Search

Reciprocal recommender systems~(RRS), conducting bilateral recommendations between two involved parties, have gained increasing attention for enhancing matching efficiency. However, the majority of existing methods in the literature still reuse conventional ranking metrics to separately assess the performance on each side of the recommendation process. These methods overlook the fact that the ranking outcomes of both sides collectively influence the effectiveness of the RRS, neglecting the necessity of a more holistic evaluation and a capable systemic solution. In this paper, we systemically revisit the task of reciprocal recommendation, by introducing the new metrics, formulation, and method. Firstly, we propose five new evaluation metrics that comprehensively and accurately assess the performance of RRS from three distinct perspectives: overall coverage, bilateral stability, and balanced ranking. These metrics provide a more holistic understanding of the system's effectiveness and enable a comprehensive evaluation. Furthermore, we formulate the RRS from a causal perspective, formulating recommendations as bilateral interventions, which can better model the decoupled effects of potential influencing factors. By utilizing the potential outcome framework, we further develop a model-agnostic causal reciprocal recommendation method that considers the causal effects of recommendations. Additionally, we introduce a reranking strategy to maximize matching outcomes, as measured by the proposed metrics. Extensive experiments on two real-world datasets from recruitment and dating scenarios demonstrate the effectiveness of our proposed metrics and approach. The code and dataset are available at: https://github.com/RUCAIBox/CRRS., Comment: KDD 2024

Continual pre-training (CPT) has been an important approach for adapting language models to specific domains or tasks. To make the CPT approach more traceable, this paper presents a technical report for continually pre-training Llama-3 (8B), which significantly enhances the Chinese language ability and scientific reasoning ability of the backbone model. To enhance the new abilities while retaining the original abilities, we design specific data mixture and curriculum strategies by utilizing existing datasets and synthesizing high-quality datasets. Specifically, we synthesize multidisciplinary scientific question and answer (QA) pairs based on related web pages, and subsequently incorporate these synthetic data to improve the scientific reasoning ability of Llama-3. We refer to the model after CPT as Llama-3-SynE (Synthetic data Enhanced Llama-3). We also present the tuning experiments with a relatively small model -- TinyLlama, and employ the derived findings to train the backbone model. Extensive experiments on a number of evaluation benchmarks show that our approach can largely improve the performance of the backbone models, including both the general abilities (+8.81 on C-Eval and +6.31 on CMMLU) and the scientific reasoning abilities (+12.00 on MATH and +4.13 on SciEval), without hurting the original capacities. Our model, data, and codes are available at https://github.com/RUC-GSAI/Llama-3-SynE., Comment: 16 pages, 10 figures, 16 tables

Adapting general large language models (LLMs) to specialized domains presents great challenges due to varied data distributions. This adaptation typically requires continual pre-training on massive domain-specific corpora to facilitate knowledge memorization, followed by training to apply this knowledge following human instructions and preferences. However, this method may result in inefficient knowledge memorization due to a lack of awareness of knowledge utilization and imposes substantial demands on LLMs to simultaneously learn knowledge utilization and format alignment with limited training samples. To facilitate the domain adaptation of LLM, we revise this process and propose a new domain adaptation framework including domain knowledge learning and general format alignment, called Mix-CPT. Specifically, we first conduct a knowledge mixture continual pre-training that concurrently focuses on knowledge memorization and utilization, allowing for mutual reinforcement. To avoid catastrophic forgetting during the continual pre-training process, we further incorporate a logit swap self-distillation constraint. Subsequently, leveraging the knowledge and capabilities acquired during continual pre-training, we efficiently perform instruction tuning and alignment with a few general training samples to achieve format alignment. Extensive experiments demonstrate that our proposed Mix-CPT framework can simultaneously improve the task-solving capabilities of LLMs on the target and general domains compared to the traditional adaptation methods., Comment: LLM, CPT, knowledge learning, format alignment; work in progress

To facilitate the research on large language models (LLMs), this paper presents a comprehensive and unified library, LLMBox, to ease the development, use, and evaluation of LLMs. This library is featured with three main merits: (1) a unified data interface that supports the flexible implementation of various training strategies, (2) a comprehensive evaluation that covers extensive tasks, datasets, and models, and (3) more practical consideration, especially on user-friendliness and efficiency. With our library, users can easily reproduce existing methods, train new models, and conduct comprehensive performance comparisons. To rigorously test LLMBox, we conduct extensive experiments in a diverse coverage of evaluation settings, and experimental results demonstrate the effectiveness and efficiency of our library in supporting various implementations related to LLMs. The detailed introduction and usage guidance can be found at https://github.com/RUCAIBox/LLMBox., Comment: Accepted by ACL 2024 Demo

Large language models (LLMs) have become the foundation of many applications, leveraging their extensive capabilities in processing and understanding natural language. While many open-source LLMs have been released with technical reports, the lack of training details hinders further research and development. This paper presents the development of YuLan, a series of open-source LLMs with $12$ billion parameters. The base model of YuLan is pre-trained on approximately $1.7$T tokens derived from a diverse corpus, including massive English, Chinese, and multilingual texts. We design a three-stage pre-training method to enhance YuLan's overall capabilities. Subsequent phases of training incorporate instruction-tuning and human alignment, employing a substantial volume of high-quality synthesized data. To facilitate the learning of complex and long-tail knowledge, we devise a curriculum-learning framework throughout across these stages, which helps LLMs learn knowledge in an easy-to-hard manner. YuLan's training is finished on Jan, 2024 and has achieved performance on par with state-of-the-art LLMs across various English and Chinese benchmarks. This paper outlines a comprehensive technical roadmap for developing LLMs from scratch. Our model and codes are available at https://github.com/RUC-GSAI/YuLan-Chat.

