论文检索

Mingrun Wei, Yeyu Yan, Dong Wang

Recently, post-training quantization (PTQ) methods for large language models (LLMs) primarily focus on tackling the challenges caused by outliers. Scaling transformation has proven to be effective while how to enhance the performance of extremely low-bitwidth (e.g., 2-bit) PTQ under it remains largely unexplored. In this work, a new PTQ framework, namely MPPQ, is established. Specifically, MPPQ first proposes an enhanced reconstruction loss based on Mixed metric supervision to mitigate the distribution inconsistency caused by quantization while providing strong regularization for learnable parameters. Secondly, we introduce a Proxy-based adaptive rounding scheme in weight quantization, which replaces the round-to-nearest (RTN) function to minimize the overall quantization errors through element-wise scaling. Furthermore, a factor coarse Pre-searching mechanism is presented to ensure proper coordination between quantization and clipping patterns, while achieving optimal initialization of clipping factors before training. Extensive experiments show that MPPQ consistently outperforms state-of-the-art methods in low-bit quantization settings. For instance, the perplexity of WikiText2 can be dramatically reduced to 8.85 (3.9 ↓ vs 12.75 of the latest method, LRQuant) for the LLaMA-2-7B model, which is quantized with W4A4.

Yufeng Wang, Jinwu Hu, Ziteng Huang, Kunyang Lin, Zitian Zhang, Peihao Chen, Yu Hu, Qianyue Wang, Zhuliang Yu, Bin Sun 等

Open-domain dialogue systems aim to generate natural and engaging conversations, providing significant practical value in real applications such as social robotics and personal assistants. The advent of large language models (LLMs) has greatly advanced this field by improving context understanding and conversational fluency. However, existing LLM-based dialogue systems often fall short in proactively understanding the user's chatting preferences and guiding conversations toward user-centered topics. This lack of user-oriented proactivity can lead users to feel unappreciated, reducing their satisfaction and willingness to continue the conversation in human-computer interactions. To address this issue, we propose a User-oriented Proactive Chatbot (UPC) to enhance the user-oriented proactivity. Specifically, we first construct a critic to evaluate this proactivity inspired by the LLM-as-a-judge strategy. Given the scarcity of high-quality training data, we then employ the critic to guide dialogues between the chatbot and user agents, generating a corpus with enhanced user-oriented proactivity. To ensure the diversity of the user backgrounds, we introduce the ISCO-800, a diverse user background dataset for constructing user agents. Moreover, considering the communication difficulty varies among users, we propose an iterative curriculum learning method that trains the chatbot from easy-to-communicate users to more challenging ones, thereby gradually enhancing its performance. Experiments demonstrate that our proposed training method is applicable to different LLMs, improving user-oriented proactivity and attractiveness in open-domain dialogues. Code and appendix are available at github.com/wang678/LLM-UPC.

Peng Wang, Shuo Li, Ruoxi Zhou, Qiguang Chen, Xiao Xu, Hao Fei, Dagang Li, Wanxiang Che, Libo Qin

Consistency identification in task-oriented dialog (CI-ToD) typically consists of three sub-tasks: User Query Inconsistency (QI) identification, Dialogue History Inconsistency (HI) identification, and Knowledge Base Inconsistency (KBI) identification, which aim to determine inconsistent relationships between system response and user query, dialogue history, and knowledge base. Previous approaches focus on the exploration of deep learning models for CI-ToD. While these models achieve remarkable progress, they still rely on large amounts of labeled data, which is hard to achieve in real-world scenarios. Motivated by this, in the paper, we aim to explore large language models for CI-ToD, which do not require any training data. In addition, we further introduce a multi-agent collaboration framework (MAC-CIToD) to model the interaction across three sub-tasks in CI-ToD, including (1) Full Connection paradigm, (2) Cycle Connection paradigm, and (3) Central Connection paradigm, which effectively builds interaction across QI, HI, and KBI. Experiments on the standard benchmark reveal that our framework achieves superior performance. Additionally, we compare MAC-CIToD with the most advanced trained approaches and find that its zero-shot performance on most metrics even surpasses that of models after training on the CI-ToD dataset.

