论文检索

Wenyi Xiao, Ziwei Huang, Leilei Gan, Wanggui He, Haoyuan Li, Zhelun Yu, Fangxun Shu, Hao Jiang, Linchao Zhu

The rapidly developing Large Vision Language Models (LVLMs) still face the hallucination phenomena where the generated responses do not align with the given contexts, significantly restricting the usages of LVLMs. Most previous work detects and mitigates hallucination at the coarse-grained level or requires expensive annotation (e.g., labeling by human experts or proprietary models). To address these issues, we propose detecting and mitigating hallucinations in LVLMs via fine-grained AI feedback. The basic idea is that we generate a small-size sentence-level hallucination annotation dataset by proprietary models, whereby we train a detection model which can perform sentence-level hallucination detection. Then, we propose a detect-then-rewrite pipeline to automatically construct preference dataset for hallucination mitigation training. Furthermore, we propose differentiating the severity of hallucinations, and introducing a Hallucination Severity-Aware Direct Preference Optimization (HSA-DPO) which prioritizes the mitigation of critical hallucination in LVLMs by incorporating the severity of hallucinations into preference learning. Extensive experiments on hallucination detection and mitigation benchmarks demonstrate that our method sets a new state-of-the-art in hallucination detection on MHaluBench, surpassing GPT-4V and Gemini, and reduces the hallucination rate by 36.1% on AMBER and 76.3% on Object HalBench compared to the base model.

Yuan Xia, Jingbo Zhou, Zhenhui Shi, Jun Chen, Haifeng Huang

The Retrieval-Augmented Language Model (RALM) has demonstrated remarkable performance on knowledge-intensive tasks by integrating external knowledge during inference, which mitigates the factual hallucinations inherited in large language models (LLMs). Despite these advancements, challenges persist in the implementation of RALMs, particularly in terms of reliability and traceability. Specifically, the irrelevant document retrieval may result in unhelpful responses or even deteriorate the performance of LLMs, while the lack of appropriate citations in outputs complicates efforts to verify the trustworthiness of the models. To this end, we propose a novel self-reasoning framework aimed at improving the reliability and traceability of RALMs, whose core idea is to leverage reasoning trajectories generated by the LLM itself. The framework involves constructing self-reasoning trajectories through three processes: a relevance-aware process, an evidence-aware selective process, and a trajectory analysis process. We evaluated our framework across four public datasets (two short-form QA datasets, one long-form QA dataset, and one fact verification dataset) to demonstrate its superiority. Our method can outperform existing state-of-the-art models and achieve performance comparable with GPT-4, using only 2,000 training samples.

Yutong Wu, Di Huang, Wenxuan Shi, Wei Wang, Yewen Pu, Lingzhe Gao, Shihao Liu, Ziyuan Nan, Kaizhao Yuan, Rui Zhang 等

Recent advancements in open-source code large language models (LLMs) have been driven by fine-tuning on the data generated from powerful closed-source LLMs, which are expensive to obtain. This paper explores whether it is possible to use a fine-tuned open-source model to generate additional data to augment its instruction-tuning dataset. We make two observations: (1) A code snippet can serve as the response to different instructions. (2) Instruction-tuned code LLMs perform better at translating code into instructions than the reverse. Based on these observations, we propose Inverse-Instruct, a data augmentation technique that uses a fine-tuned LLM to generate additional instructions of code responses from its own training dataset. The additional instruction-response pairs are added to the original dataset, and a stronger code LLM can be obtained by fine-tuning on the augmented dataset. We empirically validate Inverse-Instruct on a range of open-source code models (e.g. CodeLlama-Python and DeepSeek-Coder) and benchmarks (e.g., HumanEval(+), MBPP(+), DS-1000 and MultiPL-E), showing it consistently improves the base models.

