论文检索

Shaoqing Zhang, Zhuosheng Zhang, Kehai Chen, Rongxiang Weng, Muyun Yang, Tiejun Zhao, Min Zhang

Despite being empowered with alignment mechanisms, large language models (LLMs) are increasingly vulnerable to emerging jailbreak attacks that can compromise their alignment mechanisms. This vulnerability poses significant risks to real-world applications. Existing work faces challenges in both training efficiency and generalization capabilities (i.e., Reinforcement Learning from Human Feedback and Red-Teaming). Developing effective strategies to enable LLMs to resist continuously evolving jailbreak attempts represents a significant challenge. To address this challenge, we propose a novel defensive paradigm called GuidelineLLM, which assists LLMs in recognizing queries that may have harmful content. Before LLMs respond to a query, GuidelineLLM first identifies potential risks associated with the query, summarizes these risks into guideline suggestions, and then feeds these guidelines to the responding LLMs. Importantly, our approach eliminates the necessity for additional safety fine-tuning of the LLMs themselves; only the GuidelineLLM requires fine-tuning. This characteristic enhances the general applicability of GuidelineLLM across various LLMs. Experimental results demonstrate that GuidelineLLM can significantly reduce the attack success rate (ASR) against LLM (an average reduction of 34.17% ASR) while maintaining the usefulness of LLM in handling benign queries.

Ruizhe Zhang, Yongxin Xu, Yuzhen Xiao, Runchuan Zhu, Xinke Jiang, Xu Chu, Junfeng Zhao, Yasha Wang

By integrating external knowledge, Retrieval-Augmented Generation (RAG) has become an effective strategy for mitigating the hallucination problems that large language models (LLMs) encounter when dealing with knowledge-intensive tasks. However, in the process of integrating external non-parametric supporting evidence with internal parametric knowledge, inevitable knowledge conflicts may arise, leading to confusion in the model's responses. To enhance the knowledge selection of LLMs in various contexts, some research has focused on refining their behavior patterns through instruction-tuning. Nonetheless, due to the absence of explicit negative signals and comparative objectives, models fine-tuned in this manner may still exhibit undesirable behaviors such as contextual ignorance and contextual overinclusion. To this end, we propose a Knowledge-aware Preference Optimization strategy, dubbed KnowPO, aimed at achieving adaptive knowledge selection based on contextual relevance in real retrieval scenarios. Concretely, we proposed a general paradigm for constructing knowledge conflict datasets, which comprehensively cover various error types and learn how to avoid these negative signals through preference optimization methods. Simultaneously, we proposed a rewriting strategy and data ratio optimization strategy to address preference imbalances. Experimental results show that KnowPO outperforms previous methods for handling knowledge conflicts by over 37%, while also exhibiting robust generalization across various out-of-distribution datasets.

Lei Zhang, Yunshui Li, Jiaming Li, Xiaobo Xia, Jiaxi Yang, Run Luo, Minzheng Wang, Longze Chen, Junhao Liu, Qiang Qu 等

Some of the latest released Code Large Language Models (Code LLMs) have been trained on repository-level code data, enabling them to perceive repository structures and utilize cross-file code information. This capability allows us to directly concatenate the content of repository code files in prompts to achieve repository-level code completion. However, in real development scenarios, directly concatenating all code repository files in a prompt can easily exceed the context window of Code LLMs, leading to a significant decline in completion performance. Additionally, overly long prompts can increase completion latency, negatively impacting the user experience. In this study, we conducted extensive experiments, including completion error analysis, topology dependency analysis, and cross-file content analysis, to investigate the factors affecting repository-level code completion. Based on the conclusions drawn from these preliminary experiments, we proposed a strategy called **Hierarchical Context Pruning (HCP)** to construct high-quality completion prompts. We applied the **HCP** to six Code LLMs and evaluated them on the CrossCodeEval dataset. The experimental results showed that, compared to previous methods, the prompts constructed using our **HCP** strategy achieved higher completion accuracy on five out of six Code LLMs. Additionally, the **HCP** managed to keep the prompt length around 8k tokens (whereas the full repository code is approximately 50k tokens), significantly improving completion throughput. Our code and data will be publicly available.

