llm - 2024_06
Navigation
Papers
This survey provides an in-depth analysis of knowledge conflicts for large language models (LLMs), highlighting the complex challenges they encounter when blending contextual and parametric knowledge. Our focus is on three categories of knowledge conflicts: context-memory, inter-context, and intra-memory conflict. These conflicts can significantly impact the trustworthiness and performance of LLMs, especially in real-world applications where noise and misinformation are common. By categorizing these conflicts, exploring the causes, examining the behaviors of LLMs under such conflicts, and reviewing available solutions, this survey aims to shed light on strategies for improving the robustness of LLMs, thereby serving as a valuable resource for advancing research in this evolving area.
Large Language Models (LLMs) still struggle with natural language reasoning tasks. Motivated by the society of minds (Minsky, 1988), we propose ReConcile, a multi-model multi-agent framework designed as a round table conference among diverse LLM agents. ReConcile enhances collaborative reasoning between LLM agents via multiple rounds of discussion, learning to convince other agents to improve their answers, and employing a confidence-weighted voting mechanism that leads to a better consensus. In each round, ReConcile initiates discussion between agents via a 'discussion prompt' that consists of (a) grouped answers and explanations generated by each agent in the previous round, (b) their confidence scores, and (c) demonstrations of answer-rectifying human explanations, used for convincing other agents. Experiments on seven benchmarks demonstrate that ReConcile significantly improves LLMs' reasoning -- both individually and as a team -- surpassing prior single-agent and multi-agent baselines by up to 11.4% and even outperforming GPT-4 on three datasets. ReConcile also flexibly incorporates different combinations of agents, including API-based, open-source, and domain-specific models, leading to an 8% improvement on MATH. Finally, we analyze the individual components of ReConcile, demonstrating that the diversity originating from different models is critical to its superior performance. Code: https://github.com/dinobby/ReConcile
Recent advancements in machine learning have significantly improved the identification of disease-associated genes from gene expression datasets. However, these processes often require extensive expertise and manual effort, limiting their scalability. Large Language Model (LLM)-based agents have shown promise in automating these tasks due to their increasing problem-solving abilities. To support the evaluation and development of such methods, we introduce GenoTEX, a benchmark dataset for the automatic exploration of gene expression data, involving the tasks of dataset selection, preprocessing, and statistical analysis. GenoTEX provides annotated code and results for solving a wide range of gene identification problems, in a full analysis pipeline that follows the standard of computational genomics. These annotations are curated by human bioinformaticians who carefully analyze the datasets to ensure accuracy and reliability. To provide baselines for these tasks, we present GenoAgents, a team of LLM-based agents designed with context-aware planning, iterative correction, and domain expert consultation to collaboratively explore gene datasets. Our experiments with GenoAgents demonstrate the potential of LLM-based approaches in genomics data analysis, while error analysis highlights the challenges and areas for future improvement. We propose GenoTEX as a promising resource for benchmarking and enhancing AI-driven methods for genomics data analysis. We make our benchmark publicly available at \url{https://github.com/Liu-Hy/GenoTex}.
Large language models (LLMs) have revolutionized lots of fields of research. Although it is well-known that fine-tuning is essential for enhancing the capabilities of LLMs, existing research suggests that there is potential redundancy in the fine-tuning process and therefore proposes to update only a subset of parameters. However, these methods fail to leverage the task-specific information to identify important parameters during training. Based on the insight that gradients inherently contain information on task-specific data, we propose Gradient-Mask Tuning (GMT), a method that selectively updates parameters during training based on their gradient information. Specifically, we compute the absolute values of the gradients and apply masking to those with relatively smaller magnitudes. Our empirical results across various tasks demonstrate that GMT not only outperforms traditional fine-tuning methods but also elevates the upper limits of LLM performance. Further analysis indicates that GMT exhibits insensitivity to mask ratio and possesses computational efficiency comparable to vanilla SFT.
Recent research in Needle-in-a-Haystack (NIAH) benchmarks has explored the capabilities of Large Language Models (LLMs) in retrieving contextual information from large text documents. However, as LLMs become increasingly integrated into software development processes, it is crucial to evaluate their performance in code-based environments. As LLMs are further developed for program synthesis, we need to ensure that LLMs can understand syntax and write syntactically correct code. As a step in ensuring LLMs understand syntax, LLMs can be evaluated in their ability to find and detect syntax bugs. Our benchmark, Bug In The Code Stack (BICS), is designed to assess the ability of LLMs to identify simple syntax bugs within large source code. Our findings reveal three key insights: (1) code-based environments pose significantly more challenge compared to text-based environments for retrieval tasks, (2) there is a substantial performance disparity among different models, and (3) there is a notable correlation between longer context lengths and performance degradation, though the extent of this degradation varies between models.
Translation Quality Evaluation (TQE) is an essential step of the modern translation production process. TQE is critical in assessing both machine translation (MT) and human translation (HT) quality without reference translations. The ability to evaluate or even simply estimate the quality of translation automatically may open significant efficiency gains through process optimisation. This work examines whether the state-of-the-art large language models (LLMs) can be used for this purpose. We take OpenAI models as the best state-of-the-art technology and approach TQE as a binary classification task. On eight language pairs including English to Italian, German, French, Japanese, Dutch, Portuguese, Turkish, and Chinese, our experimental results show that fine-tuned gpt3.5 can demonstrate good performance on translation quality prediction tasks, i.e. whether the translation needs to be edited. Another finding is that simply increasing the sizes of LLMs does not lead to apparent better performances on this task by comparing the performance of three different versions of OpenAI models: curie, davinci, and gpt3.5 with 13B, 175B, and 175B parameters, respectively.
Large Language Models (LLMs) like ChatGPT or Bard have revolutionized information retrieval and captivated the audience with their ability to generate custom responses in record time, regardless of the topic. In this article, we assess the capabilities of various LLMs in producing reliable, comprehensive, and sufficiently relevant responses about historical facts in French. To achieve this, we constructed a testbed comprising numerous history-related questions of varying types, themes, and levels of difficulty. Our evaluation of responses from ten selected LLMs reveals numerous shortcomings in both substance and form. Beyond an overall insufficient accuracy rate, we highlight uneven treatment of the French language, as well as issues related to verbosity and inconsistency in the responses provided by LLMs.
Large language models (LLMs) have demonstrated their prowess in generating synthetic text and images; however, their potential for generating tabular data -- arguably the most common data type in business and scientific applications -- is largely underexplored. This paper demonstrates that LLMs, used as-is, or after traditional fine-tuning, are severely inadequate as synthetic table generators. Due to the autoregressive nature of LLMs, fine-tuning with random order permutation runs counter to the importance of modeling functional dependencies, and renders LLMs unable to model conditional mixtures of distributions (key to capturing real world constraints). We showcase how LLMs can be made to overcome some of these deficiencies by making them permutation-aware.
Large language models (LLMs) encode a vast amount of world knowledge acquired from massive text datasets. Recent studies have demonstrated that LLMs can assist an embodied agent in solving complex sequential decision making tasks by providing high-level instructions. However, interactions with LLMs can be time-consuming. In many practical scenarios, it requires a significant amount of storage space that can only be deployed on remote cloud servers. Additionally, using commercial LLMs can be costly since they may charge based on usage frequency. In this paper, we explore how to enable intelligent cost-effective interactions between a down stream task oriented agent and an LLM. We find that this problem can be naturally formulated by a Markov decision process (MDP), and propose When2Ask, a reinforcement learning based approach that learns when it is necessary to query LLMs for high-level instructions to accomplish a target task. On one side, When2Ask discourages unnecessary redundant interactions, while on the other side, it enables the agent to identify and follow useful instructions from the LLM. This enables the agent to halt an ongoing plan and transition to a more suitable one based on new environmental observations. Experiments on MiniGrid and Habitat environments that entail planning sub-goals demonstrate that When2Ask learns to solve target tasks with only a few necessary interactions with the LLM, significantly reducing interaction costs in testing environments compared with baseline methods. Our code is available at: https://github.com/ZJLAB-AMMI/LLM4RL.
Retrieval Augmented Generation (RAG) represents a significant advancement in artificial intelligence combining a retrieval phase with a generative phase, with the latter typically being powered by large language models (LLMs). The current common practices in RAG involve using "instructed" LLMs, which are fine-tuned with supervised training to enhance their ability to follow instructions and are aligned with human preferences using state-of-the-art techniques. Contrary to popular belief, our study demonstrates that base models outperform their instructed counterparts in RAG tasks by 20% on average under our experimental settings. This finding challenges the prevailing assumptions about the superiority of instructed LLMs in RAG applications. Further investigations reveal a more nuanced situation, questioning fundamental aspects of RAG and suggesting the need for broader discussions on the topic; or, as Fromm would have it, "Seldom is a glance at the statistics enough to understand the meaning of the figures".
Large Language Models (LLMs) have achieved impressive results in Machine Translation (MT). However, careful evaluations by human reveal that the translations produced by LLMs still contain multiple errors. Importantly, feeding back such error information into the LLMs can lead to self-refinement and result in improved translation performance. Motivated by these insights, we introduce a systematic LLM-based self-refinement translation framework, named \textbf{TEaR}, which stands for \textbf{T}ranslate, \textbf{E}stimate, \textbf{a}nd \textbf{R}efine, marking a significant step forward in this direction. Our findings demonstrate that 1) our self-refinement framework successfully assists LLMs in improving their translation quality across a wide range of languages, whether it's from high-resource languages to low-resource ones or whether it's English-centric or centered around other languages; 2) TEaR exhibits superior systematicity and interpretability; 3) different estimation strategies yield varied impacts, directly affecting the effectiveness of the final corrections. Additionally, traditional neural translation models and evaluation models operate separately, often focusing on singular tasks due to their limited capabilities, while general-purpose LLMs possess the capability to undertake both tasks simultaneously. We further conduct cross-model correction experiments to investigate the potential relationship between the translation and evaluation capabilities of general-purpose LLMs. Our code and data are available at https://github.com/fzp0424/self_correct_mt
LLM-based agents have emerged as promising tools, which are crafted to fulfill complex tasks by iterative planning and action. However, these agents are susceptible to undesired planning hallucinations when lacking specific knowledge for expertise-intensive tasks. To address this, preliminary attempts are made to enhance planning reliability by incorporating external workflow-related knowledge. Despite the promise, such infused knowledge is mostly disorganized and diverse in formats, lacking rigorous formalization and comprehensive comparisons. Motivated by this, we formalize different formats of workflow knowledge and present FlowBench, the first benchmark for workflow-guided planning. FlowBench covers 51 different scenarios from 6 domains, with knowledge presented in diverse formats. To assess different LLMs on FlowBench, we design a multi-tiered evaluation framework. We evaluate the efficacy of workflow knowledge across multiple formats, and the results indicate that current LLM agents need considerable improvements for satisfactory planning. We hope that our challenging benchmark can pave the way for future agent planning research.
Trust plays a pivotal role in Lecturer-Student-Collaboration, encompassing teaching and research aspects. The advent of Large Language Models (LLMs) in platforms like Open AI's ChatGPT, coupled with their cost-effectiveness and high-quality results, has led to their rapid adoption among university students. However, discerning genuine student input from LLM-generated output poses a challenge for lecturers. This dilemma jeopardizes the trust relationship between lecturers and students, potentially impacting university downstream activities, particularly collaborative research initiatives. Despite attempts to establish guidelines for student LLM use, a clear framework mutually beneficial for lecturers and students in higher education remains elusive. This study addresses the research question: How does the use of LLMs by students impact Informational and Procedural Justice, influencing Team Trust and Expected Team Performance? Methodically, we applied a quantitative construct-based survey, evaluated using techniques of Structural Equation Modelling (PLS- SEM) to examine potential relationships among these constructs. Our findings based on 23 valid respondents from Ndejje University indicate that lecturers are less concerned about the fairness of LLM use per se but are more focused on the transparency of student utilization, which significantly influences Team Trust positively. This research contributes to the global discourse on integrating and regulating LLMs and subsequent models in education. We propose that guidelines should support LLM use while enforcing transparency in Lecturer-Student- Collaboration to foster Team Trust and Performance. The study contributes valuable insights for shaping policies enabling ethical and transparent LLMs usage in education to ensure effectiveness of collaborative learning environments.