Yihan Wu, Yichen Lu, Yifan Peng, Xihua Wang, Ruihua Song, Shinji Watanabe

Audiovisual Automatic Speech Recognition (AV-ASR) aims to improve speech recognition accuracy by leveraging visual signals. It is particularly challenging in unconstrained real-world scenarios across various domains due to noisy acoustic environments, spontaneous speech, and the uncertain use of visual information. Most previous works fine-tune audio-only ASR models on audiovisual datasets, optimizing them for conventional ASR objectives. However, they often neglect visual features and common errors in unconstrained video scenarios. In this paper, we propose using a preference optimization strategy to improve speech recognition accuracy for real-world videos. First, we create preference data via simulating common errors that occurred in AV-ASR from two focals: manipulating the audio or vision input and rewriting the output transcript. Second, we propose BPO-AVASR, a Bifocal Preference Optimization method to improve AV-ASR models by leveraging both input-side and output-side preference. Extensive experiments demonstrate that our approach significantly improves speech recognition accuracy across various domains, outperforming previous state-of-the-art models on real-world video speech recognition.

Xin Wu, Yuqi Bu, Yifei Chen, Yi Cai

Despite extensive training on diverse datasets and alignment with human values, large language models (LLMs) can still generate fallacious outputs. Additionally, the validity of LLM's outputs varies significantly depending on the content. It is crucial to ensure LLMs' logical consistency across different contexts. Drawing inspiration from cognitive psychology studies, we propose a Logic Control Framework (LCF) that disentangles LLMs' hidden representations into separate content and logic spaces. Within the logic space, we use logically valid and invalid samples to construct distinct regions through contrastive learning. By moving logic representations to logically valid regions and fusing them with unchanged content representations, we significantly reduce logical fallacies in LLM outputs while maintaining content coherence. We demonstrate the effectiveness of LCF through experiments on conclusion generation and fallacy identification tasks, showing a significant improvement in logical validity and a reduction in fallacious outputs.

Xianjie Wu, Jian Yang, Linzheng Chai, Ge Zhang, Jiaheng Liu, Xeron Du, Di Liang, Daixin Shu, Xianfu Cheng, Tianzhen Sun 等

Recent advancements in Large Language Models (LLMs) have markedly enhanced the interpretation and processing of tabular data, introducing previously unimaginable capabilities. Despite these achievements, LLMs still encounter significant challenges when applied in industrial scenarios, particularly due to the increased complexity of reasoning required with real-world tabular data, underscoring a notable disparity between academic benchmarks and practical applications. To address this discrepancy, we conduct a detailed investigation into the application of tabular data in industrial scenarios and propose a comprehensive and complex benchmark TableBench, including 18 fields within four major categories of table question answering (TableQA) capabilities. Furthermore, we introduce TableLLM, trained on our meticulously constructed training set TableInstruct, achieving comparable performance with GPT-3.5. Massive experiments conducted on TableBench indicate that both open-source and proprietary LLMs still have significant room for improvement to meet real-world demands, where the most advanced model, GPT-4, achieves only a modest score compared to humans.

Jiaxing Wu, Lin Ning, Luyang Liu, Harrison Lee, Neo Wu, Chao Wang, Sushant Prakash, Shawn O'Banion, Bradley Green, Jun Xie

LLM-powered personalization agent systems employ Large Language Models (LLMs) to predict users’ behavior from their past activities. However, their effectiveness often hinges on the ability to effectively leverage extensive, long user historical data due to its inherent noise and length of such data. Existing pre-trained LLMs may generate summaries that are concise but lack the necessary context for downstream tasks, hindering their utility in personalization systems. To address these challenges, we introduce Reinforcement Learning from Prediction Feedback (RLPF). RLPF fine-tunes LLMs to generate concise, human-readable user summaries that are optimized for downstream task performance. By maximizing the usefulness of the generated summaries, RLPF effectively distills extensive user history data while preserving essential information for downstream tasks. Our empirical evaluation demonstrates significant improvements in both extrinsic downstream task utility and intrinsic summary quality, surpassing baseline methods by up to 22% and achieving an up to 84.59% win rate on Factuality, Abstractiveness, and Readability. RLPF also achieves a remarkable 74% reduction while improving performance on 16 out of 19 unseen tasks and/or datasets, showcasing its generalizability. This approach offers a promising solution for enhancing LLM personalization by effectively transforming long, noisy user histories into informative and human-readable representations.