Chen Tang, Tingrui Shen, Xinrong Gong, Chong Zhao, Tong Zhang

In Multimodal Sentiment Analysis (MSA), data noise arising from various sources can lead to uncertainty in Aleatoric Uncertainty (AU), significantly impacting model performance. Current efforts to address AU have insufficiently explored its sources. They primarily focus on modeling noise rather than implementing targeted modeling based on its origin. Consequently, these approaches struggle to effectively mitigate the influence of AU, resulting in sustained limitations in model performance. Our research identifies that the AU primarily stems from two problems: subjective bias in the annotation process and the complex set relationships of sentiment features. To specifically address them, we propose DFMU, a Distribution-based Framework for Modeling Aleatoric Uncertainty, which incorporates an uncertainty modeling block capable of encoding uncertainty distributions and adaptively adjusting optimization objectives. Furthermore, we introduce distribution-based contrastive learning with sentiment words replacement to better capture the complex relationships among features. Extensive experiments on three public MSA datasets, i.e., MOSI, MOSEI, and SIMS, demonstrate that the proposed model maintains robust performance even under high noise conditions and achieves state-of-the-art results on these popular datasets.

Dominic Simon, Rickard Ewetz

Large Language Models (LLMs) require lightweight avenues of updating stored information that has fallen out of date. Knowledge Editing (KE) approaches have been successful in updating model knowledge for simple factual queries but struggle with handling tasks that require compositional reasoning such as multi-hop question answering (MQA). We observe that existing knowledge editors leverage decompositional techniques that result in illogical reasoning processes. In this paper, we propose a knowledge editor for MQA based on semantic analysis called CHECK. Our framework is based on insights from an analogy between compilers and reasoning using LLMs. Similar to how source code is first compiled before being executed, we propose to semantically analyze reasoning chains before executing the chains to answer questions. Reasoning chains with semantic errors are revised to ensure consistency through logic optimization and re-prompting the LLM model at a higher temperature. We evaluate the effectiveness of CHECK against five state-of-the-art frameworks on four datasets and achieve an average 22.8% improved MQA accuracy.

Yimeng Ren, Yanhua Yu, Lizi Liao, Yuhu Shang, Kangkang Lu, Mingliang Yan

The task of Knowledge-Based Question Generation (KBQG) involves generating natural language questions from structured knowledge sources, posing unique challenges in balancing linguistic diversity and semantic relevance. Existing models often focus on maximizing surface-level similarity to ground-truth questions, neglecting the need for diverse syntactic forms and leading to semantic drift during generation. To overcome these challenges, we propose Refine-Reinforced Diverse Question Generation (R2DQG), a two-phase framework leveraging a generation-then-refinement paradigm. The Generator first constructs a diverse set of expressive templates using dependency parse tree similarity, capturing a wide range of syntactic patterns and styles. These templates guide the creation of question drafts, ensuring both diversity and semantic relevance. In the second phase, a Corrector module refines the drafts to mitigate semantic drift and enhance overall coherence and quality. Experiments on public datasets show that R2DQG outperforms state-of-the-art models in generating diverse, contextually accurate questions. Moreover, synthetic datasets generated by R2DQG enhance downstream QA performance, underscoring the practical utility of our approach.

Kun Peng, Chaodong Tong, Cong Cao, Hao Peng, Qian Li, Guanlin Wu, Lei Jiang, Yanbing Liu, Philip S. Yu