With the rapid development of video Multimodal Large Language Models (MLLMs), numerous benchmarks have been proposed to assess their video understanding capability. However, due to the lack of rich events in the videos, these datasets may suffer from the short-cut bias that the answers can be deduced from a few frames, without the need to watch the entire video. To address this issue, we introduce Event-Bench, an event-oriented long video understanding benchmark built on existing datasets and human annotations. Event-Bench includes six event-related tasks and 2,190 test instances to comprehensively evaluate video event understanding ability. Additionally, we propose Video Instruction Merging~(VIM), a cost-effective method that enhances video MLLMs using merged, event-intensive video instructions, addressing the scarcity of human-annotated, event-intensive data. Extensive experiments show that the best-performing model, GPT-4o, achieves an overall accuracy of 53.33, significantly outperforming the best open-source model by 41.42%. Leveraging an effective instruction synthesis method and an adaptive model architecture, VIM surpasses both state-of-the-art open-source models and GPT-4V on the Event-Bench. All code, data, and models are publicly available at https://github.com/RUCAIBox/Event-Bench., Comment: Work on progress

The emergence of in-context learning (ICL) is potentially attributed to two major abilities: task recognition (TR) for recognizing the task from demonstrations and utilizing pre-trained priors, and task learning (TL) for learning from demonstrations. However, relationships between the two abilities and how such relationships affect the emergence of ICL is unclear. In this paper, we take the first step by examining the pre-training dynamics of the emergence of ICL. With carefully designed metrics, we find that these two abilities are, in fact, competitive during pre-training. Moreover, we observe a strong negative correlation between the competition and ICL performance. Further analysis of common pre-training factors (i.e., model size, dataset size, and data curriculum) demonstrates possible ways to manage the competition. Based on these insights, we propose a simple yet effective method to better integrate these two abilities for ICL at inference time. Through adaptive ensemble learning, the performance of ICL can be significantly boosted, enabling two small models to outperform a larger one with more than twice the parameters. The code is available at https://github.com/RUCAIBox/Competitive-ICL., Comment: work in progress

Existing LLMs exhibit remarkable performance on various NLP tasks, but still struggle with complex real-world tasks, even equipped with advanced strategies like CoT and ReAct. In this work, we propose the CoAct framework, which transfers the hierarchical planning and collaboration patterns in human society to LLM systems. Specifically, our CoAct framework involves two agents: (1) A global planning agent, to comprehend the problem scope, formulate macro-level plans and provide detailed sub-task descriptions to local execution agents, which serves as the initial rendition of a global plan. (2) A local execution agent, to operate within the multi-tier task execution structure, focusing on detailed execution and implementation of specific tasks within the global plan. Experimental results on the WebArena benchmark show that CoAct can re-arrange the process trajectory when facing failures, and achieves superior performance over baseline methods on long-horizon web tasks. Code is available at https://github.com/xmhou2002/CoAct., Comment: 9 pages, 4 figures

Synthetic data has been proposed as a solution to address the issue of high-quality data scarcity in the training of large language models (LLMs). Studies have shown that synthetic data can effectively improve the performance of LLMs on downstream benchmarks. However, despite its potential benefits, our analysis suggests that there may be inherent flaws in synthetic data. The uniform format of synthetic data can lead to pattern overfitting and cause significant shifts in the output distribution, thereby reducing the model's instruction-following capabilities. Our work delves into these specific flaws associated with question-answer (Q-A) pairs, a prevalent type of synthetic data, and presents a method based on unlearning techniques to mitigate these flaws. The empirical results demonstrate the effectiveness of our approach, which can reverse the instruction-following issues caused by pattern overfitting without compromising performance on benchmarks at relatively low cost. Our work has yielded key insights into the effective use of synthetic data, aiming to promote more robust and efficient LLM training., Comment: 15 pages

Large language models (LLMs) are still struggling in aligning with human preference in complex tasks and scenarios. They are prone to overfit into the unexpected patterns or superficial styles in the training data. We conduct an empirical study that only selects the top-10\% most updated parameters in LLMs for alignment training, and see improvements in the convergence process and final performance. It indicates the existence of redundant neurons in LLMs for alignment training. To reduce its influence, we propose a low-redundant alignment method named \textbf{ALLO}, focusing on optimizing the most related neurons with the most useful supervised signals. Concretely, we first identify the neurons that are related to the human preference data by a gradient-based strategy, then identify the alignment-related key tokens by reward models for computing loss. Besides, we also decompose the alignment process into the forgetting and learning stages, where we first forget the tokens with unaligned knowledge and then learn aligned knowledge, by updating different ratios of neurons, respectively. Experimental results on 10 datasets have shown the effectiveness of ALLO. Our code and data are available at \url{https://github.com/RUCAIBox/ALLO}., Comment: 14 pages, working in progress

Hallucination detection is a challenging task for large language models (LLMs), and existing studies heavily rely on powerful closed-source LLMs such as GPT-4. In this paper, we propose an autonomous LLM-based agent framework, called HaluAgent, which enables relatively smaller LLMs (e.g. Baichuan2-Chat 7B) to actively select suitable tools for detecting multiple hallucination types such as text, code, and mathematical expression. In HaluAgent, we integrate the LLM, multi-functional toolbox, and design a fine-grained three-stage detection framework along with memory mechanism. To facilitate the effectiveness of HaluAgent, we leverage existing Chinese and English datasets to synthesize detection trajectories for fine-tuning, which endows HaluAgent with the capability for bilingual hallucination detection. Extensive experiments demonstrate that only using 2K samples for tuning LLMs, HaluAgent can perform hallucination detection on various types of tasks and datasets, achieving performance comparable to or even higher than GPT-4 without tool enhancements on both in-domain and out-of-domain datasets. We release our dataset and code at https://github.com/RUCAIBox/HaluAgent.

Medical vision-language pre-training methods mainly leverage the correspondence between paired medical images and radiological reports. Although multi-view spatial images and temporal sequences of image-report pairs are available in off-the-shelf multi-modal medical datasets, most existing methods have not thoroughly tapped into such extensive supervision signals. In this paper, we introduce the Med-ST framework for fine-grained spatial and temporal modeling to exploit information from multiple spatial views of chest radiographs and temporal historical records. For spatial modeling, Med-ST employs the Mixture of View Expert (MoVE) architecture to integrate different visual features from both frontal and lateral views. To achieve a more comprehensive alignment, Med-ST not only establishes the global alignment between whole images and texts but also introduces modality-weighted local alignment between text tokens and spatial regions of images. For temporal modeling, we propose a novel cross-modal bidirectional cycle consistency objective by forward mapping classification (FMC) and reverse mapping regression (RMR). By perceiving temporal information from simple to complex, Med-ST can learn temporal semantics. Experimental results across four distinct tasks demonstrate the effectiveness of Med-ST, especially in temporal classification tasks. Our code and model are available at https://github.com/SVT-Yang/MedST., Comment: Accepted at ICML 2024

Transformer-based large language models (LLMs) typically have a limited context window, resulting in significant performance degradation when processing text beyond the length of the context window. Extensive studies have been proposed to extend the context window and achieve length extrapolation of LLMs, but there is still a lack of in-depth interpretation of these approaches. In this study, we explore the positional information within and beyond the context window for deciphering the underlying mechanism of LLMs. By using a mean-based decomposition method, we disentangle positional vectors from hidden states of LLMs and analyze their formation and effect on attention. Furthermore, when texts exceed the context window, we analyze the change of positional vectors in two settings, i.e., direct extrapolation and context window extension. Based on our findings, we design two training-free context window extension methods, positional vector replacement and attention window extension. Experimental results show that our methods can effectively extend the context window length.