Yuqiao Wen, Behzad Shayegh, Chenyang Huang, Yanshuai Cao, Lili Mou

The ability of zero-shot translation emerges when we train a multilingual model with certain translation directions; the model can then directly translate in unseen directions. Alternatively, zero-shot translation can be accomplished by pivoting through a third language (e.g., English). In our work, we observe that both direct and pivot translations are noisy and achieve less satisfactory performance. We propose EBBS, an ensemble method with a novel bi-level beam search algorithm, where each ensemble component explores its own prediction step by step at the lower level but all components are synchronized by a "soft voting" mechanism at the upper level. Results on two popular multilingual translation datasets show that EBBS consistently outperforms direct and pivot translations, as well as existing ensemble techniques. Further, we can distill the ensemble's knowledge back to the multilingual model to improve inference efficiency; profoundly, our EBBS-distilled model can even outperform EBBS as it learns from the ensemble knowledge.

William Watson, Nicole Cho, Nishan Srishankar

Hallucination continues to be one of the most critical challenges in the institutional adoption journey of Large Language Models (LLMs). While prior studies have primarily focused on the post-generation analysis and refinement of outputs, this paper centers on the effectiveness of queries in eliciting accurate responses from LLMs. We present HalluciBot, a model that estimates the query's propensity to hallucinate before generation, without invoking any LLMs during inference. HalluciBot can serve as a proxy reward model for query rewriting, offering a general framework to estimate query quality based on accuracy and consensus. In essence, HalluciBot investigates how poorly constructed queries can lead to erroneous outputs - moreover, by employing query rewriting guided by HalluciBot's empirical estimates, we demonstrate that 95.7% output accuracy can be achieved for Multiple Choice questions. The training procedure for HalluciBot consists of perturbing 369,837 queries n times, employing n+1 independent LLM agents, sampling an output from each query, conducting a Multi-Agent Monte Carlo simulation on the sampled outputs, and training an encoder classifier. The idea of perturbation is the outcome of our ablation studies that measures the increase in output diversity (+12.5 agreement spread) by perturbing a query in lexically different but semantically similar ways. Therefore, HalluciBot paves the way to ratiocinate (76.0% test F1 score, 46.6% in saved computation on hallucinatory queries), rewrite (+30.2% positive class transition from hallucinatory to non-hallucinatory), rank (+50.6% positive class transition from hallucinatory to non-hallucinatory), and route queries to effective pipelines.

Zhuang Wang, Linna Zhou, Xuekai Chen, Zhili Zhou, Zhongliang Yang

Language steganography in social networks primarily focuses on embedding secret information into social media text efficiently to achieve covert communication. The misuse of such techniques could pose significant potential threats to public cyberspace, such as the spread of malicious code, commands, or viruses. Existing social text steganalysis techniques mainly focus on the analysis of individual social media texts. However, the information content in a single text is very limited, leading to poor detection performance in practical applications. To address this challenge, this paper proposes a social text steganalysis method that combines large-scale language models with common-sense knowledge graphs (STLC-KG). This method first uses knowledge graphs to expand the knowledge contained in the text under investigation, enriching its linguistic expression, and then utilizes large-scale language models to extract the linguistic features of the social text. The results of tests conducted on three mainstream social media platforms demonstrate that the proposed method significantly improves the performance of social text steganalysis.

Zhensheng Wang, Wenmian Yang, Kun Zhou, Yiquan Zhang, Weijia Jia

The real estate market relies heavily on structured data, such as property details, market trends, and price fluctuations. However, the lack of specialized Tabular Question Answering datasets in this domain limits the development of automated question-answering systems. To fill this gap, we introduce RETQA, the first large-scale open-domain Chinese Tabular Question Answering dataset for Real Estate. RETQA comprises 4,932 tables and 20,762 question-answer pairs across 16 sub-fields within three major domains: property information, real estate company finance information and land auction information. Compared with existing tabular question answering datasets, RETQA poses greater challenges due to three key factors: long-table structures, open-domain retrieval, and multi-domain queries. To tackle these challenges, we propose the SLUTQA framework, which integrates large language models with spoken language understanding tasks to enhance retrieval and answering accuracy. Extensive experiments demonstrate that SLUTQA significantly improves the performance of large language models on RETQA by in-context learning. RETQA and SLUTQA provide essential resources for advancing tabular question answering research in the real estate domain, addressing critical challenges in open-domain and long-table question-answering.