Aspect sentiment triplet extraction (ASTE) aims to extract triplets composed of aspect terms, opinion terms, and sentiment polarities from given sentences. The table tagging method is a popular approach to addressing this task, which encodes a sentence into a 2-dimensional table, allowing for the tagging of relations between any two words. Previous efforts have focused on designing various downstream relation learning modules to better capture interactions between tokens in the table, revealing that a stronger capability in relation capture can lead to greater improvements in the model. Motivated by this, we attempt to directly utilize transformer layers as downstream relation learning modules. Due to the powerful semantic modeling capability of transformers, it is foreseeable that this will lead to excellent improvement. However, owing to the quadratic relation between the length of the table and the length of the input sentence sequence, using transformers directly faces two challenges: overly long table sequences and unfair local attention interaction. To address these challenges, we propose a novel Table-Transformer (T-T) for the tagging-based ASTE method. Specifically, we introduce a stripe attention mechanism with a loop-shift strategy to tackle these challenges. The former modifies the global attention mechanism to only attend to a 2-dimensional local attention window, while the latter facilitates interaction between different attention windows. Extensive and comprehensive experiments demonstrate that the T-T, as a downstream relation learning module, achieves state-of-the-art performance with lower computational costs.

Seungcheol Park, Sojin Lee, Jongjin Kim, Jinsik Lee, Hyunjik Jo, U Kang

How can we accelerate large language models (LLMs) without sacrificing accuracy? The slow inference speed of LLMs hinders us to benefit from their remarkable performance in diverse applications. This is mainly because numerous sublayers are stacked together in LLMs. Sublayer pruning compresses and expedites LLMs via removing unnecessary sublayers. However, existing sublayer pruning algorithms are limited in accuracy since they naively select sublayers to prune, overlooking the different characteristics of each sublayer. In this paper, we propose SPRINT (Sublayer Pruning with Latency and Tunability Information), an accurate sublayer pruning method for LLMs. SPRINT accurately selects a target sublayer to prune by considering 1) the amount of latency reduction after pruning and 2) the tunability of sublayers. SPRINT iteratively prunes redundant sublayers and swiftly tunes the parameters of remaining sublayers. Experiments show that SPRINT achieves the best accuracy-speedup trade-off, exhibiting up to 23.88%p higher accuracy on zero-shot commonsense reasoning benchmarks compared to existing pruning algorithms.

Zhaoxi Mu, Xinyu Yang, Gang Wang

While contemporary speech separation technologies adeptly process lengthy mixed audio waveforms, they are frequently challenged by the intricacies of real-world environments, including noisy and reverberant settings, which can result in artifacts or distortions in the separated speech. To overcome these limitations, we introduce SepALM, a pioneering approach that employs audio language models (ALMs) to rectify and re-synthesize speech within the text domain following preliminary separation. SepALM comprises four core components: a separator, a corrector, a synthesizer, and an aligner. By integrating an ALM-based end-to-end error correction mechanism, we mitigate the risk of error accumulation and circumvent the optimization hurdles typically encountered in conventional methods that amalgamate automatic speech recognition (ASR) with large language models (LLMs). Additionally, we have developed Chain-of-Thought (CoT) prompting and knowledge distillation techniques to facilitate the reasoning and training processes of the ALM. Our experiments substantiate that SepALM not only elevates the precision of speech separation but also markedly bolsters adaptability in novel acoustic environments.

Behzad Mehrbakhsh, Fernando Martínez-Plumed, José Hernández-Orallo

Contamination in large language models (LLMs), and machine learning more broadly, refers to the inclusion of equal --or very similar-- examples in both training and test sets. This phenomenon usually translates into better test performance. Here we explore when this contamination is performed intentionally, for purposes that can be malicious (e.g., get better scores in evaluations) or benevolent (e.g., fix some mistakes). These interventions, usually in the form of fine-tuning memorisations, come with a budget in the size of the fine-tuning dataset. Several trade-offs appear between the breadth of the intervention (how many examples to be memorised), its depth (how many repetitions of each example) and the difficulty of the examples. By studying several LLMs and datasets, we observe some monotonic behaviour (more difficult items require more depth to be `fixed') but also some non-monotonic phenomena (very high depth levels have negative effects on non-contaminated examples). This suggests that trade-offs should be found not only in terms of the budget but also according to model specifics, the task and the item difficulty at hand.