Mathematical reasoning is an important capability of large language models~(LLMs) for real-world applications. To enhance this capability, existing work either collects large-scale math-related texts for pre-training, or relies on stronger LLMs (\eg GPT-4) to synthesize massive math problems. Both types of work generally lead to large costs in training or synthesis. To reduce the cost, based on open-source available texts, we propose an efficient way that trains a small LLM for math problem synthesis, to efficiently generate sufficient high-quality pre-training data. To achieve it, we create a dataset using GPT-4 to distill its data synthesis capability into the small LLM. Concretely, we craft a set of prompts based on human education stages to guide GPT-4, to synthesize problems covering diverse math knowledge and difficulty levels. Besides, we adopt the gradient-based influence estimation method to select the most valuable math-related texts. The both are fed into GPT-4 for creating the knowledge distillation dataset to train the small LLM. We leverage it to synthesize 6 million math problems for pre-training our JiuZhang3.0 model, which only needs to invoke GPT-4 API 9.3k times and pre-train on 4.6B data. Experimental results have shown that JiuZhang3.0 achieves state-of-the-art performance on several mathematical reasoning datasets, under both natural language reasoning and tool manipulation settings. Our code and data will be publicly released in \url{https://github.com/RUCAIBox/JiuZhang3.0}., Comment: 28 pages, SOTA math LLM using Well-trained Data Synthesis LLM

Key-value~(KV) caching is an important technique to accelerate the inference of large language models~(LLMs), but incurs significant memory overhead. To compress the size of KV cache, existing methods often compromise precision or require extra data for calibration, limiting their practicality in LLM deployment. In this paper, we introduce \textbf{DecoQuant}, a novel data-free low-bit quantization technique based on tensor decomposition methods, to effectively compress KV cache. Our core idea is to adjust the outlier distribution of the original matrix by performing tensor decomposition, so that the quantization difficulties are migrated from the matrix to decomposed local tensors. Specially, we find that outliers mainly concentrate on small local tensors, while large tensors tend to have a narrower value range. Based on this finding, we propose to apply low-bit quantization to the large tensor, while maintaining high-precision representation for the small tensor. Furthermore, we utilize the proposed quantization method to compress the KV cache of LLMs to accelerate the inference and develop an efficient dequantization kernel tailored specifically for DecoQuant. Through extensive experiments, DecoQuant demonstrates remarkable efficiency gains, showcasing up to a $\sim$75\% reduction in memory footprint while maintaining comparable generation quality., Comment: 11 pages, 6 figures

In real world, large language models (LLMs) can serve as the assistant to help users accomplish their jobs, and also support the development of advanced applications. For the wide application of LLMs, the inference efficiency is an essential concern, which has been widely studied in existing work, and numerous optimization algorithms and code libraries have been proposed to improve it. Nonetheless, users still find it challenging to compare the effectiveness of all the above methods and understand the underlying mechanisms. In this work, we perform a detailed coarse-to-fine analysis of the inference performance of various code libraries. To evaluate the overall effectiveness, we examine four usage scenarios within two practical applications. We further provide both theoretical and empirical fine-grained analyses of each module in the Transformer architecture. Our experiments yield comprehensive results that are invaluable for researchers to evaluate code libraries and improve inference strategies.

To capture user preference, transformer models have been widely applied to model sequential user behavior data. The core of transformer architecture lies in the self-attention mechanism, which computes the pairwise attention scores in a sequence. Due to the permutation-equivariant nature, positional encoding is used to enhance the attention between token representations. In this setting, the pairwise attention scores can be derived by both semantic difference and positional difference. However, prior studies often model the two kinds of difference measurements in different ways, which potentially limits the expressive capacity of sequence modeling. To address this issue, this paper proposes a novel transformer variant with complex vector attention, named EulerFormer, which provides a unified theoretical framework to formulate both semantic difference and positional difference. The EulerFormer involves two key technical improvements. First, it employs a new transformation function for efficiently transforming the sequence tokens into polar-form complex vectors using Euler's formula, enabling the unified modeling of both semantic and positional information in a complex rotation form.Secondly, it develops a differential rotation mechanism, where the semantic rotation angles can be controlled by an adaptation function, enabling the adaptive integration of the semantic and positional information according to the semantic contexts.Furthermore, a phase contrastive learning task is proposed to improve the isotropy of contextual representations in EulerFormer. Our theoretical framework possesses a high degree of completeness and generality. It is more robust to semantic variations and possesses moresuperior theoretical properties in principle. Extensive experiments conducted on four public datasets demonstrate the effectiveness and efficiency of our approach., Comment: Accepted for publication in SIGIR'24

Chain-of-Thought (CoT) prompting can enhance the reasoning capabilities of large language models (LLMs), establishing itself as a primary approach to solving complex reasoning tasks. Existing CoT synthesis approaches usually focus on simpler reasoning tasks and thus result in low-quality and inconsistent CoT prompts. In response to this challenge, we present an empirical investigation of CoT prompting and introduce CoTGenius, a novel framework designed for the automatic generation of superior CoT prompts. CoTGenius is developed based on three major evolution strategies, i.e., complicate, diversify, and specify-alongside two filtering mechanisms: evolutionary success judgement and correctness verification. We further employ CoTGenius to create an extensive CoT dataset, and subsequently fine-tune the Llama 2-Chat 7B and 13B models on this dataset. We call the resulting model ChainLM. To deal with the cumulative error issue in reasoning steps, we propose a step-level debating method, wherein multiple debaters discuss each reasoning step to arrive at the correct answer. Extensive experiments demonstrate that our ChainLM models exhibit enhanced proficiency in addressing a spectrum of complex reasoning problems compared to existing models. In addition, we conduct an in-depth analysis of the impact of data categories within CoTGenius on the model performance. We release our dataset and code at https://github.com/RUCAIBox/ChainLM., Comment: Accepted to LREC-COLING 2024

