and exploitation across domains ranging from binary analysis to web technologies. Conversely, smaller open-source models exhibit limited offensive capabilities. Our software solution and the corresponding benchmark provides a critical tool to reduce the gap between rapidly improving capabilities and robustness of cyber offense evaluations, aiding in the safer deployment and regulation of these powerful technologies.