Recently, considerable efforts have been directed towards compressing Large Language Models (LLMs), which showcase groundbreaking capabilities across diverse applications but entail significant deployment costs due to their large sizes. Meanwhile, much less attention has been given to mitigating the costs associated with deploying multiple LLMs of varying sizes despite its practical significance. Thus, this paper introduces \emph{any-precision LLM}, extending the concept of any-precision DNN to LLMs. Addressing challenges in any-precision LLM, we propose a lightweight method for any-precision quantization of LLMs, leveraging a post-training quantization framework, and develop a specialized software engine for its efficient serving. As a result, our solution significantly reduces the high costs of deploying multiple, different-sized LLMs by overlaying LLMs quantized to varying bit-widths, such as 3, 4, ..., $n$ bits, into a memory footprint comparable to a single $n$-bit LLM. All the supported LLMs with varying bit-widths demonstrate state-of-the-art model quality and inference throughput, proving itself to be a compelling option for deployment of multiple, different-sized LLMs. Our code is open-sourced and available online.
To detect stego (steganographic text) in complex scenarios, linguistic steganalysis (LS) with various motivations has been proposed and achieved excellent performance. However, with the development of generative steganography, some stegos have strong concealment, especially after the emergence of LLMs-based steganography, the existing LS has low detection or cannot detect them. We designed a novel LS with two modes called LSGC. In the generation mode, we created an LS-task "description" and used the generation ability of LLM to explain whether texts to be detected are stegos. On this basis, we rethought the principle of LS and LLMs, and proposed the classification mode. In this mode, LSGC deleted the LS-task "description" and used the "causalLM" LLMs to extract steganographic features. The LS features can be extracted by only one pass of the model, and a linear layer with initialization weights is added to obtain the classification probability. Experiments on strongly concealed stegos show that LSGC significantly improves detection and reaches SOTA performance. Additionally, LSGC in classification mode greatly reduces training time while maintaining high performance.
The rapid development of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) has exposed vulnerabilities to various adversarial attacks. This paper provides a comprehensive overview of jailbreaking research targeting both LLMs and MLLMs, highlighting recent advancements in evaluation benchmarks, attack techniques and defense strategies. Compared to the more advanced state of unimodal jailbreaking, multimodal domain remains underexplored. We summarize the limitations and potential research directions of multimodal jailbreaking, aiming to inspire future research and further enhance the robustness and security of MLLMs.
Software comments are critical for human understanding of software, and as such many comment generation techniques have been proposed. However, we find that a systematic evaluation of the factual accuracy of generated comments is rare; only subjective accuracy labels have been given. Evaluating comments generated by three Large Language Models (LLMs), we find that even for the best-performing LLM, roughly a fifth of its comments contained demonstrably inaccurate statements. While it seems code-comment consistency detection techniques should be able to detect inaccurate comments, we perform experiments demonstrating they have no statistically significant relationship with comment accuracy, underscoring the substantial difficulty of this problem. To tackle this, we propose the concept of document testing, in which a document is verified by using an LLM to generate tests based on the document, running those tests, and observing whether they pass or fail. Furthermore, we implement our concept to verify Java comments. Experiments demonstrate that our approach has a robust statistical relationship with comment accuracy, making headway into a problem where prior techniques failed. Qualitative evaluation also reveals the promise of our approach in gaining developer trust, while highlighting the limitations of our current implementation.
Large language models (LLMs) have achieved impressive human-like performance across various reasoning tasks. However, their mastery of underlying inferential rules still falls short of human capabilities. To investigate this, we propose a logic scaffolding inferential rule generation framework, to construct an inferential rule base, ULogic, comprising both primitive and compositional rules across five domains. Our analysis of GPT-series models over a rule subset reveals significant gaps in LLMs' logic understanding compared to human performance, especially in compositional and structural complex rules with certain bias patterns. We further distill these rules into a smaller-scale inference engine for flexible rule generation and enhancing downstream reasoning. Through a multi-judger evaluation, our inference engine proves effective in generating accurate, complex and abstract conclusions and premises, and improve various commonsense reasoning tasks. Overall, our work sheds light on LLMs' limitations in grasping inferential rule and suggests ways to enhance their logical reasoning abilities~\footnote{Code and data are available at \url{https://github.com/SiyuanWangw/ULogic}.}.
Large Language Models (LLMs) attempt to imitate human behavior by responding to humans in a way that pleases them, including by adhering to their values. However, humans come from diverse cultures with different values. It is critical to understand whether LLMs showcase different values to the user based on the stereotypical values of a user's known country. We prompt different LLMs with a series of advice requests based on 5 Hofstede Cultural Dimensions -- a quantifiable way of representing the values of a country. Throughout each prompt, we incorporate personas representing 36 different countries and, separately, languages predominantly tied to each country to analyze the consistency in the LLMs' cultural understanding. Through our analysis of the responses, we found that LLMs can differentiate between one side of a value and another, as well as understand that countries have differing values, but will not always uphold the values when giving advice, and fail to understand the need to answer differently based on different cultural values. Rooted in these findings, we present recommendations for training value-aligned and culturally sensitive LLMs. More importantly, the methodology and the framework developed here can help further understand and mitigate culture and language alignment issues with LLMs.
This paper introduces FedSecurity, an end-to-end benchmark that serves as a supplementary component of the FedML library for simulating adversarial attacks and corresponding defense mechanisms in Federated Learning (FL). FedSecurity eliminates the need for implementing the fundamental FL procedures, e.g., FL training and data loading, from scratch, thus enables users to focus on developing their own attack and defense strategies. It contains two key components, including FedAttacker that conducts a variety of attacks during FL training, and FedDefender that implements defensive mechanisms to counteract these attacks. FedSecurity has the following features: i) It offers extensive customization options to accommodate a broad range of machine learning models (e.g., Logistic Regression, ResNet, and GAN) and FL optimizers (e.g., FedAVG, FedOPT, and FedNOVA); ii) it enables exploring the effectiveness of attacks and defenses across different datasets and models; and iii) it supports flexible configuration and customization through a configuration file and some APIs. We further demonstrate FedSecurity's utility and adaptability through federated training of Large Language Models (LLMs) to showcase its potential on a wide range of complex applications.
Although there is a growing belief that LLMs can be used as therapists, exploring LLMs' capabilities and inefficacy, particularly from the client's perspective, is limited. This work focuses on a client-centered assessment of LLM therapists with the involvement of simulated clients, a standard approach in clinical medical education. However, there are two challenges when applying the approach to assess LLM therapists at scale. Ethically, asking humans to frequently mimic clients and exposing them to potentially harmful LLM outputs can be risky and unsafe. Technically, it can be difficult to consistently compare the performances of different LLM therapists interacting with the same client. To this end, we adopt LLMs to simulate clients and propose ClientCAST, a client-centered approach to assessing LLM therapists by client simulation. Specifically, the simulated client is utilized to interact with LLM therapists and complete questionnaires related to the interaction. Based on the questionnaire results, we assess LLM therapists from three client-centered aspects: session outcome, therapeutic alliance, and self-reported feelings. We conduct experiments to examine the reliability of ClientCAST and use it to evaluate LLMs therapists implemented by Claude-3, GPT-3.5, LLaMA3-70B, and Mixtral 8*7B. Codes are released at https://github.com/wangjs9/ClientCAST.
This study introduces the "Grade Score", a novel metric designed to evaluate the consistency and fairness of Large Language Models (LLMs) when used as multiple-choice judges with respect to order bias and choice consistency. The Grade Score combines Entropy, which measures order bias, and Mode Frequency, which assesses choice stability, offering insights into LLMs' reliability and impartiality. The study explores techniques such as prompt engineering and option sampling strategies to optimize the Grade Score, demonstrating their effectiveness in enhancing LLMs' performance. Results showcase varying performances among LLMs with respect to prompts and highlight the positive impact of including irrelevant options. The study also identifies an emergent behavior in instruction-following models, where they adapt to instructions targeting specific biases, demonstrating their adaptability. The Grade Score facilitates comparisons between LLMs and encourages ongoing research towards optimizing their decision-making processes, with potential implications for improving their reliability and fairness in various applications. All code is available on GitHub https://github.com/IoDmitri/GradeLab
Online Large Language Model (LLM) services such as ChatGPT and Claude 3 have transformed business operations and academic research by effortlessly enabling new opportunities. However, due to data-sharing restrictions, sectors such as healthcare and finance prefer to deploy local LLM applications using costly hardware resources. This scenario requires a balance between the effectiveness advantages of LLMs and significant financial burdens. Additionally, the rapid evolution of models increases the frequency and redundancy of benchmarking efforts. Existing benchmarking toolkits, which typically focus on effectiveness, often overlook economic considerations, making their findings less applicable to practical scenarios. To address these challenges, we introduce CEBench, an open-source toolkit specifically designed for multi-objective benchmarking that focuses on the critical trade-offs between expenditure and effectiveness required for LLM deployments. CEBench allows for easy modifications through configuration files, enabling stakeholders to effectively assess and optimize these trade-offs. This strategic capability supports crucial decision-making processes aimed at maximizing effectiveness while minimizing cost impacts. By streamlining the evaluation process and emphasizing cost-effectiveness, CEBench seeks to facilitate the development of economically viable AI solutions across various industries and research fields. The code and demonstration are available in \url{https://github.com/amademicnoboday12/CEBench}.
Recent empirical results have sparked a debate about whether or not Large Language Models (LLMs) are capable of Theory of Mind (ToM). While some have found LLMs to be successful on ToM evaluations such as the False Belief task (Kosinski, 2023), others have argued that LLMs solve these tasks by exploiting spurious correlations -- not representing beliefs -- since they fail on trivial alterations to these tasks (Ullman, 2023). In this paper, we introduce SCALPEL: a technique to generate targeted modifications for False Belief tasks to test different specific hypotheses about why LLMs fail. We find that modifications which make explicit common inferences -- such as that looking at a transparent object implies recognizing its contents -- preserve LLMs' performance. This suggests that LLMs' failures on modified ToM tasks could result from a lack of more general commonsense reasoning, rather than a failure to represent mental states. We argue that SCALPEL could be helpful for explaining LLM successes and failures in other cases.
In this paper, we focus on addressing the constraints faced when applying LLMs to ASR. Recent works utilize prefixLM-type models, which directly apply speech as a prefix to LLMs for ASR. We have found that optimizing speech prefixes leads to better ASR performance and propose applying RNNT loss to perform speech prefix-tuning. This is a simple approach and does not increase the model complexity or alter the inference pipeline. We also propose language-based soft prompting to further improve with frozen LLMs. Empirical analysis on realtime testset from 10 Indic languages demonstrate that our proposed speech prefix-tuning yields improvements with both frozen and fine-tuned LLMs. Our recognition results on an average of 10 Indics show that the proposed prefix-tuning with RNNT loss results in a 12\% relative improvement in WER over the baseline with a fine-tuned LLM. Our proposed approches with the frozen LLM leads to a 31\% relative improvement over basic soft-prompting prefixLM.
In this work, we introduce a novel approach that equips LLM agents with introspection, enhancing consistency and adaptability in solving complex tasks. Our approach prompts LLM agents to decompose a given task into manageable subtasks (i.e., to make a plan), and to continuously introspect upon the suitability and results of their actions. %; and when necessary, to explore ``the road not taken.'' We implement a three-fold introspective intervention: 1) anticipatory reflection on potential failures and alternative remedy before action execution, 2) post-action alignment with subtask objectives and backtracking with remedy to ensure utmost effort in plan execution, and 3) comprehensive review upon plan completion for future strategy refinement. By deploying and experimenting with this methodology -- a zero-shot approach -- within WebArena for practical tasks in web environments, our agent demonstrates superior performance with a success rate of 23.5% over existing zero-shot methods by 3.5%. The experimental results suggest that our introspection-driven approach not only enhances the agent's ability to navigate unanticipated challenges through a robust mechanism of plan execution, but also improves efficiency by reducing the number of trials and plan revisions by 45% needed to achieve a task.