Junwei Zhang, Xiaolin Li

Word Sense Disambiguation (WSD) aims to determine the meaning of target words according to the given context. The recognition of high-frequency senses has reached expectations, and the current research focus is mainly on low-frequency senses, namely Long-tail Senses (LTSs). One of the challenges in long-tail WSD is to obtain clear and distinguishable definition representations based on limited word sense definitions. Researchers try to mine word sense definition information from data from different sources to enhance the representations. Inspired by quantum theory, this paper provides a constraint mechanism for representations under non-homogeneous data to leverage the geometric relationship in its Hilbert space to constrain the value range of parameters, thereby alleviating the dependence on big data and improving the accuracy of representations. We theoretically analyze the feasibility of the constraint mechanism, and verify the WSD system based on this mechanism on the standard evaluation framework, constructed LTS datasets and cross-lingual datasets. Experimental results demonstrate the effectiveness of the scheme and achieve competitive performance.

Jinyu Zhang, Zhiyuan You, Jize Wang, Xinyi Le

Document Information Extraction (DIE) aims to extract structured information from Visually Rich Documents (VRDs). Previous full-training approaches have demonstrated strong performance but may struggle with generalization to unseen data. In contrast, training-free methods leverage powerful pre-trained models like Large Language Models (LLMs) to address various downstream tasks with only a few examples. Nonetheless, training-free methods for DIE encounter two primary challenges: (1) understanding the complex relationship between layout and textual elements in VRDs, and (2) providing accurate guidance to pre-trained models. To address these challenges, we propose SAmple-centric In-context Learning (SAIL). SAIL introduces a fine-grained entity-level textual similarity to facilitate in-depth text analysis by LLMs and incorporates layout similarity to enhance the analysis of layouts in VRDs. Moreover, SAIL formulates a unified In-Context Learning (ICL) prompt template for various sample-centric examples, enabling tailored prompts that deliver precise guidance to pre-trained models for each sample. Extensive experiments on FUNSD, CORD, and SROIE benchmarks with various base models (e.g., LLMs) indicate that our SAIL outperforms training-free baselines, even closer to the full-training methods, showing the superiority and generalization of our method.

Hongxuan Zhang, Yao Zhao, Jiaqi Zheng, Chenyi Zhuang, Jinjie Gu, Guihai Chen

The emergence of long-context text applications utilizing large language models (LLMs) has presented significant scalability challenges, particularly in memory footprint. The linear growth of the Key-Value (KV) cache, which stores attention keys and values to reduce redundant computations, can significantly increase memory usage and may prevent models from functioning properly in memory-constrained environments. To address this issue, we propose a novel approach called Cache Sparse Representation (CSR), which converts the KV cache by transforming the dense Key-Value cache tensor into sparse indexes and weights, offering a more memory-efficient representation during LLM inference. Furthermore, we introduce NeuralDict, a novel neural network-based method to automatically generate the dictionary used in our sparse representation. Our extensive experiments demonstrate that CSR matches the performance of state-of-the-art KV cache quantization algorithms while ensuring robust functionality in memory-constrained environments.

Hanyu Zhang, Xiting Wang, Chengao Li, Xiang Ao, Qing He

As large language models (LLMs) are widely deployed across various domains, the ability to control their generated outputs has become more critical. This control involves aligning LLMs outputs with human values and ethical principles or customizing LLMs on specific topics or styles for individual users. Existing controlled generation methods either require significant computational resources and extensive trial-and-error or provide coarse-grained control. In this paper, we propose Generation with Concept Activation Vector (GCAV), a lightweight model control framework that ensures accurate control without requiring resource-extensive fine-tuning. Specifically, GCAV first trains a concept activation vector for specified concepts to be controlled, such as toxicity. During inference, GCAV steers the concept vector in LLMs, for example, by removing the toxicity concept vector from the activation layers. Control experiments from different perspectives, including toxicity reduction, sentiment control, linguistic style, and topic control, demonstrate that our framework achieves state-of-the-art performance with granular control, allowing for fine-grained adjustments of both the steering layers and the steering magnitudes for individual samples.