Chengcheng Mai, Yuxiang Wang, Ziyu Gong, Hanxiang Wang, Yihua Huang

Document-level relation extraction (Doc-RE) aims to extract relations between entities across multiple sentences. Therefore, Doc-RE requires more comprehensive reasoning abilities like humans, involving complex cross-sentence interactions between entities, contexts, and external general knowledge, compared to the sentence-level RE. However, most existing Doc-RE methods focus on optimizing single reasoning ability, but lack the ability to utilize external knowledge for comprehensive reasoning on long documents. To solve these problems, a knowledge retrieval augmented method, named KnowRA, was proposed with comprehensive reasoning to autonomously determine whether to accept external knowledge to assist Doc-RE. Firstly, we constructed a document graph for semantic encoding and integrated the co-reference resolution model to augment the co-reference reasoning ability. Then, we expanded the document graph into a document knowledge graph by retrieving the external knowledge base for common-sense reasoning and a novel knowledge filtration method was presented to filter out irrelevant knowledge. Finally, we proposed the axis attention mechanism to build direct and indirect associations with intermediary entities for achieving cross-sentence logical reasoning. Extensive experiments conducted on two datasets verified the effectiveness of our method compared to the state-of-the-art baselines. Our code is available at https://anonymous.4open.science/r/KnowRA.

Xinhe Li, Jiajun Liu, Peng Wang

Recent studies have demonstrated that Large Language Models (LLMs) have strong mathematical reasoning abilities but rely on hundreds of billions of parameters. To tackle the challenge of poor reasoning in Small Language Models (SLMs), existing methods typically leverage LLMs to generate massive amounts of data for cramming training. In psychology, they are akin to System 1 thinking, which resolves reasoning problems rapidly based on experience and intuition. However, human learning also requires System 2 thinking, where knowledge is first acquired and then reinforced through practice. Inspired by such two distinct modes of thinking, we propose a novel method based on the multi-LoRA Interaction for mathematical reasoning Distillation (LoRID). First, we input the question and reasoning of each sample into an LLM to create knowledge-enhanced datasets. Subsequently, we train a LoRA block on the student model as an Intuitive Reasoner (IR), which directly generates Chain-of-Thoughts for problem-solving. Then, to imitate System 2 thinking, we train the Knowledge Generator (KG) and Deep Reasoner (DR), respectively. The former outputs only knowledge after receiving problems, while the latter uses that knowledge to perform reasoning. Finally, to address the randomness in the generation of IR and DR, we evaluate whether their outputs are consistent, and the inference process needs to be iterated if not. This step can enhance the mathematical reasoning ability of SLMs through mutual feedback. Experimental results show that LoRID achieves state-of-the-art performance, especially on the GSM8K dataset, where it outperforms the second-best method by 2.3%, 16.1%, 2.4%, 12.3%, and 1.8% accuracy across the five base models, respectively. Meanwhile, we select four strong baselines as System 1, and after integrating them with our method, the reasoning ability of student models is consistently and significantly improved. The datasets and codes are available at https://github.com/Xinhe-Li/LoRID.

Qian Li, Xuchen Li, Zongyu Chang, Yuzheng Zhang, Cheng Ji, Shangguang Wang

Satellite technology has significantly influenced our daily lives, manifested in applications such as navigation and communication. With its development, a vast amount of multimodal satellite commonsense data has been generated, thus leading to an urgent demand for conversation about satellite data. However, existing large language models suffer from prevalent hallucinations and poor comprehensibility on multimodal satellite data due to their high professional content threshold and partial information opacity. To address these issues, we propose a multimodal satellite knowledge retrieval-augmented iterative alignment framework (Sat-RIA) for satellite commonsense conversation. We first construct multi-view retrieval expert knowledge to reduce hallucinations and enhance the interpretability of responses, which incorporates the satellite expert database, satellite rule, satellite image database, and a satellite knowledge graph. We next design commonsense conversation instructions to make the answers more legible and understandable. Furthermore, the retrieval-augmented iterative alignment module refines response precision by aligning outputs with task-specific standards through multi-stage evaluations. Finally, we construct satellite multi-turn dialogue and visual question-answer datasets for a more comprehensive evaluation of satellite commonsense conversation. Experimental results demonstrate that Sat-RIA outperforms existing large language models and provides more comprehensible answers with fewer hallucinations.