This research focuses on how Large Language Models (LLMs) can help with (path) planning for mobile embodied agents such as robots, in a human-in-the-loop and interactive manner. A novel framework named LLM A*, aims to leverage the commonsense of LLMs, and the utility-optimal A* is proposed to facilitate few-shot near-optimal path planning. Prompts are used for two main purposes: 1) to provide LLMs with essential information like environments, costs, heuristics, etc.; 2) to communicate human feedback on intermediate planning results to LLMs. This approach takes human feedback on board and renders the entire planning process transparent (akin to a `white box') to humans. Moreover, it facilitates code-free path planning, thereby fostering the accessibility and inclusiveness of artificial intelligence techniques to communities less proficient in coding. Comparative analysis against A* and RL demonstrates that LLM A* exhibits greater efficiency in terms of search space and achieves paths comparable to A* while outperforming RL. The interactive nature of LLM A* also makes it a promising tool for deployment in collaborative human-robot tasks. Codes and Supplemental Materials can be found at GitHub: https://github.com/speedhawk/LLM-A-.
Training on model-generated synthetic data is a promising approach for finetuning LLMs, but it remains unclear when it helps or hurts. In this paper, we investigate this question for math reasoning via an empirical study, followed by building a conceptual understanding of our observations. First, we find that while the typical approach of finetuning a model on synthetic correct or positive problem-solution pairs generated by capable models offers modest performance gains, sampling more correct solutions from the finetuned learner itself followed by subsequent fine-tuning on this self-generated data $\textbf{doubles}$ the efficiency of the same synthetic problems. At the same time, training on model-generated positives can amplify various spurious correlations, resulting in flat or even inverse scaling trends as the amount of data increases. Surprisingly, we find that several of these issues can be addressed if we also utilize negative responses, i.e., model-generated responses that are deemed incorrect by a final answer verifier. Crucially, these negatives must be constructed such that the training can appropriately recover the utility or advantage of each intermediate step in the negative response. With this per-step scheme, we are able to attain consistent gains over only positive data, attaining performance similar to amplifying the amount of synthetic data by $\mathbf{8 \times}$. We show that training on per-step negatives can help to unlearn spurious correlations in the positive data, and is equivalent to advantage-weighted reinforcement learning (RL), implying that it inherits robustness benefits of RL over imitating positive data alone.
We study the potential of using large language models (LLMs) as an interactive optimizer for solving maximization problems in a text space using natural language and numerical feedback. Inspired by the classical optimization literature, we classify the natural language feedback into directional and non-directional, where the former is a generalization of the first-order feedback to the natural language space. We find that LLMs are especially capable of optimization when they are provided with {directional feedback}. Based on this insight, we design a new LLM-based optimizer that synthesizes directional feedback from the historical optimization trace to achieve reliable improvement over iterations. Empirically, we show our LLM-based optimizer is more stable and efficient in solving optimization problems, from maximizing mathematical functions to optimizing prompts for writing poems, compared with existing techniques.
We conceptualize the process of understanding as information compression, and propose a method for ranking large language models (LLMs) based on lossless data compression. We demonstrate the equivalence of compression length under arithmetic coding with cumulative negative log probabilities when using a large language model as a prior, that is, the pre-training phase of the model is essentially the process of learning the optimal coding length. At the same time, the evaluation metric compression ratio can be obtained without actual compression, which greatly saves overhead. In this paper, we use five large language models as priors for compression, then compare their performance on challenging natural language processing tasks, including sentence completion, question answering, and coreference resolution. Experimental results show that compression ratio and model performance are positively correlated, so it can be used as a general metric to evaluate large language models.
Reranking is a critical component in recommender systems, playing an essential role in refining the output of recommendation algorithms. Traditional reranking models have focused predominantly on accuracy, but modern applications demand consideration of additional criteria such as diversity and fairness. Existing reranking approaches often fail to harmonize these diverse criteria effectively at the model level. Moreover, these models frequently encounter challenges with scalability and personalization due to their complexity and the varying significance of different reranking criteria in diverse scenarios. In response, we introduce a comprehensive reranking framework enhanced by LLM, designed to seamlessly integrate various reranking criteria while maintaining scalability and facilitating personalized recommendations. This framework employs a fully connected graph structure, allowing the LLM to simultaneously consider multiple aspects such as accuracy, diversity, and fairness through a coherent Chain-of-Thought (CoT) process. A customizable input mechanism is also integrated, enabling the tuning of the language model's focus to meet specific reranking needs. We validate our approach using three popular public datasets, where our framework demonstrates superior performance over existing state-of-the-art reranking models in balancing multiple criteria. The code for this implementation is publicly available.
Modular vision-language models (Vision-LLMs) align pretrained image encoders with (frozen) large language models (LLMs) and post-hoc condition LLMs to `understand' the image input. With the abundance of readily available high-quality English image-text data as well as strong monolingual English LLMs, the research focus has been on English-only Vision-LLMs. Multilingual vision-language models are still predominantly obtained via expensive end-to-end pretraining, resulting in comparatively smaller models, trained on limited multilingual image data supplemented with text-only multilingual corpora. We present mBLIP, the first Vision-LLM leveraging multilingual LLMs, which we obtain in a computationally efficient manner on consumer-level hardware. To this end, we \textit{re-align} an image encoder previously tuned to an English LLM to a new, multilingual LLM using only a few million multilingual training examples derived from a mix of vision-and-language tasks, which we obtain by machine-translating high-quality English data to 95 languages. On the IGLUE benchmark and XM3600, mBLIP yields results competitive with state-of-the-art models and it greatly outperforms strong English-only Vision-LLMs like Llava 1.5. We release our model, code, and train data at \url{https://github.com/gregor-ge/mBLIP}.
LLMs offer valuable capabilities, yet they can be utilized by malicious users to disseminate deceptive information and generate fake news. The growing prevalence of LLMs poses difficulties in crafting detection approaches that remain effective across various text domains. Additionally, the absence of precautionary measures for AI-generated news on online social platforms is concerning. Therefore, there is an urgent need to improve people's ability to differentiate between news articles written by humans and those produced by LLMs. By providing cues in human-written and LLM-generated news, we can help individuals increase their skepticism towards fake LLM-generated news. This paper aims to elucidate simple markers that help individuals distinguish between articles penned by humans and those created by LLMs. To achieve this, we initially collected a dataset comprising 39k news articles authored by humans or generated by four distinct LLMs with varying degrees of fake. We then devise a metric named Entropy-Shift Authorship Signature (ESAS) based on the information theory and entropy principles. The proposed ESAS ranks terms or entities, like POS tagging, within news articles based on their relevance in discerning article authorship. We demonstrate the effectiveness of our metric by showing the high accuracy attained by a basic method, i.e., TF-IDF combined with logistic regression classifier, using a small set of terms with the highest ESAS score. Consequently, we introduce and scrutinize these top ESAS-ranked terms to aid individuals in strengthening their skepticism towards LLM-generated fake news.
In the rapidly evolving Next-Generation Networking (NGN) era, the adoption of zero-trust architectures has become increasingly crucial to protect security. However, provisioning zero-trust services in NGNs poses significant challenges, primarily due to the environmental complexity and dynamics. Motivated by these challenges, this paper explores efficient zero-trust service provisioning using hierarchical micro-segmentations. Specifically, we model zero-trust networks via hierarchical graphs, thereby jointly considering the resource- and trust-level features to optimize service efficiency. We organize such zero-trust networks through micro-segmentations, which support granular zero-trust policies efficiently. To generate the optimal micro-segmentation, we present the Large Language Model-Enhanced Graph Diffusion (LEGD) algorithm, which leverages the diffusion process to realize a high-quality generation paradigm. Additionally, we utilize policy boosting and Large Language Models (LLM) to enable LEGD to optimize the generation policy and understand complicated graphical features. Moreover, realizing the unique trustworthiness updates or service upgrades in zero-trust NGN, we further present LEGD-Adaptive Maintenance (LEGD-AM), providing an adaptive way to perform task-oriented fine-tuning on LEGD. Extensive experiments demonstrate that the proposed LEGD achieves 90% higher efficiency in provisioning services compared with other baselines. Moreover, the LEGD-AM can reduce the service outage time by over 50%.
LLMs are increasingly powerful and widely used to assist users in a variety of tasks. This use risks the introduction of LLM biases to consequential decisions such as job hiring, human performance evaluation, and criminal sentencing. Bias in NLP systems along the lines of gender and ethnicity has been widely studied, especially for specific stereotypes (e.g., Asians are good at math). In this paper, we investigate bias along less-studied but still consequential, dimensions, such as age and beauty, measuring subtler correlated decisions that LLMs make between social groups and unrelated positive and negative attributes. We ask whether LLMs hold wide-reaching biases of positive or negative sentiment for specific social groups similar to the "what is beautiful is good" bias found in people in experimental psychology. We introduce a template-generated dataset of sentence completion tasks that asks the model to select the most appropriate attribute to complete an evaluative statement about a person described as a member of a specific social group. We also reverse the completion task to select the social group based on an attribute. We report the correlations that we find for 4 cutting-edge LLMs. This dataset can be used as a benchmark to evaluate progress in more generalized biases and the templating technique can be used to expand the benchmark with minimal additional human annotation.
Machine learning techniques applied to the problem of financial market forecasting struggle with dynamic regime switching, or underlying correlation and covariance shifts in true (hidden) market variables. Drawing inspiration from the success of reinforcement learning in robotics, particularly in agile locomotion adaptation of quadruped robots to unseen terrains, we introduce an innovative approach that leverages world knowledge of pretrained LLMs (aka. 'privileged information' in robotics) and dynamically adapts them using intrinsic, natural market rewards using LLM alignment technique we dub as "Reinforcement Learning from Market Feedback" (**RLMF**). Strong empirical results demonstrate the efficacy of our method in adapting to regime shifts in financial markets, a challenge that has long plagued predictive models in this domain. The proposed algorithmic framework outperforms best-performing SOTA LLM models on the existing (FLARE) benchmark stock-movement (SM) tasks by more than 15\% improved accuracy. On the recently proposed NIFTY SM task, our adaptive policy outperforms the SOTA best performing trillion parameter models like GPT-4. The paper details the dual-phase, teacher-student architecture and implementation of our model, the empirical results obtained, and an analysis of the role of language embeddings in terms of Information Gain.
Knowledge tagging for questions plays a crucial role in contemporary intelligent educational applications, including learning progress diagnosis, practice question recommendations, and course content organization. Traditionally, these annotations are always conducted by pedagogical experts, as the task requires not only a strong semantic understanding of both question stems and knowledge definitions but also deep insights into connecting question-solving logic with corresponding knowledge concepts. With the recent emergence of advanced text encoding algorithms, such as pre-trained language models, many researchers have developed automatic knowledge tagging systems based on calculating the semantic similarity between the knowledge and question embeddings. In this paper, we explore automating the task using Large Language Models (LLMs), in response to the inability of prior encoding-based methods to deal with the hard cases which involve strong domain knowledge and complicated concept definitions. By showing the strong performance of zero- and few-shot results over math questions knowledge tagging tasks, we demonstrate LLMs' great potential in conquering the challenges faced by prior methods. Furthermore, by proposing a reinforcement learning-based demonstration retriever, we successfully exploit the great potential of different-sized LLMs in achieving better performance results while keeping the in-context demonstration usage efficiency high.
Recently, large language models (LLMs) have shown surprising performance in task-specific workloads as well as general tasks with the given prompts. However, to achieve unprecedented performance, recent LLMs use billions to trillions of parameters, which hinder the wide adaptation of those models due to their extremely large compute and memory requirements. To resolve the issue, various model compression methods are being actively investigated. In this work, we propose SDQ (Sparse Decomposed Quantization) to exploit both structured sparsity and quantization to achieve both high compute and memory efficiency. From our evaluations, we observe that SDQ can achieve 4x effective compute throughput with <1% quality drop.