Congzhi Zhang, Linhai Zhang, Jialong Wu, Yulan He, Deyu Zhou

Despite the notable advancements of existing prompting methods, such as In-Context Learning and Chain-of-Thought for Large Language Models (LLMs), they still face challenges related to various biases. Traditional debiasing methods primarily focus on the model training stage, including approaches based on data augmentation and reweighting, yet they struggle with the complex biases inherent in LLMs. To address such limitations, the causal relationship behind the prompting methods is uncovered using a structural causal model, and a novel causal prompting method based on front-door adjustment is proposed to effectively mitigate LLMs biases. In specific, causal intervention is achieved by designing the prompts without accessing the parameters and logits of LLMs. The chain-of-thought generated by LLM is employed as the mediator variable and the causal effect between input prompts and output answers is calculated through front-door adjustment to mitigate model biases. Moreover, to accurately represent the chain-of-thoughts and estimate the causal effects, contrastive learning is used to fine-tune the encoder of chain-of-thought by aligning its space with that of the LLM. Experimental results show that the proposed causal prompting approach achieves excellent performance across seven natural language processing datasets on both open-source and closed-source LLMs.

Chen Zhang, Dading Chong, Feng Jiang, Chengguang Tang, Anningzhe Gao, Guohua Tang, Haizhou Li

In natural human-to-human conversations, participants often receive feedback signals from one another based on their follow-up reactions. These reactions can include verbal responses, facial expressions, changes in emotional state, and other non-verbal cues. Similarly, in human-machine interactions, the machine can leverage the user's follow-up utterances as feedback signals to assess whether it has appropriately addressed the user's request. Therefore, we propose using the likelihood of follow-up utterances as rewards to differentiate preferred responses from less favored ones, without relying on human or commercial LLM-based preference annotations. Our proposed reward mechanism, ``Follow-up Likelihood as Reward" (FLR), matches the performance of strong reward models trained on large-scale human or GPT-4 annotated data on 8 pairwise-preference and 4 rating-based benchmarks. Building upon the FLR mechanism, we propose to automatically mine preference data from the online generations of a base policy model. The preference data are subsequently used to boost the helpfulness of the base model through direct alignment from preference (DAP) methods, such as direct preference optimization (DPO). Lastly, we demonstrate that fine-tuning the language model that provides follow-up likelihood with natural language feedback significantly enhances FLR's performance on reward modeling benchmarks and effectiveness in aligning the base policy model's helpfulness.

Xupeng Zha, Huan Zhao, Guanghui Ye, Zixing Zhang

Conversational Emotion Recognition (CER) has recently been explored through conversational context modeling to learn the emotion distribution, i.e., the likelihood over emotion categories associated with each utterance. While these methods have shown promising results in emotion classification, they often focus on the interactions between utterances (utterance-view) and overlook shifts in the speaker's emotions (emotion-view). This emphasis on homogeneous view modeling limits their overall effectiveness. To address this limitation, we propose DVL-CER, a novel Dual-View Learning approach for CER. DVL-CER integrates both the utterance-view and emotion-view using two projection heads, enabling cross-view projection of emotion distributions. Our approach offers several key advantages: (1) We introduce an emotion-view that captures shifts in a speaker's emotions from initial to subsequent states within a conversation. This view enriches the conversation modeling and supports seamless integration with various CER baseline models. (2) Our dual-view projection learning strategy flexibly balances consistency and independence between the two heterogeneous views, promoting view-specific adaptation learning and incorporating the emotion verification capability within CER. We validate DVL-CER through extensive experiments on two widely-used datasets, IEMOCAP and EmoryNLP. The results demonstrate that DVL-CER achieves state-of-the-art performance, delivering robust and high-quality emotion distributions compared with existing CER methods and other dual-view learning strategies.