In online video platforms, reading or writing comments on interesting videos has become an essential part of the video watching experience. However, existing video recommender systems mainly model users' interaction behaviors with videos, lacking consideration of comments in user behavior modeling. In this paper, we propose a novel recommendation approach called LSVCR by leveraging user interaction histories with both videos and comments, so as to jointly conduct personalized video and comment recommendation. Specifically, our approach consists of two key components, namely sequential recommendation (SR) model and supplemental large language model (LLM) recommender. The SR model serves as the primary recommendation backbone (retained in deployment) of our approach, allowing for efficient user preference modeling. Meanwhile, we leverage the LLM recommender as a supplemental component (discarded in deployment) to better capture underlying user preferences from heterogeneous interaction behaviors. In order to integrate the merits of the SR model and the supplemental LLM recommender, we design a twostage training paradigm. The first stage is personalized preference alignment, which aims to align the preference representations from both components, thereby enhancing the semantics of the SR model. The second stage is recommendation-oriented fine-tuning, in which the alignment-enhanced SR model is fine-tuned according to specific objectives. Extensive experiments in both video and comment recommendation tasks demonstrate the effectiveness of LSVCR. Additionally, online A/B testing on the KuaiShou platform verifies the actual benefits brought by our approach. In particular, we achieve a significant overall gain of 4.13% in comment watch time.

In this paper, we study the harmlessness alignment problem of multimodal large language models (MLLMs). We conduct a systematic empirical analysis of the harmlessness performance of representative MLLMs and reveal that the image input poses the alignment vulnerability of MLLMs. Inspired by this, we propose a novel jailbreak method named HADES, which hides and amplifies the harmfulness of the malicious intent within the text input, using meticulously crafted images. Experimental results show that HADES can effectively jailbreak existing MLLMs, which achieves an average Attack Success Rate (ASR) of 90.26% for LLaVA-1.5 and 71.60% for Gemini Pro Vision. Our code and data will be publicly released., Comment: Work in progress

Visual instruction tuning is the key to building multimodal large language models (MLLMs), which greatly improves the reasoning capabilities of large language models (LLMs) in vision scenario. However, existing MLLMs mostly rely on a mixture of multiple highly diverse visual instruction datasets for training (even more than a million instructions), which may introduce data redundancy. To investigate this issue, we conduct a series of empirical studies, which reveal a significant redundancy within the visual instruction datasets, and show that greatly reducing the amount of several instruction dataset even do not affect the performance. Based on the findings, we propose a new data selection approach TIVE, to eliminate redundancy within visual instruction data. TIVE first estimates the task-level and instance-level value of the visual instructions based on computed gradients. Then, according to the estimated values, TIVE determines the task proportion within the visual instructions, and selects representative instances to compose a smaller visual instruction subset for training. Experiments on LLaVA-1.5 show that our approach using only about 7.5% data can achieve comparable performance as the full-data fine-tuned model across seven benchmarks, even surpassing it on four of the benchmarks. Our code and data will be publicly released.

The rise of generative models has driven significant advancements in recommender systems, leaving unique opportunities for enhancing users' personalized recommendations. This workshop serves as a platform for researchers to explore and exchange innovative concepts related to the integration of generative models into recommender systems. It primarily focuses on five key perspectives: (i) improving recommender algorithms, (ii) generating personalized content, (iii) evolving the user-system interaction paradigm, (iv) enhancing trustworthiness checks, and (v) refining evaluation methodologies for generative recommendations. With generative models advancing rapidly, an increasing body of research is emerging in these domains, underscoring the timeliness and critical importance of this workshop. The related research will introduce innovative technologies to recommender systems and contribute to fresh challenges in both academia and industry. In the long term, this research direction has the potential to revolutionize the traditional recommender paradigms and foster the development of next-generation recommender systems.

With the rapid development of recommender systems, there is increasing side information that can be employed to improve the recommendation performance. Specially, we focus on the utilization of the associated \emph{textual data} of items (eg product title) and study how text features can be effectively fused with ID features in sequential recommendation. However, there exists distinct data characteristics for the two kinds of item features, making a direct fusion method (eg adding text and ID embeddings as item representation) become less effective. To address this issue, we propose a novel {\ul \emph{Te}}xt-I{\ul \emph{D}} semantic fusion approach for sequential {\ul \emph{Rec}}ommendation, namely \textbf{\our}. The core idea of our approach is to conduct a sequence-level semantic fusion approach by better integrating global contexts. The key strategy lies in that we transform the text embeddings and ID embeddings by Fourier Transform from \emph{time domain} to \emph{frequency domain}. In the frequency domain, the global sequential characteristics of the original sequences are inherently aggregated into the transformed representations, so that we can employ simple multiplicative operations to effectively fuse the two kinds of item features. Our fusion approach can be proved to have the same effects of contextual convolution, so as to achieving sequence-level semantic fusion. In order to further improve the fusion performance, we propose to enhance the discriminability of the text embeddings from the text encoder, by adaptively injecting positional information via a mixture-of-experts~(MoE) modulation method. Our implementation is available at this repository: \textcolor{magenta}{\url{https://github.com/RUCAIBox/TedRec}}., Comment: 8 pages, 5 figures

Automatic prompt optimization is an important approach to improving the performance of large language models (LLMs). Recent research demonstrates the potential of using LLMs as prompt optimizers, which can generate improved task prompts via iterative refinement. In this paper, we propose a novel perspective to investigate the design of LLM-based prompt optimizers, by drawing an analogy with gradient-based model optimizers. To connect these two approaches, we identify two pivotal factors in model parameter learning: update direction and update method. Focused on the two aspects, we borrow the theoretical framework and learning methods from gradient-based optimization to design improved strategies for LLM-based prompt optimizers. By systematically analyzing a rich set of improvement strategies, we further develop a capable Gradient-inspired LLM-based Prompt Optimizer called GPO. At each step, it first retrieves relevant prompts from the optimization trajectory as the update direction. Then, it utilizes the generation-based refinement strategy to perform the update, while controlling the edit distance through a cosine-based decay strategy. Extensive experiments demonstrate the effectiveness and efficiency of GPO. In particular, GPO brings an additional improvement of up to 56.8% on Big-Bench Hard and 55.3% on MMLU compared to baseline methods.