Large language models (LLMs) have shown an impressive ability to perform tasks believed to require thought processes. When the model does not document an explicit thought process, it becomes difficult to understand the processes occurring within its hidden layers and to determine if these processes can be referred to as reasoning. We introduce a novel and interpretable analysis of internal multi-hop reasoning processes in LLMs. We demonstrate that the prediction process for compositional reasoning questions can be modeled using a simple linear transformation between two semantic category spaces. We show that during inference, the middle layers of the network generate highly interpretable embeddings that represent a set of potential intermediate answers for the multi-hop question. We use statistical analyses to show that a corresponding subset of tokens is activated in the model's output, implying the existence of parallel reasoning paths. These observations hold true even when the model lacks the necessary knowledge to solve the task. Our findings can help uncover the strategies that LLMs use to solve reasoning tasks, offering insights into the types of thought processes that can emerge from artificial intelligence. Finally, we also discuss the implication of cognitive modeling of these results.
Judging the equivalence between two SQL queries is a fundamental problem with many practical applications in data management and SQL generation (i.e., evaluating the quality of generated SQL queries in text-to-SQL task). While the research community has reasoned about SQL equivalence for decades, it poses considerable difficulties and no complete solutions exist. Recently, Large Language Models (LLMs) have shown strong reasoning capability in conversation, question answering and solving mathematics challenges. In this paper, we study if LLMs can be used to determine the equivalence between SQL queries under two notions of SQL equivalence (semantic equivalence and relaxed equivalence). To assist LLMs in generating high quality responses, we present two prompting techniques: Miniature & Mull and Explain & Compare. The former technique is used to evaluate the semantic equivalence in which it asks LLMs to execute a query on a simple database instance and then explore if a counterexample exists by modifying the database. The latter technique is used to evaluate the relaxed equivalence in which it asks LLMs to explain the queries and then compare if they contain significant logical differences. Our experiments demonstrate using our techniques, LLMs is a promising tool to help data engineers in writing semantically equivalent SQL queries, however challenges still persist, and is a better metric for evaluating SQL generation than the popular execution accuracy.
Retrieval-augmented generation (RAG) has emerged as a promising solution to mitigate the limitations of large language models (LLMs), such as hallucinations and outdated information. However, it remains unclear how LLMs handle knowledge conflicts arising from different augmented retrieved passages, especially when these passages originate from the same source and have equal trustworthiness. In this work, we conduct a comprehensive evaluation of LLM-generated answers to questions that have varying answers based on contradictory passages from Wikipedia, a dataset widely regarded as a high-quality pre-training resource for most LLMs. Specifically, we introduce WikiContradict, a benchmark consisting of 253 high-quality, human-annotated instances designed to assess LLM performance when augmented with retrieved passages containing real-world knowledge conflicts. We benchmark a diverse range of both closed and open-source LLMs under different QA scenarios, including RAG with a single passage, and RAG with 2 contradictory passages. Through rigorous human evaluations on a subset of WikiContradict instances involving 5 LLMs and over 3,500 judgements, we shed light on the behaviour and limitations of these models. For instance, when provided with two passages containing contradictory facts, all models struggle to generate answers that accurately reflect the conflicting nature of the context, especially for implicit conflicts requiring reasoning. Since human evaluation is costly, we also introduce an automated model that estimates LLM performance using a strong open-source language model, achieving an F-score of 0.8. Using this automated metric, we evaluate more than 1,500 answers from seven LLMs across all WikiContradict instances. To facilitate future work, we release WikiContradict on: https://ibm.biz/wikicontradict.
Analogical reasoning is considered core to human learning and cognition. Recent studies have compared the analogical reasoning abilities of human subjects and Large Language Models (LLMs) on abstract symbol manipulation tasks, such as letter string analogies. However, these studies largely neglect analogical reasoning over semantically meaningful symbols, such as natural language words. This ability to draw analogies that link language to non-linguistic domains, which we term semantic structure-mapping, is thought to play a crucial role in language acquisition and broader cognitive development. We test human subjects and LLMs on analogical reasoning tasks that require the transfer of semantic structure and content from one domain to another. Advanced LLMs match human performance across many task variations. However, humans and LLMs respond differently to certain task variations and semantic distractors. Overall, our data suggest that LLMs are approaching human-level performance on these important cognitive tasks, but are not yet entirely human like.
How can we construct an automated debate judge to evaluate an extensive, vibrant, multi-turn debate? This task is challenging, as judging a debate involves grappling with lengthy texts, intricate argument relationships, and multi-dimensional assessments. At the same time, current research mainly focuses on short dialogues, rarely touching upon the evaluation of an entire debate. In this paper, by leveraging Large Language Models (LLMs), we propose Debatrix, which makes the analysis and assessment of multi-turn debates more aligned with majority preferences. Specifically, Debatrix features a vertical, iterative chronological analysis and a horizontal, multi-dimensional evaluation collaboration. To align with real-world debate scenarios, we introduced the PanelBench benchmark, comparing our system's performance to actual debate outcomes. The findings indicate a notable enhancement over directly using LLMs for debate evaluation. Source code and benchmark data are available online at https://github.com/ljcleo/debatrix .
Humans often express their communicative intents indirectly or non-literally, which requires their interlocutors -- human or AI -- to understand beyond the literal meaning of words. While most existing work has focused on discriminative evaluations, we present a new approach to generatively evaluate large language models' (LLMs') intention understanding by examining their responses to non-literal utterances. Ideally, an LLM should respond in line with the true intention of a non-literal utterance, not its literal interpretation. Our findings show that LLMs struggle to generate pragmatically relevant responses to non-literal language, achieving only 50-55% accuracy on average. While explicitly providing oracle intentions significantly improves performance (e.g., 75% for Mistral-Instruct), this still indicates challenges in leveraging given intentions to produce appropriate responses. Using chain-of-thought to make models spell out intentions yields much smaller gains (60% for Mistral-Instruct). These findings suggest that LLMs are not yet effective pragmatic interlocutors, highlighting the need for better approaches for modeling intentions and utilizing them for pragmatic generation.
Large Language Models (LLMs) have shown superior capability to solve reasoning problems with programs. While being a promising direction, most of such frameworks are trained and evaluated in settings with a prior knowledge of task requirements. However, as LLMs become more capable, it is necessary to assess their reasoning abilities in more realistic scenarios where many real-world problems are open-ended with ambiguous scope, and often require multiple formalisms to solve. To investigate this, we introduce the task of reasoning in the wild, where an LLM is tasked to solve a reasoning problem of unknown type by identifying the subproblems and their corresponding formalisms, and writing a program to solve each subproblem, guided by a tactic. We create a large tactic-guided trajectory dataset containing detailed solutions to a diverse set of reasoning problems, ranging from well-defined single-form reasoning (e.g., math, logic), to ambiguous and hybrid ones (e.g., commonsense, combined math and logic). This allows us to test various aspects of LLMs reasoning at the fine-grained level such as the selection and execution of tactics, and the tendency to take undesired shortcuts. In experiments, we highlight that existing LLMs fail significantly on problems with ambiguous and mixed scope, revealing critical limitations and overfitting issues (e.g. accuracy on GSM8K drops by at least 50\%). We further show the potential of finetuning a local LLM on the tactic-guided trajectories in achieving better performance. Project repo is available at github.com/gblackout/Reason-in-the-Wild
Recent advancements in Reinforcement Learning with Human Feedback (RLHF) have significantly impacted the alignment of Large Language Models (LLMs). The sensitivity of reinforcement learning algorithms such as Proximal Policy Optimization (PPO) has led to new line work on Direct Policy Optimization (DPO), which treats RLHF in a supervised learning framework. The increased practical use of these RLHF methods warrants an analysis of their vulnerabilities. In this work, we investigate the vulnerabilities of DPO to poisoning attacks under different scenarios and compare the effectiveness of preference poisoning, a first of its kind. We comprehensively analyze DPO's vulnerabilities under different types of attacks, i.e., backdoor and non-backdoor attacks, and different poisoning methods across a wide array of language models, i.e., LLama 7B, Mistral 7B, and Gemma 7B. We find that unlike PPO-based methods, which, when it comes to backdoor attacks, require at least 4\% of the data to be poisoned to elicit harmful behavior, we exploit the true vulnerabilities of DPO more simply so we can poison the model with only as much as 0.5\% of the data. We further investigate the potential reasons behind the vulnerability and how well this vulnerability translates into backdoor vs non-backdoor attacks.
Vulnerability detection is essential for software quality assurance. In recent years, deep learning models (especially large language models) have shown promise in vulnerability detection. In this work, we propose a novel LLM-based vulnerability detection technique Vul-RAG, which leverages knowledge-level retrieval-augmented generation (RAG) framework to detect vulnerability for the given code in three phases. First, Vul-RAG constructs a vulnerability knowledge base by extracting multi-dimension knowledge via LLMs from existing CVE instances; second, for a given code snippet, Vul-RAG} retrieves the relevant vulnerability knowledge from the constructed knowledge base based on functional semantics; third, Vul-RAG leverages LLMs to check the vulnerability of the given code snippet by reasoning the presence of vulnerability causes and fixing solutions of the retrieved vulnerability knowledge. Our evaluation of Vul-RAG on our constructed benchmark PairVul shows that Vul-RAG substantially outperforms all baselines by 12.96\%/110\% relative improvement in accuracy/pairwise-accuracy. In addition, our user study shows that the vulnerability knowledge generated by Vul-RAG can serve as high-quality explanations which can improve the manual detection accuracy from 0.60 to 0.77.
As the field of AI continues to evolve, a significant dimension of this progression is the development of Large Language Models and their potential to enhance multi-agent artificial intelligence systems. This paper explores the cooperative capabilities of Large Language Model-augmented Autonomous Agents (LAAs) using the well-known Meltin Pot environments along with reference models such as GPT4 and GPT3.5. Preliminary results suggest that while these agents demonstrate a propensity for cooperation, they still struggle with effective collaboration in given environments, emphasizing the need for more robust architectures. The study's contributions include an abstraction layer to adapt Melting Pot game scenarios for LLMs, the implementation of a reusable architecture for LLM-mediated agent development - which includes short and long-term memories and different cognitive modules, and the evaluation of cooperation capabilities using a set of metrics tied to the Melting Pot's "Commons Harvest" game. The paper closes, by discussing the limitations of the current architectural framework and the potential of a new set of modules that fosters better cooperation among LAAs.
Large language models (LLMs) have transformed the field of natural language processing, but they remain susceptible to jailbreaking attacks that exploit their capabilities to generate unintended and potentially harmful content. Existing token-level jailbreaking techniques, while effective, face scalability and efficiency challenges, especially as models undergo frequent updates and incorporate advanced defensive measures. In this paper, we introduce JailMine, an innovative token-level manipulation approach that addresses these limitations effectively. JailMine employs an automated "mining" process to elicit malicious responses from LLMs by strategically selecting affirmative outputs and iteratively reducing the likelihood of rejection. Through rigorous testing across multiple well-known LLMs and datasets, we demonstrate JailMine's effectiveness and efficiency, achieving a significant average reduction of 86% in time consumed while maintaining high success rates averaging 95%, even in the face of evolving defensive strategies. Our work contributes to the ongoing effort to assess and mitigate the vulnerability of LLMs to jailbreaking attacks, underscoring the importance of continued vigilance and proactive measures to enhance the security and reliability of these powerful language models.
Large language models (LLMs) iteratively generate text token by token, with memory usage increasing with the length of generated token sequences. The unpredictability of generation lengths makes it difficult to estimate the time and memory needed to process requests, posing a challenge for effective request scheduling. Conventional sequence-level scheduling (SLS) serves requests in a first-come first-served (FCFS) manner with static batching where requests with short generation lengths are delayed until those with long ones have finished generation, which hurts computational efficiency. Besides, to avoid out-of-memory (OOM) errors, SLS batches requests with a small batch size, which limits throughput. Recently proposed iteration-level scheduling (ILS) enhances computational efficiency with continuous batching to return completed requests timely and dynamically add new requests for processing. However, many ILS schedulers limit the number of parallel-processing requests to avoid OOM errors while achieving a fast inference speed, which compromises throughput. Moreover, existing SLS and ILS schedulers fail to balance the workload across multiple deployed LLM instances. To tackle these challenges, we propose slice-level scheduling (SCLS). By splitting the predefined maximal generation length limit into slices and serving batches slice by slice, it provides a precise range of serving time and memory usage for batched requests, laying the foundation for effective scheduling. Experiments confirm that compared with SLS and ILS schedulers, SCLS can improve throughput by up to 315.8% and greatly mitigate load imbalance with proposed batching and offloading algorithms.