Qian Li, Cheng Ji, Shu Guo, Kun Peng, Qianren Mao, Shangguang Wang

Multi-modal relation extraction (MMRE) is a challenging task that seeks to identify relationships between entities with textual and visual attributes. However, existing methods struggle to handle the complexities posed by multiple entity pairs within a single sentence that share similar contextual information (e.g., identical text and image content). These scenarios amplify the difficulty of distinguishing relationships and hinder accurate extraction. To address these limitations, we propose the variational multi-modal hypergraph attention network (VM-HAN), a novel and robust framework for MMRE. Unlike previous approaches, VM-HAN constructs a multi-modal hypergraph for each sentence-image pair, explicitly modeling high-order intra-/inter-modal correlations among different entity pairs in the same context. This design enables a more detailed and nuanced understanding of entity relationships by capturing intricate cross-modal interactions that are often overlooked. Additionally, we introduce the variational hypergraph attention network (V-HAN). This variational attention mechanism dynamically refines the hypergraph structure, enabling the model to effectively handle the inherent ambiguity and complexity of multi-modal data. Comprehensive experiments on benchmark MMRE datasets demonstrate that VM-HAN achieves state-of-the-art performance, significantly surpassing existing methods in both accuracy and efficiency.

Mingxin Li, Wenhao Wang, Hongru Ji, Xianghua Li, Chao Gao

Large Language Models (LLMs) demonstrate strong capabilities in video understanding. However, it exhibits hallucinations and factual errors in video description. On the one hand, existing Multimodal Large Language Models (MLLMs) are primarily trained by combining language models and vision models, with their visual understanding capabilities depending on the performance of the backbone. Moreover, video descriptions often suffer from incomplete content and the possibility of errors. Given the proven assessment of the strong reasoning capabilities of LLMs, this paper proposes ERSR, a novel Entity and Relationship based Self-Enhanced Reasoning method for imperfect video understanding. Specifically, an entities and relationships strategy is designed to perform scene graphs based on the limited observed entity relationships, thereby enhancing video descriptions. Furthermore, by providing question feedbacks, a self-enhanced forward and feedback reasoning strategy is provided to enhance reasoning logic. Finally, the prediction question answering results are re-validated through rethinking and verifying using the LLMs. Extensive experiments show that the proposed method achieves competitive results on real-world video understanding datasets, with an overall improvement of no less than 1.4%.

Jie Li, Shifei Ding, Lili Guo, Xuan Li

Emotion Recognition in Conversation (ERC) aims to detect the emotions of individual utterances within a conversation. Generating efficient and modality-specific representations for each utterance remains a significant challenge. Previous studies have proposed various models to integrate features extracted using different modality-specific encoders. However, they neglect the varying contributions of modalities to this task and introduce high complexity by aligning modalities at the frame level. To address these challenges, we propose the Multi-modal Anchor Gated Transformer with Knowledge Distillation (MAGTKD) for the ERC task. Specifically, prompt learning is employed to enhance textual modality representations, while knowledge distillation is utilized to strengthen representations of weaker modalities. Furthermore, we introduce a multi-modal anchor gated transformer to effectively integrate utterance-level representations across modalities. Extensive experiments on the IEMOCAP and MELD datasets demonstrate the effectiveness of knowledge distillation in enhancing modality representations and achieve state-of-the-art performance in emotion recognition. Our code is available at: https://github.com/JieLi-dd/MAGTKD.