Yujing Wang, Hainan Zhang, Sijia Wen, Wangjie Qiu, Binghui Guo

Federated learning is susceptible to model poisoning attacks, especially those meticulously crafted for servers. Traditional defense methods mainly focus on updating assessments or robust aggregation against manually crafted myopic attacks. When facing advanced attacks, their defense stability is notably insufficient. Therefore, it is imperative to develop adaptive defenses against such advanced poisoning attacks. We find that benign clients exhibit significantly higher data distribution stability than malicious clients in federated learning in both CV and NLP tasks. Therefore, the malicious clients can be recognized by observing the stability of their data distribution. In this paper, we propose AdaAggRL, an RL-based Adaptive Aggregation method, to defend against sophisticated poisoning attacks. Specifically, we first utilize distribution learning to simulate the clients' data distributions. Then, we use maximum mean discrepancy (MMD) to calculate the pairwise similarity of the current local model data distribution, its historical data distribution, and global model data distribution. Finally, we use policy learning to adaptively determine the aggregation weights based on the above similarities. Experiments on four real-world datasets demonstrate that the proposed defense model significantly outperforms widely adopted defense models for sophisticated attacks.

Yujing Wang, Hainan Zhang, Liang Pang, Binghui Guo, Hongwei Zheng, Zhiming Zheng

In a real-world RAG system, the current query often involves spoken ellipses and ambiguous references from dialogue contexts, necessitating query rewriting to better describe user's information needs. However, traditional context-based rewriting has minimal enhancement on downstream generation tasks due to the lengthy process from query rewriting to response generation. Some researchers try to utilize reinforcement learning with generation feedback to assist the rewriter, but this sparse rewards provide little guidance in most cases, leading to unstable training and generation results.We find that user's needs are also reflected in the gold documents, retrieved documents and ground-truth. Therefore, by feeding back these multi-aspect dense rewards to query rewriting, more stable and satisfactory responses can be achieved. In this paper, we propose a novel query rewriting method MaFeRw, which improves RAG performance by integrating multi-aspect feedback from both the retrieval process and generated results. Specifically, we first use manual data to train a T5 model for the rewriter initialization. Next, we design three metrics as reinforcement learning feedback: the similarity between the rewritten query and the gold document, the ranking metrics, and ROUGE between the generation and the ground truth. Inspired by RLAIF, we train three kinds of reward models for the above metrics to achieve more efficient training. Finally, we combine the scores of these reward models as feedback, and use PPO algorithm to explore the optimal query rewriting strategy.Experimental results on two conversational RAG datasets demonstrate that MaFeRw achieves superior generation metrics and more stable training compared to baselines.

Yueqian Wang, Xiaojun Meng, Yuxuan Wang, Jianxin Liang, Qun Liu, Dongyan Zhao

Multi-modal multi-party conversation (MMC) is a less studied yet important topic of research due to that it well fits real-world scenarios and thus potentially has more widely-used applications. Compared with the traditional multi-modal conversations, MMC requires stronger character-centered understanding abilities as there are many interlocutors appearing in both the visual and textual context. To facilitate the study of this problem, we present Friends-MMC in this paper, an MMC dataset that contains 24,000+ unique utterances paired with video context. To explore the character-centered understanding of the dialogue, we also annotate the speaker of each utterance, the names and bounding bboxes of faces that appear in the video. Based on this Friends-MMC dataset, we further study two fundamental MMC tasks: conversation speaker identification and conversation response prediction, both of which have the multi-party nature with the video or image as visual context. For conversation speaker identification, we demonstrate the inefficiencies of existing methods such as pre-trained models, and propose a simple yet effective baseline method that leverages an optimization solver to utilize the context of two modalities to achieve better performance. For conversation response prediction, we fine-tune generative dialogue models on Friend-MMC, and analyze the benefits of speaker information. The code and dataset will be publicly available, and thus we call for more attention on modelling speaker information when understanding conversations.