Due to the excellent capacities of large language models (LLMs), it becomes feasible to develop LLM-based agents for reliable user simulation. Considering the scarcity and limit (e.g., privacy issues) of real user data, in this paper, we conduct large-scale user simulation for web search, to improve the analysis and modeling of user search behavior. Specially, we propose BASES, a novel user simulation framework with LLM-based agents, designed to facilitate comprehensive simulations of web search user behaviors. Our simulation framework can generate unique user profiles at scale, which subsequently leads to diverse search behaviors. To demonstrate the effectiveness of BASES, we conduct evaluation experiments based on two human benchmarks in both Chinese and English, demonstrating that BASES can effectively simulate large-scale human-like search behaviors. To further accommodate the research on web search, we develop WARRIORS, a new large-scale dataset encompassing web search user behaviors, including both Chinese and English versions, which can greatly bolster research in the field of information retrieval. Our code and data will be publicly released soon.

Considering the limited internal parametric knowledge, retrieval-augmented generation (RAG) has been widely used to extend the knowledge scope of large language models (LLMs). Despite the extensive efforts on RAG research, in existing methods, LLMs cannot precisely assess the relevance of retrieved documents, thus likely leading to misleading or even incorrect utilization of external knowledge (i.e., retrieved documents). To address this issue, in this paper, we propose REAR, a RElevance-Aware Retrieval-augmented approach for open-domain question answering (QA). As the key motivation, we aim to enhance the self-awareness of source relevance for LLMs, so as to adaptively utilize external knowledge in RAG systems. Specially, we develop a new architecture for LLM based RAG system, by incorporating a specially designed rank head that precisely assesses the relevance of retrieved documents. Furthermore, we propose an improved training method based on bi-granularity relevance fusion and noise-resistant training. By combining the improvements in both architecture and training, our proposed REAR can better utilize external knowledge by effectively perceiving the relevance of retrieved documents. Experiments on four open-domain QA tasks show that REAR significantly outperforms previous a number of competitive RAG approaches. Our code and data can be accessed at https://github.com/RUCAIBox/REAR.

The ability of the foundation models heavily relies on large-scale, diverse, and high-quality pretraining data. In order to improve data quality, researchers and practitioners often have to manually curate datasets from difference sources and develop dedicated data cleansing pipeline for each data repository. Lacking a unified data processing framework, this process is repetitive and cumbersome. To mitigate this issue, we propose a data processing framework that integrates a Processing Module which consists of a series of operators at different granularity levels, and an Analyzing Module which supports probing and evaluation of the refined data. The proposed framework is easy to use and highly flexible. In this demo paper, we first introduce how to use this framework with some example use cases and then demonstrate its effectiveness in improving the data quality with an automated evaluation with ChatGPT and an end-to-end evaluation in pretraining the GPT-2 model. The code and demonstration videos are accessible on GitHub., Comment: 6 pages, 2 figures; accepted by SIGIR'24 demo track

In this paper, we aim to improve the reasoning ability of large language models (LLMs) over knowledge graphs (KGs) to answer complex questions. Inspired by existing methods that design the interaction strategy between LLMs and KG, we propose an autonomous LLM-based agent framework, called KG-Agent, which enables a small LLM to actively make decisions until finishing the reasoning process over KGs. In KG-Agent, we integrate the LLM, multifunctional toolbox, KG-based executor, and knowledge memory, and develop an iteration mechanism that autonomously selects the tool then updates the memory for reasoning over KG. To guarantee the effectiveness, we leverage program language to formulate the multi-hop reasoning process over the KG, and synthesize a code-based instruction dataset to fine-tune the base LLM. Extensive experiments demonstrate that only using 10K samples for tuning LLaMA-7B can outperform state-of-the-art methods using larger LLMs or more data, on both in-domain and out-domain datasets. Our code and data will be publicly released., Comment: work in progress; efficient 7B LLM-based agent

Reinforcement learning (RL) has been widely used in training large language models (LLMs) for preventing unexpected outputs, eg reducing harmfulness and errors. However, existing RL methods mostly adopt the instance-level reward, which is unable to provide fine-grained supervision for complex reasoning tasks, and can not focus on the few key tokens that lead to the incorrectness. To address it, we propose a new RL method named RLMEC that incorporates a generative model as the reward model, which is trained by the erroneous solution rewriting task under the minimum editing constraint, and can produce token-level rewards for RL training. Based on the generative reward model, we design the token-level RL objective for training and an imitation-based regularization for stabilizing RL process. And the both objectives focus on the learning of the key tokens for the erroneous solution, reducing the effect of other unimportant tokens. The experiment results on mathematical tasks and question-answering tasks have demonstrated the effectiveness of our approach. Our code and data are available at https://github.com/RUCAIBox/RLMEC., Comment: 18 pages, Findings of ACL2024

Recently, large language models such as ChatGPT have showcased remarkable abilities in solving general tasks, demonstrating the potential for applications in recommender systems. To assess how effectively LLMs can be used in recommendation tasks, our study primarily focuses on employing LLMs as recommender systems through prompting engineering. We propose a general framework for utilizing LLMs in recommendation tasks, focusing on the capabilities of LLMs as recommenders. To conduct our analysis, we formalize the input of LLMs for recommendation into natural language prompts with two key aspects, and explain how our framework can be generalized to various recommendation scenarios. As for the use of LLMs as recommenders, we analyze the impact of public availability, tuning strategies, model architecture, parameter scale, and context length on recommendation results based on the classification of LLMs. As for prompt engineering, we further analyze the impact of four important components of prompts, \ie task descriptions, user interest modeling, candidate items construction and prompting strategies. In each section, we first define and categorize concepts in line with the existing literature. Then, we propose inspiring research questions followed by experiments to systematically analyze the impact of different factors on two public datasets. Finally, we summarize promising directions to shed lights on future research., Comment: 40 pages, under review

Recently, multi-task instruction tuning has been applied into sentence representation learning, which endows the capability of generating specific representations with the guidance of task instruction, exhibiting strong generalization ability on new tasks. However, these methods mostly neglect the potential interference problems across different tasks and instances, which may affect the training and convergence of the model. To address it, we propose a data curriculum method, namely Data-CUBE, that arranges the orders of all the multi-task data for training, to minimize the interference risks from the two views. In the task level, we aim to find the optimal task order to minimize the total cross-task interference risk, which is exactly the traveling salesman problem, hence we utilize a simulated annealing algorithm to find its solution. In the instance level, we measure the difficulty of all instances per task, then divide them into the easy-to-difficult mini-batches for training. Experiments on MTEB sentence representation evaluation tasks show that our approach can boost the performance of state-of-the-art methods. Our code and data are publicly available at the link: \url{https://github.com/RUCAIBox/Data-CUBE}., Comment: 14 pages, working in progress