Large Language Models (LLMs) tend to be unreliable in the factuality of their answers. To address this problem, NLP researchers have proposed a range of techniques to estimate LLM's confidence over facts. However, due to the lack of a systematic comparison, it is not clear how the different methods compare to one another. To fill this gap, we present a survey and empirical comparison of estimators of factual confidence. We define an experimental framework allowing for fair comparison, covering both fact-verification and question answering. Our experiments across a series of LLMs indicate that trained hidden-state probes provide the most reliable confidence estimates, albeit at the expense of requiring access to weights and training data. We also conduct a deeper assessment of factual confidence by measuring the consistency of model behavior under meaning-preserving variations in the input. We find that the confidence of LLMs is often unstable across semantically equivalent inputs, suggesting that there is much room for improvement of the stability of models' parametric knowledge. Our code is available at (https://github.com/amazon-science/factual-confidence-of-llms).
The Large Language Model (LLM) has gained significant popularity and is extensively utilized across various domains. Most LLM deployments occur within cloud data centers, where they encounter substantial response delays and incur high costs, thereby impacting the Quality of Services (QoS) at the network edge. Leveraging vector database caching to store LLM request results at the edge can substantially mitigate response delays and cost associated with similar requests, which has been overlooked by previous research. Addressing these gaps, this paper introduces a novel Vector database-assisted cloud-Edge collaborative LLM QoS Optimization (VELO) framework. Firstly, we propose the VELO framework, which ingeniously employs vector database to cache the results of some LLM requests at the edge to reduce the response time of subsequent similar requests. Diverging from direct optimization of the LLM, our VELO framework does not necessitate altering the internal structure of LLM and is broadly applicable to diverse LLMs. Subsequently, building upon the VELO framework, we formulate the QoS optimization problem as a Markov Decision Process (MDP) and devise an algorithm grounded in Multi-Agent Reinforcement Learning (MARL) to decide whether to request the LLM in the cloud or directly return the results from the vector database at the edge. Moreover, to enhance request feature extraction and expedite training, we refine the policy network of MARL and integrate expert demonstrations. Finally, we implement the proposed algorithm within a real edge system. Experimental findings confirm that our VELO framework substantially enhances user satisfaction by concurrently diminishing delay and resource consumption for edge users utilizing LLMs.
Knowledge Graph (KG) inductive reasoning, which aims to infer missing facts from new KGs that are not seen during training, has been widely adopted in various applications. One critical challenge of KG inductive reasoning is handling low-resource scenarios with scarcity in both textual and structural aspects. In this paper, we attempt to address this challenge with Large Language Models (LLMs). Particularly, we utilize the state-of-the-art LLMs to generate a graph-structural prompt to enhance the pre-trained Graph Neural Networks (GNNs), which brings us new methodological insights into the KG inductive reasoning methods, as well as high generalizability in practice. On the methodological side, we introduce a novel pretraining and prompting framework ProLINK, designed for low-resource inductive reasoning across arbitrary KGs without requiring additional training. On the practical side, we experimentally evaluate our approach on 36 low-resource KG datasets and find that ProLINK outperforms previous methods in three-shot, one-shot, and zero-shot reasoning tasks, exhibiting average performance improvements by 20%, 45%, and 147%, respectively. Furthermore, ProLINK demonstrates strong robustness for various LLM promptings as well as full-shot scenarios.
Large language models (LLMs) trained on datasets of publicly available source code have established a new state of the art in code generation tasks. However, these models are mostly unaware of the code that exists within a specific project, preventing the models from making good use of existing APIs. Instead, LLMs often invent, or "hallucinate", non-existent APIs or produce variants of already existing code. This paper presents De-Hallucinator, a technique that grounds the predictions of an LLM through a novel combination of retrieving suitable API references and iteratively querying the model with increasingly suitable context information in the prompt. The approach exploits the observation that predictions by LLMs often resemble the desired code, but they fail to correctly refer to already existing APIs. De-Hallucinator automatically identifies project-specific API references related to the model's initial predictions and adds these references into the prompt. Unlike retrieval-augmented generation (RAG), our approach uses the initial prediction(s) by the model to iteratively retrieve increasingly suitable API references. Our evaluation applies the approach to two tasks: predicting API usages in Python and generating tests in JavaScript. We show that De-Hallucinator consistently improves the generated code across five LLMs. In particular, the approach improves the edit distance by 23.3-50.6% and the recall of correctly predicted API usages by 23.9-61.0% for code completion, and improves the number of fixed tests that initially failed because of hallucinations by 63.2%, resulting in a 15.5% increase in statement coverage for test generation.
Multilingual Large Language Models (LLMs) achieve remarkable levels of zero-shot cross-lingual transfer performance. We speculate that this is predicated on their ability to align languages without explicit supervision from parallel sentences. While representations of translationally equivalent sentences in different languages are known to be similar after convergence, however, it remains unclear how such cross-lingual alignment emerges during pre-training of LLMs. Our study leverages intrinsic probing techniques, which identify which subsets of neurons encode linguistic features, to correlate the degree of cross-lingual neuron overlap with the zero-shot cross-lingual transfer performance for a given model. In particular, we rely on checkpoints of BLOOM, a multilingual autoregressive LLM, across different training steps and model scales. We observe a high correlation between neuron overlap and downstream performance, which supports our hypothesis on the conditions leading to effective cross-lingual transfer. Interestingly, we also detect a degradation of both implicit alignment and multilingual abilities in certain phases of the pre-training process, providing new insights into the multilingual pretraining dynamics.
Text spotting has seen tremendous progress in recent years yielding performant techniques which can extract text at the character, word or line level. However, extracting blocks of text from images (block-level text spotting) is relatively unexplored. Blocks contain more context than individual lines, words or characters and so block-level text spotting would enhance downstream applications, such as translation, which benefit from added context. We propose a novel method, BTS-LLM (Block-level Text Spotting with LLMs), to identify text at the block level. BTS-LLM has three parts: 1) detecting and recognizing text at the line level, 2) grouping lines into blocks and 3) finding the best order of lines within a block using a large language model (LLM). We aim to exploit the strong semantic knowledge in LLMs for accurate block-level text spotting. Consequently if the text spotted is semantically meaningful but has been corrupted during text recognition, the LLM is also able to rectify mistakes in the text and produce a reconstruction of it.
Considering the challenges faced by large language models (LLMs) in logical reasoning and planning, prior efforts have sought to augment LLMs with access to external solvers. While progress has been made on simple reasoning problems, solving classical constraint satisfaction problems, such as the Boolean Satisfiability Problem (SAT) and Graph Coloring Problem (GCP), remains difficult for off-the-shelf solvers due to their intricate expressions and exponential search spaces. In this paper, we propose a novel differential logic layer-aided language modeling (DiLA) approach, where logical constraints are integrated into the forward and backward passes of a network layer, to provide another option for LLM tool learning. In DiLA, LLM aims to transform the language description to logic constraints and identify initial solutions of the highest quality, while the differential logic layer focuses on iteratively refining the LLM-prompted solution. Leveraging the logic layer as a bridge, DiLA enhances the logical reasoning ability of LLMs on a range of reasoning problems encoded by Boolean variables, guaranteeing the efficiency and correctness of the solution process. We evaluate the performance of DiLA on two classic reasoning problems and empirically demonstrate its consistent outperformance against existing prompt-based and solver-aided approaches.
Natural language (NL) has long been the predominant format for human cognition and communication, and by extension, has been similarly pivotal in the development and application of Large Language Models (LLMs). Yet, besides NL, LLMs have seen various non-NL formats during pre-training, such as code and logical expression. NL's status as the optimal format for LLMs, particularly in single-LLM reasoning and multi-agent communication, has not been thoroughly examined. In this work, we challenge the default use of NL by exploring the utility of non-NL formats in these contexts. We show that allowing LLMs to autonomously select the most suitable format before reasoning or communicating leads to a 3.3 to 5.7\% improvement in reasoning efficiency for different LLMs, and up to a 72.7\% reduction in token usage in multi-agent communication, all while maintaining communicative effectiveness. Our comprehensive analysis further reveals that LLMs can devise a format from limited task instructions and that the devised format is effectively transferable across different LLMs. Intriguingly, the structured communication format decided by LLMs exhibits notable parallels with established agent communication languages, suggesting a natural evolution towards efficient, structured communication in agent communication. Our code is released at \url{https://github.com/thunlp/AutoForm}.
A dog whistle is a form of coded communication that carries a secondary meaning to specific audiences and is often weaponized for racial and socioeconomic discrimination. Dog whistling historically originated from United States politics, but in recent years has taken root in social media as a means of evading hate speech detection systems and maintaining plausible deniability. In this paper, we present an approach for word-sense disambiguation of dog whistles from standard speech using Large Language Models (LLMs), and leverage this technique to create a dataset of 16,550 high-confidence coded examples of dog whistles used in formal and informal communication. Silent Signals is the largest dataset of disambiguated dog whistle usage, created for applications in hate speech detection, neology, and political science. The dataset can be found at https://huggingface.co/datasets/SALT-NLP/silent_signals.
Accurate text summarization is one of the most common and important tasks performed by Large Language Models, where the costs of human review for an entire document may be high, but the costs of errors in summarization may be even greater. We propose Detecting Errors through Ensembling Prompts (DEEP) - an end-to-end large language model framework for detecting factual errors in text summarization. Our framework uses a diverse set of LLM prompts to identify factual inconsistencies, treating their outputs as binary features, which are then fed into ensembling models. We then calibrate the ensembled models to produce empirically accurate probabilities that a text is factually consistent or free of hallucination. We demonstrate that prior models for detecting factual errors in summaries perform significantly worse without optimizing the thresholds on subsets of the evaluated dataset. Our framework achieves state-of-the-art (SOTA) balanced accuracy on the AggreFact-XSUM FTSOTA, TofuEval Summary-Level, and HaluEval Summarization benchmarks in detecting factual errors within transformer-generated text summaries. It does so without any fine-tuning of the language model or reliance on thresholding techniques not available in practical settings.
Configuration settings are essential for tailoring software behavior to meet specific performance requirements. However, incorrect configurations are widespread, and identifying those that impact system performance is challenging due to the vast number and complexity of possible settings. In this work, we present PerfSense, a lightweight framework that leverages Large Language Models (LLMs) to efficiently identify performance-sensitive configurations with minimal overhead. PerfSense employs LLM agents to simulate interactions between developers and performance engineers using advanced prompting techniques such as prompt chaining and retrieval-augmented generation (RAG). Our evaluation of seven open-source Java systems demonstrates that PerfSense achieves an average accuracy of 64.77% in classifying performance-sensitive configurations, outperforming both our LLM baseline (50.36%) and the previous state-of-the-art method (61.75%). Notably, our prompt chaining technique improves recall by 10% to 30% while maintaining similar precision levels. Additionally, a manual analysis of 362 misclassifications reveals common issues, including LLMs' misunderstandings of requirements (26.8%). In summary, PerfSense significantly reduces manual effort in classifying performance-sensitive configurations and offers valuable insights for future LLM-based code analysis research.
NLP in mental health has been primarily social media focused. Real world practitioners also have high case loads and often domain specific variables, of which modern LLMs lack context. We take a dataset made by recruiting 644 participants, including individuals diagnosed with Bipolar Disorder (BD), Schizophrenia (SZ), and Healthy Controls (HC). Participants undertook tasks derived from a standardized mental health instrument, and the resulting data were transcribed and annotated by experts across five clinical variables. This paper demonstrates the application of contemporary language models in sequence-to-sequence tasks to enhance mental health research. Specifically, we illustrate how these models can facilitate the deployment of mental health instruments, data collection, and data annotation with high accuracy and scalability. We show that small models are capable of annotation for domain-specific clinical variables, data collection for mental-health instruments, and perform better then commercial large models.