Andong Li, Tong Lei, Zhihang Sun, Rilin Chen, Erwei Yin, Xiaodong Li, Chengshi Zheng

Despite the rapid development of neural vocoders in recent years, they usually suffer from some intrinsic challenges like opaque modeling, and parameter-performance trade-off. In this study, we propose an innovative time-frequency (T-F) domain-based neural vocoder to resolve the above-mentioned challenges. To be specific, we bridge the connection between the classical signal range-null decomposition (RND) theory and vocoder task, and the reconstruction of target spectrogram can be decomposed into the superimposition between the range-space and null-space, where the former is enabled by a linear domain shift from the original mel-scale domain to the target linear-scale domain, and the latter is instantiated via a learnable network for further spectral detail generation. Accordingly, we propose a novel dual-path framework, where the spectrum is hierarchically encoded/decoded, and the cross- and narrow-band modules are elaborately devised for efficient sub-band and sequential modeling. Comprehensive experiments are conducted on the LJSpeech and LibriTTS benchmarks. Quantitative and qualitative results show that while enjoying lightweight network parameters, the proposed approach yields state-of-the-art performance among existing advanced methods. Our code and the pretrained model weights are available at https://github.com/Andong-Li-speech/RNDVoC.

Tong Lei, Zhiyu Zhang, Rilin Chen, Meng Yu, Jing Lu, Chengshi Zheng, Dong Yu, Andong Li

While previous diffusion-based neural vocoders typically follow a noise-to-data generation pipe-line, the linear-degradation prior of the mel-spectrogram is often neglected, resulting in limited generation quality. By revisiting the vocoding task and excavating its connection with the signal restoration task, this paper proposes a time-frequency (T-F) domain-based neural vocoder with the Schrödinger Bridge, called BridgeVoC, which is the first to follow the data-to-data generation paradigm. Specifically, the mel-spectrogram can be projected into the target linear-scale domain and regarded as a degraded spectral representation with a deficient rank distribution. Based on this, the Schrödinger Bridge is leveraged to establish a connection between the degraded and target data distributions. During the inference stage, starting from the degraded representation, the target spectrum can be gradually restored rather than generated from a Gaussian noise process. Quantitative experiments on LJSpeech and LibriTTS show that BridgeVoC achieves faster inference and surpasses existing diffusion-based vocoder baselines, while also matching or exceeding non-diffusion state-of-the-art methods across evaluation metrics.

Jemin Lee, Sihyeong Park, Jinse Kwon, Jihun Oh, Yongin Kwon

Quantization has gained attention as a promising solution for the cost-effective deployment of large and small language models. However, most prior work has been limited to perplexity or basic knowledge tasks and lacks a comprehensive evaluation of recent models like Llama-3.3. In this paper, we conduct a comprehensive evaluation of instruction-tuned models spanning 1B to 405B parameters, applying four quantization methods across 13 datasets. Our findings reveal that (1) quantized models generally surpass smaller FP16 baselines, yet they often struggle with instruction-following and hallucination detection; (2) FP8 consistently emerges as the most robust option across tasks, and AWQ tends to outperform GPTQ in weight-only quantization; (3) smaller models can suffer severe accuracy drops at 4-bit quantization, while 70B-scale models maintain stable performance; (4) notably, \textit{hard} tasks do not always experience the largest accuracy losses, indicating that quantization magnifies a model’s inherent weaknesses rather than simply correlating with task difficulty; and (5) an LLM-based judge (MT-Bench) highlights significant performance declines in Coding and STEM tasks, though it occasionally reports improvements in reasoning.

Manzong Huang, Chenyang Bu, Yi He, Xindong Wu

Knowledge Graph (KG)-augmented Large Language Models (LLMs) have recently propelled significant advances in complex reasoning tasks, thanks to their broad domain knowledge and contextual awareness. Unfortunately, current methods often assume KGs to be complete, which is impractical given the inherent limitations of KG construction and the potential loss of contextual cues when converting unstructured text into entity-relation triples. In response, this paper proposes the Triple Context Restoration and Query-driven Feedback (TCR-QF) framework, which reconstructs the textual context underlying each triple to mitigate information loss, while dynamically refining the KG structure by iteratively incorporating query-relevant missing knowledge. Experiments on five benchmark question-answering datasets substantiate the effectiveness of TCR-QF in KG and LLM integration, where itachieves a 29.1% improvement in Exact Match and a 15.5% improvement in F1 over its state-of-the-art GraphRAG competitors. The code is publicly available at https://github.com/HFUT-DMiC-Lab/TCR-QF.git.