Jinshan Zeng, Xianglong Yu, Xianchao Tong, Wenyan Xiao

Text readability assessment involves categorizing texts based on readers' comprehension levels. Hybrid automatic readability assessment (ARA) models, combining deep and linguistic features, have recently attracted rising attention due to their impressive performance. However, existing hybrid ARA models generally ignore the specific-intrinsic information of deep and linguistic representations, and cannot fully explore their common-intrinsic information. In this paper, we introduce a self-supervised collaborative information bottleneck (SCIB) module for ARA to address these issues. Specifically, we collaboratively consider both specific-intrinsic and common-intrinsic information of the linguistic representation and various levels of deep representations including the document-, sentence- and word-level deep representations, and yield their refined representations via a self-supervised information bottleneck scheme. Extensive experiments are conducted on four English and two Chinese corpora to demonstrate the effectiveness of the proposed model. Experimental results show that the proposed model outperforms state-of-the-art models in terms of four important evaluation metrics, and the suggested SCIB module can effectively capture the specific- and common-intrinsic information.

Amir Zandieh, Majid Daliri, Insu Han

Serving LLMs requires substantial memory due to the storage requirements of Key-Value (KV) embeddings in the KV cache, which grows with sequence length. An effective approach to compress KV cache is quantization. However, traditional quantization methods face significant memory overhead due to the need to store quantization constants (at least a zero point and a scale) in full precision per data block. Depending on the block size, this overhead can add 1 or 2 bits per quantized number. We introduce QJL, a new quantization approach that consists of a Johnson-Lindenstrauss (JL) transform followed by sign-bit quantization. In contrast to existing methods, QJL eliminates memory overheads by removing the need for storing quantization constants. We propose an asymmetric estimator for the inner product of two vectors and demonstrate that applying QJL to one vector and a standard JL transform without quantization to the other provides an unbiased estimator with minimal distortion. We have developed an efficient implementation of the QJL sketch and its corresponding inner product estimator, incorporating a lightweight CUDA kernel for optimized computation. When applied across various LLMs and NLP tasks to quantize the KV cache to only 3 bits, QJL demonstrates a more than fivefold reduction in KV cache memory usage without compromising accuracy, all while achieving faster runtime.

Shengbin Yue, Siyuan Wang, Wei Chen, Xuanjing Huang, Zhongyu Wei

Recent advancements in Large Language Models (LLMs) have led to significant breakthroughs in various natural language processing tasks. However, generating factually consistent responses in knowledge-intensive scenarios remains a challenge due to issues such as hallucination, difficulty in acquiring long-tailed knowledge, and limited memory expansion. This paper introduces SMART, a novel multi-agent framework that leverages external knowledge to enhance the interpretability and factual consistency of LLM-generated responses. SMART comprises four specialized agents, each performing a specific sub-trajectory action to navigate complex knowledge-intensive tasks. We propose a multi-agent co-training paradigm, Long-Short Trajectory Learning, which ensures synergistic collaboration among agents while maintaining fine-grained execution by each agent. Extensive experiments on five knowledge-intensive tasks demonstrate SMART's superior performance compared to widely adopted knowledge internalization and knowledge enhancement methods. Our framework can extend beyond knowledge-intensive tasks to more complex scenarios.

Lin Yuan, Jun Xu, Honghao Gui, Mengshu Sun, Zhiqiang Zhang, Lei Liang, Jun Zhou

High-quality, large-scale instructions are crucial for aligning large language models (LLMs), however, there is a severe shortage of instruction in the field of natural language understanding (NLU). Previous works on constructing NLU instructions mainly focus on information extraction (IE), neglecting tasks such as machine reading comprehension, question answering, and text classification. Furthermore, the lack of diversity in the data has led to a decreased generalization ability of trained LLMs in other NLU tasks and a noticeable decline in the fundamental model's general capabilities. To address this issue, we propose Hum, a large-scale, high-quality synthetic instruction corpus for NLU tasks, designed to enhance the NLU capabilities of LLMs. Specifically, Hum includes IE (either close IE or open IE), machine reading comprehension, text classification, and instruction generalist tasks, thereby enriching task diversity. Additionally, we introduce a human-LLMs collaborative mechanism to synthesize instructions, which enriches instruction diversity by incorporating guidelines, preference rules, and format variants. We conduct extensive experiments on 5 NLU tasks and 28 general capability evaluation datasets for LLMs. Experimental results show that Hum enhances the NLU capabilities of six LLMs by an average of 3.1%, with no significant decline observed in other general capabilities.