Automatically interpreting CT scans can ease the workload of radiologists. However, this is challenging mainly due to the scarcity of adequate datasets and reference standards for evaluation. This study aims to bridge this gap by introducing a novel evaluation framework, named ``GPTRadScore''. This framework assesses the capabilities of multi-modal LLMs, such as GPT-4 with Vision (GPT-4V), Gemini Pro Vision, LLaVA-Med, and RadFM, in generating descriptions for prospectively-identified findings. By employing a decomposition technique based on GPT-4, GPTRadScore compares these generated descriptions with gold-standard report sentences, analyzing their accuracy in terms of body part, location, and type of finding. Evaluations demonstrated a high correlation with clinician assessments and highlighted its potential over traditional metrics, such as BLEU, METEOR, and ROUGE. Furthermore, to contribute to future studies, we plan to release a benchmark dataset annotated by clinicians. Using GPTRadScore, we found that while GPT-4V and Gemini Pro Vision fare better, their performance revealed significant areas for improvement, primarily due to limitations in the dataset used for training these models. To demonstrate this potential, RadFM was fine-tuned and it resulted in significant accuracy improvements: location accuracy rose from 3.41\% to 12.8\%, body part accuracy from 29.12\% to 53\%, and type accuracy from 9.24\% to 30\%, thereby validating our hypothesis.
As the demand for more personalized recommendation grows and a dramatic boom in commercial scenarios arises, the study on multi-scenario recommendation (MSR) has attracted much attention, which uses the data from all scenarios to simultaneously improve their recommendation performance. However, existing methods tend to integrate insufficient scenario knowledge and neglect learning personalized cross-scenario preferences, thus leading to suboptimal performance and inadequate interpretability. Meanwhile, though large language model (LLM) has shown great capability of reasoning and capturing semantic information, the high inference latency and high computation cost of tuning hinder its implementation in industrial recommender systems. To fill these gaps, we propose an effective efficient interpretable LLM-enhanced paradigm LLM4MSR in this work. Specifically, we first leverage LLM to uncover multi-level knowledge including scenario correlations and users' cross-scenario interests from the designed scenario- and user-level prompt without fine-tuning the LLM, then adopt hierarchical meta networks to generate multi-level meta layers to explicitly improves the scenario-aware and personalized recommendation capability. Our experiments on KuaiSAR-small, KuaiSAR, and Amazon datasets validate two significant advantages of LLM4MSR: (i) the effectiveness and compatibility with different multi-scenario backbone models (achieving 1.5%, 1%, and 40% AUC improvement on three datasets), (ii) high efficiency and deployability on industrial recommender systems, and (iii) improved interpretability. The implemented code and data is available to ease reproduction.
Large Language Models (LLMs) such as ChatGPT and GitHub Copilot have revolutionized automated code generation in software engineering. However, as these models are increasingly utilized for software development, concerns have arisen regarding the security and quality of the generated code. These concerns stem from LLMs being primarily trained on publicly available code repositories and internet-based textual data, which may contain insecure code. This presents a significant risk of perpetuating vulnerabilities in the generated code, creating potential attack vectors for exploitation by malicious actors. Our research aims to tackle these issues by introducing a framework for secure behavioral learning of LLMs through In-Content Learning (ICL) patterns during the code generation process, followed by rigorous security evaluations. To achieve this, we have selected four diverse LLMs for experimentation. We have evaluated these coding LLMs across three programming languages and identified security vulnerabilities and code smells. The code is generated through ICL with curated problem sets and undergoes rigorous security testing to evaluate the overall quality and trustworthiness of the generated code. Our research indicates that ICL-driven one-shot and few-shot learning patterns can enhance code security, reducing vulnerabilities in various programming scenarios. Developers and researchers should know that LLMs have a limited understanding of security principles. This may lead to security breaches when the generated code is deployed in production systems. Our research highlights LLMs are a potential source of new vulnerabilities to the software supply chain. It is important to consider this when using LLMs for code generation. This research article offers insights into improving LLM security and encourages proactive use of LLMs for code generation to ensure software system safety.
Stance detection holds great potential for enhancing the quality of online political discussions, as it has shown to be useful for summarizing discussions, detecting misinformation, and evaluating opinion distributions. Usually, transformer-based models are used directly for stance detection, which require large amounts of data. However, the broad range of debate questions in online political discussion creates a variety of possible scenarios that the model is faced with and thus makes data acquisition for model training difficult. In this work, we show how to leverage LLM-generated synthetic data to train and improve stance detection agents for online political discussions:(i) We generate synthetic data for specific debate questions by prompting a Mistral-7B model and show that fine-tuning with the generated synthetic data can substantially improve the performance of stance detection. (ii) We examine the impact of combining synthetic data with the most informative samples from an unlabelled dataset. First, we use the synthetic data to select the most informative samples, second, we combine both these samples and the synthetic data for fine-tuning. This approach reduces labelling effort and consistently surpasses the performance of the baseline model that is trained with fully labeled data. Overall, we show in comprehensive experiments that LLM-generated data greatly improves stance detection performance for online political discussions.
Recently, large language models (LLMs) have demonstrated their effectiveness in various natural language processing (NLP) tasks. However, the lack of tourism knowledge limits the performance of LLMs in tourist attraction presentations and travel planning. To address this challenge, we constructed a supervised fine-tuning dataset for the culture and tourism domain, named Cultour. This dataset consists of three parts: tourism knowledge base QA data, travelogues data, and tourism diversity QA data. Additionally, we propose TourLLM, a Qwen-based model supervised fine-tuned with Cultour, to improve the quality of the information provided about attractions and travel planning. To evaluate the performance of TourLLM, we employed both automatic and human evaluation, and we proposed a human evaluation criterion named CRA (Consistency, Readability, Availability). The experimental results demonstrate the effectiveness of the responses generated by the TourLLM. Our proposed Cultour is accessible at https://github.com/mrweiqk/Cultour.
In software engineering processes, systems are first specified using a modeling language such as UML. These initial designs are often collaboratively created, many times in meetings where different domain experts use whiteboards, paper or other types of quick supports to create drawings and blueprints that then will need to be formalized. These proper, machine-readable, models are key to ensure models can be part of automated processes (e.g. input of a low-code generation pipeline, a model-based testing system, ...). But going from hand-drawn diagrams to actual models is a time-consuming process that sometimes ends up with such drawings just added as informal images to the software documentation, reducing their value a lot. To avoid this tedious task, we explore the usage of Large Language Models (LLM) to generate the formal representation of (UML) models from a given drawing. More specifically, we have evaluated the capabilities of different LLMs to convert images of UML class diagrams into the actual models represented in the images. While the results are good enough to use such an approach as part of a model-driven engineering pipeline we also highlight some of their current limitations and the need to keep the human in the loop to overcome those limitations.
Large Language Models (LLMs) have achieved remarkable success across various tasks, yet their ability to learn incrementally without forgetting remains underexplored. Incremental learning (IL) is crucial as it enables models to acquire new knowledge while retaining previously learned information, akin to human learning. Existing benchmarks for IL are insufficient due to data leakage issues and the overqualification of LLMs. To address these challenges, we introduce Concept-1K, a novel dataset comprising 1,023 recently emerged concepts across diverse domains. The concepts in Concept-1K are discrete, interpretable units of knowledge that allow for fine-grained analysis of learning and forgetting processes. Using Concept-1K as a testbed, we aim to answer the question: ``Can LLMs learn new concepts incrementally without forgetting like humans?'' Our investigation reveals that LLMs still suffer from catastrophic forgetting and that LoRA, despite fine-tuning fewer parameters, may lead to more forgetting on training data. Additionally, we explore the roles of in-context learning, model scale, buffer size, and pretraining in IL performance. These findings highlight the strengths and limitations of LLMs in IL scenarios and provide a robust benchmark for future research.
We present CodeNav, an LLM agent that navigates and leverages previously unseen code repositories to solve user queries. In contrast to tool-use LLM agents that require ``registration'' of all relevant tools via manual descriptions within the LLM context, CodeNav automatically indexes and searches over code blocks in the target codebase, finds relevant code snippets, imports them, and uses them to iteratively generate a solution with execution feedback. To highlight the core-capabilities of CodeNav, we first showcase three case studies where we use CodeNav for solving complex user queries using three diverse codebases. Next, on three benchmarks, we quantitatively compare the effectiveness of code-use (which only has access to the target codebase) to tool-use (which has privileged access to all tool names and descriptions). Finally, we study the effect of varying kinds of tool and library descriptions on code-use performance, as well as investigate the advantage of the agent seeing source code as opposed to natural descriptions of code. All code will be made open source under a permissive license.
Recent studies show that large language models (LLMs) improve their performance through self-feedback on certain tasks while degrade on others. We discovered that such a contrary is due to LLM's bias in evaluating their own output. In this paper, we formally define LLM's self-bias - the tendency to favor its own generation - using two statistics. We analyze six LLMs (GPT-4, GPT-3.5, Gemini, LLaMA2, Mixtral and DeepSeek) on translation, constrained text generation, and mathematical reasoning tasks. We find that self-bias is prevalent in all examined LLMs across multiple languages and tasks. Our analysis reveals that while the self-refine pipeline improves the fluency and understandability of model outputs, it further amplifies self-bias. To mitigate such biases, we discover that larger model size and external feedback with accurate assessment can significantly reduce bias in the self-refine pipeline, leading to actual performance improvement in downstream tasks. The code and data are released at https://github.com/xu1998hz/llm_self_bias.
Large Language Models (LLMs) have achieved remarkable progress in language understanding and generation. Custom LLMs leveraging textual features have been applied to recommendation systems, demonstrating improvements across various recommendation scenarios. However, most existing methods perform untrained recommendation based on pre-trained knowledge (e.g., movie recommendation), and the auto-regressive generation of LLMs leads to slow inference speeds, making them less effective in real-time recommendations.To address this, we propose a framework for news recommendation using LLMs, named \textit{CherryRec}, which ensures the quality of recommendations while accelerating the recommendation process. Specifically, we employ a Knowledge-aware News Rapid Selector to retrieve candidate options based on the user's interaction history. The history and retrieved items are then input as text into a fine-tuned LLM, the Content-aware News Llm Evaluator, designed to enhance news recommendation capabilities. Finally, the Value-aware News Scorer integrates the scores to compute the CherryRec Score, which serves as the basis for the final recommendation.We validate the effectiveness of the proposed framework by comparing it with state-of-the-art baseline methods on benchmark datasets. Our experimental results consistently show that CherryRec outperforms the baselines in both recommendation performance and efficiency.The project resource can be accessed at: \url{https://github.com/xxxxxx}
This paper introduces a novel privacy-preservation framework named PFID for LLMs that addresses critical privacy concerns by localizing user data through model sharding and singular value decomposition. When users are interacting with LLM systems, their prompts could be subject to being exposed to eavesdroppers within or outside LLM system providers who are interested in collecting users' input. In this work, we proposed a framework to camouflage user input, so as to alleviate privacy issues. Our framework proposes to place model shards on the client and the public server, we sent compressed hidden states instead of prompts to and from servers. Clients have held back information that can re-privatized the hidden states so that overall system performance is comparable to traditional LLMs services. Our framework was designed to be communication efficient, computation can be delegated to the local client so that the server's computation burden can be lightened. We conduct extensive experiments on machine translation tasks to verify our framework's performance.
While Large language models (LLMs) have demonstrated considerable capabilities across various natural language tasks, they often fall short of the performance achieved by domain-specific state-of-the-art models. One potential approach to enhance domain-specific capabilities of LLMs involves fine-tuning them using corresponding datasets. However, this method can be both resource and time-intensive, and not applicable to closed-source commercial LLMs. In this paper, we propose Preference Adaptation for Enhancing Domain-specific Abilities of LLMs (PANDA), a method designed to augment the domain-specific capabilities of LLMs by leveraging insights from the response preference of expert models without requiring fine-tuning. Our experimental results reveal that PANDA significantly enhances the domain-specific ability of LLMs on text classification and interactive decision tasks. Moreover, LLM with PANDA even outperforms the expert model that being learned on 4 tasks of ScienceWorld. This finding highlights the potential of exploring tuning-free approaches to achieve weak-to-strong generalization.