In the era of large language models (LLMs), hallucination (i.e., the tendency to generate factually incorrect content) poses great challenge to trustworthy and reliable deployment of LLMs in real-world applications. To tackle the LLM hallucination, three key questions should be well studied: how to detect hallucinations (detection), why do LLMs hallucinate (source), and what can be done to mitigate them (mitigation). To address these challenges, this work presents a systematic empirical study on LLM hallucination, focused on the the three aspects of hallucination detection, source and mitigation. Specially, we construct a new hallucination benchmark HaluEval 2.0, and designs a simple yet effective detection method for LLM hallucination. Furthermore, we zoom into the different training or utilization stages of LLMs and extensively analyze the potential factors that lead to the LLM hallucination. Finally, we implement and examine a series of widely used techniques to mitigate the hallucinations in LLMs. Our work has led to several important findings to understand the hallucination origin and mitigate the hallucinations in LLMs. Our code and data can be accessed at https://github.com/RUCAIBox/HaluEval-2.0., Comment: 24 pages, 8 figures, 13 tables

Pre-trained recommendation models (PRMs) have attracted widespread attention recently. However, their totally different model structure, huge model size and computation cost hinder their application in practical recommender systems. Hence, it is highly essential to explore how to practically utilize PRMs in real-world recommendations. In this paper, we propose a novel joint knowledge distillation from different pre-trained recommendation models named PRM-KD for recommendation, which takes full advantages of diverse PRMs as teacher models for enhancing student models efficiently. Specifically, PRM-KD jointly distills diverse informative knowledge from multiple representative PRMs such as UniSRec, Recformer, and UniM^2Rec. The knowledge from the above PRMs are then smartly integrated into the student recommendation model considering their confidence and consistency. We further verify the universality of PRM-KD with various types of student models, including sequential recommendation, feature interaction, and graph-based models. Extensive experiments on five real-world datasets demonstrate the effectiveness and efficacy of PRM-KD, which could be viewed as an economical shortcut in practically and conveniently making full use of different PRMs in online systems.

Question Answering over Knowledge Graph (KGQA) aims to seek answer entities for the natural language question from a large-scale Knowledge Graph~(KG). To better perform reasoning on KG, recent work typically adopts a pre-trained language model~(PLM) to model the question, and a graph neural network~(GNN) based module to perform multi-hop reasoning on the KG. Despite the effectiveness, due to the divergence in model architecture, the PLM and GNN are not closely integrated, limiting the knowledge sharing and fine-grained feature interactions. To solve it, we aim to simplify the above two-module approach, and develop a more capable PLM that can directly support subgraph reasoning for KGQA, namely ReasoningLM. In our approach, we propose a subgraph-aware self-attention mechanism to imitate the GNN for performing structured reasoning, and also adopt an adaptation tuning strategy to adapt the model parameters with 20,000 subgraphs with synthesized questions. After adaptation, the PLM can be parameter-efficient fine-tuned on downstream tasks. Experiments show that ReasoningLM surpasses state-of-the-art models by a large margin, even with fewer updated parameters and less training data. Our codes and data are publicly available at~\url{https://github.com/RUCAIBox/ReasoningLM}., Comment: EMNLP-23-Main; simple but effective SOTA on CWQ under a weak-supervised setting

Scaling of neural networks has recently shown great potential to improve the model capacity in various fields. Specifically, model performance has a power-law relationship with model size or data size, which provides important guidance for the development of large-scale models. However, there is still limited understanding on the scaling effect of user behavior models in recommender systems, where the unique data characteristics (e.g. data scarcity and sparsity) pose new challenges to explore the scaling effect in recommendation tasks. In this work, we focus on investigating the scaling laws in large sequential recommendation models. Specially, we consider a pure ID-based task formulation, where the interaction history of a user is formatted as a chronological sequence of item IDs. We don't incorporate any side information (e.g. item text), because we would like to explore how scaling law holds from the perspective of user behavior. With specially improved strategies, we scale up the model size to 0.8B parameters, making it feasible to explore the scaling effect in a diverse range of model sizes. As the major findings, we empirically show that scaling law still holds for these trained models, even in data-constrained scenarios. We then fit the curve for scaling law, and successfully predict the test loss of the two largest tested model scales. Furthermore, we examine the performance advantage of scaling effect on five challenging recommendation tasks, considering the unique issues (e.g. cold start, robustness, long-term preference) in recommender systems. We find that scaling up the model size can greatly boost the performance on these challenging tasks, which again verifies the benefits of large recommendation models.

Recently, large language models (LLMs) have shown great potential in recommender systems, either improving existing recommendation models or serving as the backbone. However, there exists a large semantic gap between LLMs and recommender systems, since items to be recommended are often indexed by discrete identifiers (item ID) out of the LLM's vocabulary. In essence, LLMs capture language semantics while recommender systems imply collaborative semantics, making it difficult to sufficiently leverage the model capacity of LLMs for recommendation. To address this challenge, in this paper, we propose a new LLM-based recommendation model called LC-Rec, which can better integrate language and collaborative semantics for recommender systems. Our approach can directly generate items from the entire item set for recommendation, without relying on candidate items. Specifically, we make two major contributions in our approach. For item indexing, we design a learning-based vector quantization method with uniform semantic mapping, which can assign meaningful and non-conflicting IDs (called item indices) for items. For alignment tuning, we propose a series of specially designed tuning tasks to enhance the integration of collaborative semantics in LLMs. Our fine-tuning tasks enforce LLMs to deeply integrate language and collaborative semantics (characterized by the learned item indices), so as to achieve an effective adaptation to recommender systems. Extensive experiments demonstrate the effectiveness of our method, showing that our approach can outperform a number of competitive baselines including traditional recommenders and existing LLM-based recommenders. Our code is available at https://github.com/RUCAIBox/LC-Rec/., Comment: Accepted by ICDE 2024

Alignment with human preference is a desired property of large language models (LLMs). Currently, the main alignment approach is based on reinforcement learning from human feedback (RLHF). Despite the effectiveness of RLHF, it is intricate to implement and train, thus recent studies explore how to develop alternative alignment approaches based on supervised fine-tuning (SFT). A major limitation of SFT is that it essentially does imitation learning, which cannot fully understand what are the expected behaviors. To address this issue, we propose an improved alignment approach named FIGA. Different from prior methods, we incorporate fine-grained (i.e., token or phrase level) quality signals that are derived by contrasting good and bad responses. Our approach has made two major contributions. Firstly, we curate a refined alignment dataset that pairs initial responses and the corresponding revised ones. Secondly, we devise a new loss function can leverage fine-grained quality signals to instruct the learning of LLMs for alignment. Extensive experiments have demonstrated the effectiveness of our approaches by comparing a number of competitive baselines.