Hongwei Yuan, Xiu Tang, Ke Chen, Lidan Shou, Gang Chen, Huan Li

Large language models (LLMs) have significantly advanced the performance of various natural language processing tasks, including text-to-SQL. Current LLM-based text-to-SQL schemes mainly focus on improving the understanding of natural language questions (NLQs) or refining the quality of generated SQLs. While these strategies are effective, they often address specific, nuanced aspects. In contrast, humans approach text-to-SQL with a holistic view, applying transitional logical reasoning across multiple steps to arrive at the final answer. We believe LLMs can leverage human cognitive processes to achieve greater accuracy in text-to-SQL. In this paper, we present COGSQL, a framework featuring a suite of tailored models and strategies aimed at replicating human cognitive processes for enhanced LLM-based text-to-SQL. COGSQL consists of three key modules: (1) SQL preparation: we employ a coarse-to-fine schema linking and syntax keyword prediction, akin to how human recall and align key concepts for better understanding. (2) SQL generation: we introduce a concept-enhanced chain-of-thought prompting, enhancing NLQ interpretation and SQL composition of LLMs, similar to humans drafting SQL query. (3) SQL correction: we develop NLQ consistency and result consistency techniques to correct various errors, mirroring how humans evaluate and refine reasoning. We conduct extensive experiments using diverse benchmarks and LLMs. The results and analysis verify the effectiveness and generalizability of COGSQL.

Hongbang Yuan, Zhuoran Jin, Pengfei Cao, Yubo Chen, Kang Liu, Jun Zhao

LLM have achieved success in many fields but still troubled by problematic content in the training corpora. LLM unlearning aims at reducing their influence and avoid undesirable behaviours. However, existing unlearning methods remain vulnerable to adversarial queries and the unlearned knowledge resurfaces after the manually designed attack queries. As part of a red-team effort to proactively assess the vulnerabilities of unlearned models, we design Dynamic Unlearning Attack (DUA), a dynamic and automated framework to attack these models and evaluate their robustness. It optimizes adversarial suffixes to reintroduce the unlearned knowledge in various scenarios. We find that unlearned knowledge can be recovered in 55.2% of the questions, even without revealing the unlearned model's parameters. In response to this vulnerability, we propose Latent Adversarial Unlearning (LAU), a universal framework that effectively enhances the robustness of the unlearned process. It formulates the unlearning process as a min-max optimization problem and resolves it through two stages: an attack stage, where perturbation vectors are trained and added to the latent space of LLMs to recover the unlearned knowledge, and a defense stage, where previously trained perturbation vectors are used to enhance unlearned model's robustness. With our LAU framework, we obtain two robust unlearning methods, AdvGA and AdvNPO. We conduct extensive experiments across multiple unlearning benchmarks and various models, and demonstrate that they improve the unlearning effectiveness by over 53.5%, cause only less than a 11.6% reduction in neighboring knowledge, and have almost no impact on the model's general capabilities.

Dingbo Yuan, Yipeng Chen, Guodong Liu, Chenchen Li, Chengfu Tang, Dongxu Zhang, Zhenkui Wang, Xudong Wang, Song Liu

Recent years have witnessed a profound evolution in the abilities of Large Language Model, which has significantly boosted the proliferation of role-playing agents and platforms. Nonetheless, there is a conspicuous absence of systematic and comprehensive evaluations of role-playing abilities which are truly aligned with users' interaction scenarios in real-world. To address this gap, we have devised DMT-RoleBench, a benchmark designed to evaluate the role-playing abilities of large language models and agents based on dynamic multi-turn dialogues. Compared with existed role-playing benchmarks, DMT-RoleBench boasts several principal advantages: (1) It contains a more diverse role types and system prompts of different formats. (2) We propose an innovative evaluation paradigm to assess role-playing abilities based on dynamically generating multi-turn dialogues constrained by specific evaluation intents and topics, which is well aligned with users' interaction scenarios in real-world. (3) We define a three-tiered metric system and provide DMT-RM, which is a reward model aligned with human annotations, to annotate the dialogues. And we propose DMT-Score to calculate the final scores based on the annotated dialogues. Our experiments and analysis of leading models equipped with role-playing abilities have demonstrated the effectiveness of DMT-RoleBench.