Large Language Models (LLMs) and Multi-Modal LLMs (MLLMs) have played a critical role in numerous applications. However, current LLMs are vulnerable to prompt-based attacks, with jailbreaking attacks enabling LLMs to generate harmful content, while hijacking attacks manipulate the model to perform unintended tasks, underscoring the necessity for detection methods. Unfortunately, existing detecting approaches are usually tailored to specific attacks, resulting in poor generalization in detecting various attacks across different modalities. To address it, we propose JailGuard, a universal detection framework for jailbreaking and hijacking attacks across LLMs and MLLMs. JailGuard operates on the principle that attacks are inherently less robust than benign ones, regardless of method or modality. Specifically, JailGuard mutates untrusted inputs to generate variants and leverages the discrepancy of the variants' responses on the model to distinguish attack samples from benign samples. We implement 18 mutators for text and image inputs and design a mutator combination policy to further improve detection generalization. To evaluate the effectiveness of JailGuard, we build the first comprehensive multi-modal attack dataset, containing 11,000 data items across 15 known attack types. The evaluation suggests that JailGuard achieves the best detection accuracy of 86.14%/82.90% on text and image inputs, outperforming state-of-the-art methods by 11.81%-25.73% and 12.20%-21.40%.
Existing operations research (OR) models and tools play indispensable roles in smart-city operations, yet their practical implementation is limited by the complexity of modeling and deficiencies in optimization proficiency. To generate more relevant and accurate solutions to users' requirements, we propose a large language model (LLM)-based agent ("City-LEO") that enhances the efficiency and transparency of city management through conversational interactions. Specifically, to accommodate diverse users' requirements and enhance computational tractability, City-LEO leverages LLM's logical reasoning capabilities on prior knowledge to scope down large-scale optimization problems efficiently. In the human-like decision process, City-LEO also incorporates End-to-end (E2E) model to synergize the prediction and optimization. The E2E framework be conducive to coping with environmental uncertainties and involving more query-relevant features, and then facilitates transparent and interpretable decision-making process. In case study, we employ City-LEO in the operations management of e-bike sharing (EBS) system. The numerical results demonstrate that City-LEO has superior performance when benchmarks against the full-scale optimization problem. With less computational time, City-LEO generates more satisfactory and relevant solutions to the users' requirements, and achieves lower global suboptimality without significantly compromising accuracy. In a broader sense, our proposed agent offers promise to develop LLM-embedded OR tools for smart-city operations management.
Chain-of-thought (CoT), tree-of-thought (ToT), and related techniques work surprisingly well in practice for some complex reasoning tasks with Large Language Models (LLMs), but why? This work seeks the underlying reasons by conducting experimental case studies and linking the performance benefits to well-established sample and computational complexity principles in machine learning. We experimented with 6 reasoning tasks, ranging from grade school math, air travel planning, ..., to Blocksworld. The results suggest that (i) both CoT and ToT benefit significantly from task decomposition, which breaks a complex reasoning task into a sequence of steps with low sample complexity and explicitly outlines the reasoning structure, and (ii) for computationally hard reasoning tasks, the more sophisticated tree structure of ToT outperforms the linear structure of CoT. These findings provide useful guidelines for the use of LLM in solving reasoning tasks in practice.
Recently, both closed-source LLMs and open-source communities have made significant strides, outperforming humans in various general domains. However, their performance in specific professional fields such as medicine, especially within the open-source community, remains suboptimal due to the complexity of medical knowledge. We propose Aquila-Med, a bilingual medical LLM based on Aquila, addressing these challenges through continue pre-training, supervised fine-tuning (SFT), and reinforcement learning from human feedback (RLHF). We construct a large-scale Chinese and English medical dataset for continue pre-training and a high-quality SFT dataset, covering extensive medical specialties. Additionally, we develop a high-quality Direct Preference Optimization (DPO) dataset for further alignment. Aquila-Med achieves notable results across single-turn, multi-turn dialogues, and medical multiple-choice questions, demonstrating the effectiveness of our approach. We open-source the datasets and the entire training process, contributing valuable resources to the research community. Our models and datasets will released at https://huggingface.co/BAAI/AquilaMed-RL.
Recently, Large Language Models (LLMs) attained impressive performance in math and reasoning benchmarks. However, they still often struggle with logic problems and puzzles that are relatively easy for humans. To further investigate this, we introduce a new benchmark, SearchBench, containing 11 unique search problem types, each equipped with automated pipelines to generate an arbitrary number of instances and analyze the feasibility, correctness, and optimality of LLM-generated solutions. We show that even the most advanced LLMs fail to solve these problems end-to-end in text, e.g. GPT4 solves only 1.4%. SearchBench problems require considering multiple pathways to the solution as well as backtracking, posing a significant challenge to auto-regressive models. Instructing LLMs to generate code that solves the problem helps, but only slightly, e.g., GPT4's performance rises to 11.7%. In this work, we show that in-context learning with A* algorithm implementations enhances performance. The full potential of this promoting approach emerges when combined with our proposed Multi-Stage-Multi-Try method, which breaks down the algorithm implementation into two stages and verifies the first stage against unit tests, raising GPT-4's performance above 57%.
Intermediate Distillation: Data-Efficient Distillation from Black-Box LLMs for Information Retrieval
Recent research has explored distilling knowledge from large language models (LLMs) to optimize retriever models, especially within the retrieval-augmented generation (RAG) framework. However, most existing training methods rely on extracting supervision signals from LLMs' weights or their output probabilities, which is not only resource-intensive but also incompatible with black-box LLMs. In this paper, we introduce \textit{Intermediate Distillation}, a data-efficient knowledge distillation training scheme that treats LLMs as black boxes and distills their knowledge via an innovative LLM-ranker-retriever pipeline, solely using LLMs' ranking generation as the supervision signal. Extensive experiments demonstrate that our proposed method can significantly improve the performance of retriever models with only 1,000 training instances. Moreover, our distilled retriever model significantly boosts performance in question-answering tasks within the RAG framework, demonstrating the potential of LLMs to economically and effectively train smaller models.
Recent work shows that causal facts can be effectively extracted from LLMs through prompting, facilitating the creation of causal graphs for causal inference tasks. However, it is unclear if this success is limited to explicitly-mentioned causal facts in the pretraining data which the model can memorize. Thus, this work investigates: Can LLMs infer causal relations from other relational data in text? To disentangle the role of memorized causal facts vs inferred causal relations, we finetune LLMs on synthetic data containing temporal, spatial and counterfactual relations, and measure whether the LLM can then infer causal relations. We find that: (a) LLMs are susceptible to inferring causal relations from the order of two entity mentions in text (e.g. X mentioned before Y implies X causes Y); (b) if the order is randomized, LLMs still suffer from the post hoc fallacy, i.e. X occurs before Y (temporal relation) implies X causes Y. We also find that while LLMs can correctly deduce the absence of causal relations from temporal and spatial relations, they have difficulty inferring causal relations from counterfactuals, questioning their understanding of causality.
In the context of text classification, the financial burden of annotation exercises for creating training data is a critical issue. Active learning techniques, particularly those rooted in uncertainty sampling, offer a cost-effective solution by pinpointing the most instructive samples for manual annotation. Similarly, Large Language Models (LLMs) such as GPT-3.5 provide an alternative for automated annotation but come with concerns regarding their reliability. This study introduces a novel methodology that integrates human annotators and LLMs within an Active Learning framework. We conducted evaluations on three public datasets. IMDB for sentiment analysis, a Fake News dataset for authenticity discernment, and a Movie Genres dataset for multi-label classification.The proposed framework integrates human annotation with the output of LLMs, depending on the model uncertainty levels. This strategy achieves an optimal balance between cost efficiency and classification performance. The empirical results show a substantial decrease in the costs associated with data annotation while either maintaining or improving model accuracy.
This study empirically tests the $\textit{Narrative Economics}$ hypothesis, which posits that narratives (ideas that are spread virally and affect public beliefs) can influence economic fluctuations. We introduce two curated datasets containing posts from X (formerly Twitter) which capture economy-related narratives (Data will be shared upon paper acceptance). Employing Natural Language Processing (NLP) methods, we extract and summarize narratives from the tweets. We test their predictive power for $\textit{macroeconomic}$ forecasting by incorporating the tweets' or the extracted narratives' representations in downstream financial prediction tasks. Our work highlights the challenges in improving macroeconomic models with narrative data, paving the way for the research community to realistically address this important challenge. From a scientific perspective, our investigation offers valuable insights and NLP tools for narrative extraction and summarization using Large Language Models (LLMs), contributing to future research on the role of narratives in economics.
Each LLM serving request goes through two phases. The first is prefill which processes the entire input prompt and produces the first output token and the second is decode which generates the rest of output tokens, one-at-a-time. Prefill iterations have high latency but saturate GPU compute due to parallel processing of the input prompt. In contrast, decode iterations have low latency but also low compute utilization because a decode iteration processes only a single token per request. This makes batching highly effective for decodes and consequently for overall throughput. However, batching multiple requests leads to an interleaving of prefill and decode iterations which makes it challenging to achieve both high throughput and low latency. We introduce an efficient LLM inference scheduler, Sarathi-Serve, to address this throughput-latency tradeoff. Sarathi-Serve introduces chunked-prefills which splits a prefill request into near equal sized chunks and creates stall-free schedules that adds new requests in a batch without pausing ongoing decodes. Stall-free scheduling unlocks the opportunity to improve throughput with large batch sizes while minimizing the effect of batching on latency. Furthermore, uniform batches in Sarathi-Serve ameliorate the imbalance between iterations resulting in minimal pipeline bubbles. Our techniques yield significant improvements in inference performance across models and hardware under tail latency constraints. For Mistral-7B on single A100 GPUs, we achieve 2.6x higher serving capacity and up to 3.7x higher serving capacity for the Yi-34B model on two A100 GPUs as compared to vLLM. When used with pipeline parallelism on Falcon-180B, Sarathi-Serve provides up to 5.6x gain in the end-to-end serving capacity. The source code for Sarathi-Serve is available at https://github.com/microsoft/sarathi-serve.
Despite their vast capabilities, Large Language Models (LLMs) often struggle with generating reliable outputs, frequently producing high-confidence inaccuracies known as hallucinations. Addressing this challenge, our research introduces InternalInspector, a novel framework designed to enhance confidence estimation in LLMs by leveraging contrastive learning on internal states including attention states, feed-forward states, and activation states of all layers. Unlike existing methods that primarily focus on the final activation state, InternalInspector conducts a comprehensive analysis across all internal states of every layer to accurately identify both correct and incorrect prediction processes. By benchmarking InternalInspector against existing confidence estimation methods across various natural language understanding and generation tasks, including factual question answering, commonsense reasoning, and reading comprehension, InternalInspector achieves significantly higher accuracy in aligning the estimated confidence scores with the correctness of the LLM's predictions and lower calibration error. Furthermore, InternalInspector excels at HaluEval, a hallucination detection benchmark, outperforming other internal-based confidence estimation methods in this task.
With the proliferation of red-teaming strategies for Large Language Models (LLMs), the deficiency in the literature about improving the safety and robustness of LLM defense strategies is becoming increasingly pronounced. This paper introduces the LLM-based \textbf{sentinel} model as a plug-and-play prefix module designed to reconstruct the input prompt with just a few ($<30$) additional tokens, effectively reducing toxicity in responses from target LLMs. The sentinel model naturally overcomes the \textit{parameter inefficiency} and \textit{limited model accessibility} for fine-tuning large target models. We employ an interleaved training regimen using Proximal Policy Optimization (PPO) to optimize both red team and sentinel models dynamically, incorporating a value head-sharing mechanism inspired by the multi-agent centralized critic to manage the complex interplay between agents. Our extensive experiments across text-to-text and text-to-image demonstrate the effectiveness of our approach in mitigating toxic outputs, even when dealing with larger models like \texttt{Llama-2}, \texttt{GPT-3.5} and \texttt{Stable-Diffusion}, highlighting the potential of our framework in enhancing safety and robustness in various applications.