There is a rapidly-growing research interest in modeling user preferences via pre-training multi-domain interactions for recommender systems. However, Existing pre-trained multi-domain recommendations mostly select the item texts to be bridges across domains, and simply explore the user behaviors in target domains. Hence, they ignore other informative multi-modal item contents (e.g., visual information), and also lack of thorough consideration of user behaviors from all interactive domains. To address these issues, in this paper, we propose to pre-train universal multi-modal item content presentation for multi-domain recommendation, called UniM^2Rec, which could smoothly learn the multi-modal item content presentations and the multi-modal user preferences from all domains. With the pre-trained multi-domain recommendation model, UniM^2Rec could be efficiently and effectively transferred to new target domains in practice. Extensive experiments conducted on five real-world datasets in target domains demonstrate the superiority of the proposed method over existing competitive methods, especially for the real-world recommendation scenarios that usually struggle with seriously missing or noisy item contents.

Large language models~(LLMs) have greatly advanced the frontiers of artificial intelligence, attaining remarkable improvement in model capacity. To assess the model performance, a typical approach is to construct evaluation benchmarks for measuring the ability level of LLMs in different aspects. Despite that a number of high-quality benchmarks have been released, the concerns about the appropriate use of these benchmarks and the fair comparison of different models are increasingly growing. Considering these concerns, in this paper, we discuss the potential risk and impact of inappropriately using evaluation benchmarks and misleadingly interpreting the evaluation results. Specially, we focus on a special issue that would lead to inappropriate evaluation, \ie \emph{benchmark leakage}, referring that the data related to evaluation sets is occasionally used for model training. This phenomenon now becomes more common since pre-training data is often prepared ahead of model test. We conduct extensive experiments to study the effect of benchmark leverage, and find that it can dramatically boost the evaluation results, which would finally lead to an unreliable assessment of model performance. To improve the use of existing evaluation benchmarks, we finally present several guidelines for both LLM developers and benchmark maintainers. We hope this work can draw attention to appropriate training and evaluation of LLMs., Comment: 11 pages

Visual instruction tuning is an essential approach to improving the zero-shot generalization capability of Multi-modal Large Language Models (MLLMs). A surge of visual instruction datasets with various focuses and characteristics have been proposed recently, enabling MLLMs to achieve surprising results on evaluation benchmarks. To develop more capable MLLMs, in this paper, we aim to investigate a more fundamental question: ``what makes for good visual instructions?''. By conducting a comprehensive empirical study, we find that instructions focused on complex visual reasoning tasks are particularly effective in improving the performance of MLLMs on evaluation benchmarks. Building upon this finding, we design a systematic approach to automatically creating high-quality complex visual reasoning instructions. Our approach employs a synthesis-complication-reformulation paradigm, leveraging multiple stages to gradually increase the complexity of the instructions while guaranteeing quality. Based on this approach, we create the synthetic visual reasoning instruction dataset consisting of 32K examples, namely ComVint, and fine-tune four MLLMs on it. Experimental results demonstrate that our dataset consistently enhances the performance of all the compared MLLMs, e.g., improving the performance of MiniGPT-4 and BLIP-2 on MME-Cognition by 32.6% and 28.8%, respectively. Our code and data are publicly available at the link: https://github.com/RUCAIBox/ComVint., Comment: Work in progress

Recently, there has been an emergence of employing LLM-powered agents as believable human proxies, based on their remarkable decision-making capability. However, existing studies mainly focus on simulating human dialogue. Human non-verbal behaviors, such as item clicking in recommender systems, although implicitly exhibiting user preferences and could enhance the modeling of users, have not been deeply explored. The main reasons lie in the gap between language modeling and behavior modeling, as well as the incomprehension of LLMs about user-item relations. To address this issue, we propose AgentCF for simulating user-item interactions in recommender systems through agent-based collaborative filtering. We creatively consider not only users but also items as agents, and develop a collaborative learning approach that optimizes both kinds of agents together. Specifically, at each time step, we first prompt the user and item agents to interact autonomously. Then, based on the disparities between the agents' decisions and real-world interaction records, user and item agents are prompted to reflect on and adjust the misleading simulations collaboratively, thereby modeling their two-sided relations. The optimized agents can also propagate their preferences to other agents in subsequent interactions, implicitly capturing the collaborative filtering idea. Overall, the optimized agents exhibit diverse interaction behaviors within our framework, including user-item, user-user, item-item, and collective interactions. The results show that these agents can demonstrate personalized behaviors akin to those of real-world individuals, sparking the development of next-generation user behavior simulation.

Humans often interact with large language models (LLMs) in multi-turn interaction to obtain desired answers or more information. However, most existing studies overlook the multi-turn instruction following ability of LLMs, in terms of training dataset, training method, and evaluation benchmark. In this paper, we introduce Parrot, a solution aiming to enhance multi-turn instruction following for LLMs. First, we introduce an efficient but effective method for collecting multi-turn instructions that feature human-like queries, such as anaphora and ellipsis. Second, we propose a context-aware preference optimization strategy to further enhance LLMs for complex queries in multi-turn interaction. Moreover, to quantitatively evaluate LLMs in multi-turn instruction following, we manually build a multi-turn benchmark derived from existing ones. Extensive experiments show that Parrot improves current LLMs by up to 7.2% in multi-turn instruction following. Our dataset and codes will be open-sourced to facilitate future research.

Large language models (LLMs) have achieved dramatic proficiency over NLP tasks with normal length. Recently, multiple studies have committed to extending the context length and enhancing the long text modeling capabilities of LLMs. To comprehensively evaluate the long context ability of LLMs, we propose BAMBOO, a multi-task long context benchmark. BAMBOO has been designed with four principles: comprehensive capacity evaluation, avoidance of data contamination, accurate automatic evaluation, and different length levels. It consists of 10 datasets from 5 different long text understanding tasks, i.e. question answering, hallucination detection, text sorting, language modeling, and code completion, to cover core capacities and various domains of LLMs. We conduct experiments with five long context models on BAMBOO and further discuss four key research questions of long text. We also qualitatively analyze current long context models and point out future directions for enhancing long text modeling capacities. We release our data, prompts, and code at https://github.com/RUCAIBox/BAMBOO., Comment: Accepted for the Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING) 2024