Yijiong Yu, Yongfeng Huang, Zhixiao Qi, Zhe Zhou

As Large Language Models (LLMs) continue to evolve, more are being designed to handle long-context inputs. Despite this advancement, most of them still face challenges in accurately handling long-context tasks, often showing the "lost in the middle" issue. We identify that insufficient retrieval capability is one of the important reasons for this issue. To tackle this challenge, we propose a novel approach to design training data for long-context tasks, aiming at augmenting LLMs' proficiency in extracting key information from long context. Specially, we incorporate an additional part named "paraphrasing the original text" when constructing the answer of training samples and then fine-tuning the model. Experimenting on LongBench and NaturalQuestions Multi-document-QA dataset with models of Llama and Qwen series, our method achieves an improvement of up to 8.48% and 4.48% in average scores, respectively, showing effectiveness in improving the model’s performance on long-context tasks.

Jiaxing Yu, Xinda Wu, Yunfei Xu, Tieyao Zhang, Songruoyao Wu, Le Ma, Kejun Zhang

Lyric-to-melody generation aims to automatically create melodies based on given lyrics, requiring the capture of complex and subtle correlations between them. However, previous works usually suffer from two main challenges: 1) lyric-melody alignment modeling, which is often simplified to one-syllable/word-to-one-note alignment, while others have the problem of low alignment accuracy; 2) lyric-melody harmony modeling, which usually relies heavily on intermediates or strict rules, limiting model's capabilities and generative diversity. In this paper, we propose SongGLM, a lyric-to-melody generation system that leverages 2D alignment encoding and multi-task pre-training based on the General Language Model (GLM) to guarantee the alignment and harmony between lyrics and melodies. Specifically, 1) we introduce a unified symbolic song representation for lyrics and melodies with word-level and phrase-level (2D) alignment encoding to capture the lyric-melody alignment; 2) we design a multi-task pre-training framework with hierarchical blank infilling objectives (n-gram, phrase, and long span), and incorporate lyric-melody relationships into the extraction of harmonized n-grams to ensure the lyric-melody harmony. We also construct a large-scale lyric-melody paired dataset comprising over 200,000 English song pieces for pre-training and fine-tuning. The objective and subjective results indicate that SongGLM can generate melodies from lyrics with significant improvements in both alignment and harmony, outperforming all the previous baseline methods.

Ji Won Yoon

Existing knowledge distillation (KD) studies for streaming automatic speech recognition (ASR) adopt a non-streaming model as the teacher and a streaming model as the student, respectively. Since the non-streaming teacher usually has less emission latency compared to the streaming student, the teacher's prediction is typically shifted by $\tau$ frames, where the parameter $\tau$ is selected heuristically. In this paper, we observe that this manual shifting is sub-optimal and propose a novel framework, namely Heuristic-free KD. Instead of leveraging knowledge from the non-streaming teacher model, we employ a self-distillation setup, distilling the knowledge within the streaming architecture itself. Since the teacher and student share the same streaming ASR backbone, the alignment mismatch issue can be effectively mitigated without requiring any time shifting by $\tau$. Additionally, we incorporate full-context textual information as an auxiliary multi-modal input for the proposed teacher. Although the streaming architecture lacks future context, the additional linguistic input enables it to generate more accurate knowledge for self-distillation. We empirically demonstrate that the proposed KD approach significantly improves the performance of the streaming ASR model, outperforming conventional methods that rely on the offline teacher and heuristic parameter.