Serving systems for Large Language Models (LLMs) are often optimized to improve quality of service (QoS) and throughput. However, due to the lack of open-sourced LLM serving workloads, these systems are frequently evaluated under unrealistic workload assumptions. Consequently, performance may degrade when these systems are deployed in real-world scenarios. This work presents BurstGPT, an LLM serving workload with 5.29 million traces from regional Azure OpenAI GPT services over 121 days. BurstGPT captures realistic LLM serving characteristics through detailed tracing of: (1) Concurrency of requests: It traces burstiness variations of requests in Azure OpenAI GPT services, revealing diversified concurrency patterns in different services and model types. (2) Response Lengths of requests: It traces the auto-regressive serving processes of GPT models, showing statistical relations between requests and their responses. (3) Failures of requests: It traces failures of conversation and API services, showing intensive resource needs and limited resource availability of such services in Azure. Details of the characteristics can serve multiple purposes in LLM serving optimizations, such as system evaluation and trace provisioning. In our demo evaluation with BurstGPT, we observe that frequent variations in BurstGPT reveal declines in efficiency, stability, or reliability in realistic LLM serving. We identify that the generalization of KV cache management and request scheduling optimization is not guaranteed for different workloads, especially when systems are poorly optimized for unrealistic workloads. We have made the dataset publicly available to encourage further research at https://github.com/HPMLL/BurstGPT.
Recent advancements in language technology and Artificial Intelligence have resulted in numerous Language Models being proposed to perform various tasks in the legal domain ranging from predicting judgments to generating summaries. Despite their immense potential, these models have been proven to learn and exhibit societal biases and make unfair predictions. In this study, we explore the ability of Large Language Models (LLMs) to perform legal tasks in the Indian landscape when social factors are involved. We present a novel metric, $\beta$-weighted $\textit{Legal Safety Score ($LSS_{\beta}$)}$, which encapsulates both the fairness and accuracy aspects of the LLM. We assess LLMs' safety by considering its performance in the $\textit{Binary Statutory Reasoning}$ task and its fairness exhibition with respect to various axes of disparities in the Indian society. Task performance and fairness scores of LLaMA and LLaMA--2 models indicate that the proposed $LSS_{\beta}$ metric can effectively determine the readiness of a model for safe usage in the legal sector. We also propose finetuning pipelines, utilising specialised legal datasets, as a potential method to mitigate bias and improve model safety. The finetuning procedures on LLaMA and LLaMA--2 models increase the $LSS_{\beta}$, improving their usability in the Indian legal domain. Our code is publicly released.
Large language models (LLMs) have shown to pose social and ethical risks such as generating toxic language or facilitating malicious use of hazardous knowledge. Machine unlearning is a promising approach to improve LLM safety by directly removing harmful behaviors and knowledge. In this paper, we propose "SPlit, UNlearn, MerGE" (SPUNGE), a framework that can be used with any unlearning method to amplify its effectiveness. SPUNGE leverages data attributes during unlearning by splitting unlearning data into subsets based on specific attribute values, unlearning each subset separately, and merging the unlearned models. We empirically demonstrate that SPUNGE significantly improves the performance of two recent unlearning methods on state-of-the-art LLMs while maintaining their general capabilities on standard academic benchmarks.
Abstraction ability is crucial in human intelligence, which can also benefit various tasks in NLP study. Existing work shows that LLMs are deficient in abstract ability, and how to improve it remains unexplored. In this work, we design the framework AbsInstruct to enhance LLMs' abstraction ability through instruction tuning. The framework builds instructions with in-depth explanations to assist LLMs in capturing the underlying rationale of abstraction. Meanwhile, we introduce a plausibility estimator to select instructions that are more consistent with the abstraction knowledge of LLMs to be aligned. Then, our framework combines abstraction instructions with general-purpose ones to build a hybrid dataset. Extensive experiments and analyses demonstrate that our framework can considerably enhance LLMs' abstraction ability with strong generalization performance while maintaining their general instruction-following abilities.
The pervasive use of textual formats in the documentation of software requirements presents a great opportunity for applying large language models (LLMs) to software engineering tasks. High-quality software requirements not only enhance the manual software development process but also position organizations to fully harness the potential of the emerging LLMs technology. This paper introduces a tailored LLM for automating the generation of code snippets from well-structured requirements documents. This LLM is augmented with knowledge, heuristics, and instructions that are pertinent to the software development process, requirements analysis, object-oriented design, and test-driven development, effectively emulating the expertise of a seasoned software engineer. We introduce a "Progressive Prompting" method that allows software engineers to engage with this LLM in a stepwise manner. Through this approach, the LLM incrementally tackles software development tasks by interpreting the provided requirements to extract functional requirements, using these to create object-oriented models, and subsequently generating unit tests and code based on the object-oriented designs. We demonstrate the LLM's proficiency in comprehending intricate user requirements and producing robust design and code solutions through a case study focused on the development of a web project. This study underscores the potential of integrating LLMs into the software development workflow to significantly enhance both efficiency and quality. The tailored LLM is available at https://chat.openai.com/g/g-bahoiKzkB-software-engineer-gpt.
The increasing size of large language models (LLMs) challenges their usage on resource-constrained platforms. For example, memory on modern GPUs is insufficient to hold LLMs that are hundreds of Gigabytes in size. Offloading is a popular method to escape this constraint by storing weights of an LLM model to host CPU memory and SSD, then loading each weight to GPU before every use. In our case study of offloaded inference, we found that due to the low bandwidth between storage devices and GPU, the latency of transferring large model weights from its offloaded location to GPU memory becomes the critical bottleneck with actual compute taking nearly 0% of runtime. To effectively reduce the weight transfer latency, we propose a novel sparse format that compresses the unstructured sparse pattern of pruned LLM weights to non-zero values with high compression ratio and low decompression overhead. Endor achieves this by expressing the positions of non-zero elements with a bitmap. Compared to offloaded inference using the popular Huggingface Accelerate, applying Endor accelerates OPT-66B by 1.70x and Llama2-70B by 1.78x. When direct weight transfer from SSD to GPU is leveraged, Endor achieves 2.25x speedup on OPT-66B and 2.37x speedup on Llama2-70B.
This article gives an overview of the field of LLM text recognition. Different approaches and implemented detectors for the recognition of LLM-generated text are presented. In addition to discussing the implementations, the article focuses on benchmarking the detectors. Although there are numerous software products for the recognition of LLM-generated text, with a focus on ChatGPT-like LLMs, the quality of the recognition (recognition rate) is not clear. Furthermore, while it can be seen that scientific contributions presenting their novel approaches strive for some kind of comparison with other approaches, the construction and independence of the evaluation dataset is often not comprehensible. As a result, discrepancies in the performance evaluation of LLM detectors are often visible due to the different benchmarking datasets. This article describes the creation of an evaluation dataset and uses this dataset to investigate the different detectors. The selected detectors are benchmarked against each other.
Ensuring that large language models (LLMs) reflect diverse user values and preferences is crucial as their user bases expand globally. It is therefore encouraging to see the growing interest in LLM personalization within the research community. However, current works often rely on the LLM-as-a-Judge approach for evaluation without thoroughly examining its validity. In this paper, we investigate the reliability of LLM-as-a-Personalized-Judge, asking LLMs to judge user preferences based on personas. Our findings suggest that directly applying LLM-as-a-Personalized-Judge is less reliable than previously assumed, showing low and inconsistent agreement with human ground truth. The personas typically used are often overly simplistic, resulting in low predictive power. To address these issues, we introduce verbal uncertainty estimation into the LLM-as-a-Personalized-Judge pipeline, allowing the model to express low confidence on uncertain judgments. This adjustment leads to much higher agreement (above 80%) on high-certainty samples for binary tasks. Through human evaluation, we find that the LLM-as-a-Personalized-Judge achieves comparable performance to third-party humans evaluation and even surpasses human performance on high-certainty samples. Our work indicates that certainty-enhanced LLM-as-a-Personalized-Judge offers a promising direction for developing more reliable and scalable methods for evaluating LLM personalization.
The growing popularity of Large Language Models has sparked interest in context compression for Large Language Models (LLMs). However, the performance of previous methods degrades dramatically as compression ratios increase, sometimes even falling to the closed-book level. This decline can be attributed to the loss of key information during the compression process. Our preliminary study supports this hypothesis, emphasizing the significance of retaining key information to maintain model performance under high compression ratios. As a result, we introduce Query-Guided Compressor (QGC), which leverages queries to guide the context compression process, effectively preserving key information within the compressed context. Additionally, we employ a dynamic compression strategy. We validate the effectiveness of our proposed QGC on the Question Answering task, including NaturalQuestions, TriviaQA, and HotpotQA datasets. Experimental results show that QGC can consistently perform well even at high compression ratios, which also offers significant benefits in terms of inference cost and throughput.
SemEval-2024 Task 8 is focused on multigenerator, multidomain, and multilingual black-box machine-generated text detection. Such a detection is important for preventing a potential misuse of large language models (LLMs), the newest of which are very capable in generating multilingual human-like texts. We have coped with this task in multiple ways, utilizing language identification and parameter-efficient fine-tuning of smaller LLMs for text classification. We have further used the per-language classification-threshold calibration to uniquely combine fine-tuned models predictions with statistical detection metrics to improve generalization of the system detection performance. Our submitted method achieved competitive results, ranking at the fourth place, just under 1 percentage point behind the winner.
Large Language Models trained on code corpora (code-LLMs) have demonstrated impressive performance in various coding assistance tasks. However, despite their increased size and training dataset, code-LLMs still have limitations such as suggesting codes with syntactic errors, variable misuse etc. Some studies argue that code-LLMs perform well on coding tasks because they use self-attention and hidden representations to encode relations among input tokens. However, previous works have not studied what code properties are not encoded by code-LLMs. In this paper, we conduct a fine-grained analysis of attention maps and hidden representations of code-LLMs. Our study indicates that code-LLMs only encode relations among specific subsets of input tokens. Specifically, by categorizing input tokens into syntactic tokens and identifiers, we found that models encode relations among syntactic tokens and among identifiers, but they fail to encode relations between syntactic tokens and identifiers. We also found that fine-tuned models encode these relations poorly compared to their pre-trained counterparts. Additionally, larger models with billions of parameters encode significantly less information about code than models with only a few hundred million parameters.
The last decade has seen blossoming research in deep learning theory attempting to answer, "Why does deep learning generalize?" A powerful shift in perspective precipitated this progress: the study of overparametrized models in the interpolation regime. In this paper, we argue that another perspective shift is due, since some of the desirable qualities of LLMs are not a consequence of good statistical generalization and require a separate theoretical explanation. Our core argument relies on the observation that AR probabilistic models are inherently non-identifiable: models zero or near-zero KL divergence apart -- thus, equivalent test loss -- can exhibit markedly different behaviors. We support our position with mathematical examples and empirical observations, illustrating why non-identifiability has practical relevance through three case studies: (1) the non-identifiability of zero-shot rule extrapolation; (2) the approximate non-identifiability of in-context learning; and (3) the non-identifiability of fine-tunability. We review promising research directions focusing on LLM-relevant generalization measures, transferability, and inductive biases.
Our study aims to identify behavior patterns in cultural values exhibited by large language models (LLMs). The studied variants include question ordering, prompting language, and model size. Our experiments reveal that each tested LLM can efficiently behave with different cultural values. More interestingly: (i) LLMs exhibit relatively consistent cultural values when presented with prompts in a single language. (ii) The prompting language e.g., Chinese or English, can influence the expression of cultural values. The same question can elicit divergent cultural values when the same LLM is queried in a different language. (iii) Differences in sizes of the same model (e.g., Llama2-7B vs 13B vs 70B) have a more significant impact on their demonstrated cultural values than model differences (e.g., Llama2 vs Mixtral). Our experiments reveal that query language and model size of LLM are the main factors resulting in cultural value differences.
This paper presents an analysis of open-source large language models (LLMs) and their application in Retrieval-Augmented Generation (RAG) tasks, specific for enterprise-specific data sets scraped from their websites. With the increasing reliance on LLMs in natural language processing, it is crucial to evaluate their performance, accessibility, and integration within specific organizational contexts. This study examines various open-source LLMs, explores their integration into RAG frameworks using enterprise-specific data, and assesses the performance of different open-source embeddings in enhancing the retrieval and generation process. Our findings indicate that open-source LLMs, combined with effective embedding techniques, can significantly improve the accuracy and efficiency of RAG systems, offering a viable alternative to proprietary solutions for enterprises.