The field of urban spatial-temporal prediction is advancing rapidly with the development of deep learning techniques and the availability of large-scale datasets. However, challenges persist in accessing and utilizing diverse urban spatial-temporal datasets from different sources and stored in different formats, as well as determining effective model structures and components with the proliferation of deep learning models. This work addresses these challenges and provides three significant contributions. Firstly, we introduce "atomic files", a unified storage format designed for urban spatial-temporal big data, and validate its effectiveness on 40 diverse datasets, simplifying data management. Secondly, we present a comprehensive overview of technological advances in urban spatial-temporal prediction models, guiding the development of robust models. Thirdly, we conduct extensive experiments using diverse models and datasets, establishing a performance leaderboard and identifying promising research directions. Overall, this work effectively manages urban spatial-temporal data, guides future efforts, and facilitates the development of accurate and efficient urban spatial-temporal prediction models. It can potentially make long-term contributions to urban spatial-temporal data management and prediction, ultimately leading to improved urban living standards., Comment: 14 pages, 3 figures, VLDB under review

Autonomous agents have long been a prominent research focus in both academic and industry communities. Previous research in this field often focuses on training agents with limited knowledge within isolated environments, which diverges significantly from human learning processes, and thus makes the agents hard to achieve human-like decisions. Recently, through the acquisition of vast amounts of web knowledge, large language models (LLMs) have demonstrated remarkable potential in achieving human-level intelligence. This has sparked an upsurge in studies investigating LLM-based autonomous agents. In this paper, we present a comprehensive survey of these studies, delivering a systematic review of the field of LLM-based autonomous agents from a holistic perspective. More specifically, we first discuss the construction of LLM-based autonomous agents, for which we propose a unified framework that encompasses a majority of the previous work. Then, we present a comprehensive overview of the diverse applications of LLM-based autonomous agents in the fields of social science, natural science, and engineering. Finally, we delve into the evaluation strategies commonly used for LLM-based autonomous agents. Based on the previous studies, we also present several challenges and future directions in this field. To keep track of this field and continuously update our survey, we maintain a repository of relevant references at https://github.com/Paitesanshi/LLM-Agent-Survey., Comment: 35 pages, 5 figures, 3 tables, has been accepted by frontiers of computer science (FCS), doi={10.1007/s11704-024-40231-1}

Interpreting ancient Chinese has been the key to comprehending vast Chinese literature, tradition, and civilization. In this paper, we propose Erya for ancient Chinese translation. From a dataset perspective, we collect, clean, and classify ancient Chinese materials from various sources, forming the most extensive ancient Chinese resource to date. From a model perspective, we devise Erya training method oriented towards ancient Chinese. We design two jointly-working tasks: disyllabic aligned substitution (DAS) and dual masked language model (DMLM). From an evaluation perspective, we build a benchmark to judge ancient Chinese translation quality in different scenarios and evaluate the ancient Chinese translation capacities of various existing models. Our model exhibits remarkable zero-shot performance across five domains, with over +12.0 BLEU against GPT-3.5 models and better human evaluation results than ERNIE Bot. Subsequent fine-tuning further shows the superior transfer capability of Erya model with +6.2 BLEU gain. We release all the above-mentioned resources at https://github.com/RUCAIBox/Erya., Comment: Accepted by NLPCC 2023

Conversational recommender systems (CRS) aim to provide the recommendation service via natural language conversations. To develop an effective CRS, high-quality CRS datasets are very crucial. However, existing CRS datasets suffer from the long-tail issue, \ie a large proportion of items are rarely (or even never) mentioned in the conversations, which are called long-tail items. As a result, the CRSs trained on these datasets tend to recommend frequent items, and the diversity of the recommended items would be largely reduced, making users easier to get bored. To address this issue, this paper presents \textbf{LOT-CRS}, a novel framework that focuses on simulating and utilizing a balanced CRS dataset (\ie covering all the items evenly) for improving \textbf{LO}ng-\textbf{T}ail recommendation performance of CRSs. In our approach, we design two pre-training tasks to enhance the understanding of simulated conversation for long-tail items, and adopt retrieval-augmented fine-tuning with label smoothness strategy to further improve the recommendation of long-tail items. Extensive experiments on two public CRS datasets have demonstrated the effectiveness and extensibility of our approach, especially on long-tail recommendation., Comment: work in progress

Knowledge-intensive tasks (e.g., open-domain question answering (QA)) require a substantial amount of factual knowledge and often rely on external information for assistance. Recently, large language models (LLMs) (e.g., ChatGPT), have demonstrated impressive prowess in solving a wide range of tasks with world knowledge, including knowledge-intensive tasks. However, it remains unclear how well LLMs are able to perceive their factual knowledge boundaries, particularly how they behave when incorporating retrieval augmentation. In this study, we present an initial analysis of the factual knowledge boundaries of LLMs and how retrieval augmentation affects LLMs on open-domain QA. Specially, we focus on three primary research questions and analyze them by examining QA performance, priori judgement and posteriori judgement of LLMs. We show evidence that LLMs possess unwavering confidence in their capabilities to respond to questions and the accuracy of their responses. Furthermore, retrieval augmentation proves to be an effective approach in enhancing LLMs' awareness of knowledge boundaries, thereby improving their judgemental abilities. Additionally, we also find that LLMs have a propensity to rely on the provided retrieval results when formulating answers, while the quality of these results significantly impacts their reliance. The code to reproduce this work is available at https://github.com/RUCAIBox/LLM-Knowledge-Boundary.

Despite the superior performance, Large Language Models~(LLMs) require significant computational resources for deployment and use. To overcome this issue, quantization methods have been widely applied to reduce the memory footprint of LLMs as well as increasing the inference rate. However, a major challenge is that low-bit quantization methods often lead to performance degradation. It is important to understand how quantization impacts the capacity of LLMs. Different from previous studies focused on overall performance, this work aims to investigate the impact of quantization on \emph{emergent abilities}, which are important characteristics that distinguish LLMs from small language models. Specially, we examine the abilities of in-context learning, chain-of-thought reasoning, and instruction-following in quantized LLMs. Our empirical experiments show that these emergent abilities still exist in 4-bit quantization models, while 2-bit models encounter severe performance degradation on the test of these abilities. To improve the performance of low-bit models, we conduct two special experiments: (1) fine-gained impact analysis that studies which components (or substructures) are more sensitive to quantization, and (2) performance compensation through model fine-tuning. Our work derives a series of important findings to understand the impact of quantization on emergent abilities, and sheds lights on the possibilities of extremely low-bit quantization for LLMs., Comment: 15 pages, 4 figures