Yudong Wang, Damai Dai, Zhe Yang, Jingyuan Ma, Zhifang Sui

Most work treats large language models as black boxes without an in-depth understanding of their internal working mechanism. To explain the internal representations of LLMs, we utilize a gradient-based metric to assess the activation level of model parameters. Based on this metric, we obtain three preliminary findings. (1) When the inputs are in the same domain, parameters in the shallow layers will be activated densely, which means a larger portion of parameters will have great impacts on the outputs. In contrast, parameters in the deep layers are activated sparsely. (2) When the inputs are across different domains, parameters in shallow layers exhibit higher similarity in the activation behavior than in deep layers. (3) In deep layers, the similarity of the distributions of activated parameters is positively correlated to the empirical data relevance. Further, we develop three validation experiments to solidify these findings. (1) Firstly, starting from the first finding, we attempt to configure different sparsities for different layers and find this method can benefit model pruning. (2) Secondly, we find that a pruned model based on one calibration set can better handle tasks related to the calibration task than those not related, which validates the second finding. (3) Thirdly, Based on the STS-B and SICK benchmarks, we find that two sentences with consistent semantics tend to share similar parameter activation patterns in deep layers, which aligns with our third finding. Our work sheds light on the behavior of parameter activation in LLMs, and we hope these findings will have the potential to inspire more practical applications.

Xingmei Wang, Jinghan Liu, Jiaxiang Meng, Boquan Li, Zijian Liu

Open-set speaker recognition is to identify whether the voices are from the same speaker. One challenge of speaker recognition is collecting large amounts of high-quality data. Based on the promising results of image classification, one intuitively feasible solution is semi-supervised learning (SSL) which uses confidence thresholds to assign pseudo labels for unlabeled data. However, we empirically demonstrated that applying SSL methods to speaker recognition is non-trivial. These methods focus solely on inter-class discrepancy as thresholds to select pseudo labels, overlooking intra-class compactness, which is particularly important for open-set speaker recognition tasks. Motivated by this, we propose Int*-Match, a semi-supervised speaker recognition method selecting reliable pseudo labels with intra-class compactness and inter-class discrepancy for speaker recognition. In particular, we use the inter-class discrepancy of labeled data as the threshold for pseudo-label selection and adjust the threshold based on the intra-class compactness of the pseudo labels dynamically and adaptively. Our systematic experiments demonstrate the superiority of Int*-Match, presenting an outstanding Equal Error Rate (EER) of 1.00% on the VoxCeleb1 original test set, which is merely 0.06% below the performance achieved by fully supervised learning.

Tongguan Wang, Mingmin Wu, Guixin Su, Dongyu Su, Yuxue Hu, Zhongqiang Huang, Ying Sha

The performance of various tasks of natural language processing has greatly improved with the emergence of large language models. However, there is still much room for improvement in understanding certain specific linguistic phenomena, such as Chinese idioms, which are usually composed of four characters. Chinese idioms are difficult to understand due to semantic gaps between their literal and actual meanings. Researchers have proposed the Chinese idiom reading comprehension task to examine the ability of large language models to represent and understand Chinese idioms. The task requires choosing the correct Chinese idiom from a list of candidates to complete the sentence. The current research mainly focuses on text-based idiom comprehension. Nevertheless, there are many idiom application scenarios that combine images and text, and we believe that the corresponding images are beneficial for the model's understanding of the idioms. Therefore, to address the above problems, we first construct a large-scale Multimodal Chinese Idiom Reading Comprehension dataset (MChIRC), which contains a total of 44,433 image-text pairs covering 2,926 idioms. Then, we propose a Dual-Contrastive Idiom Graph Network (DCIGN), which employs a dual-contrastive learning module to align the text and image features corresponding to the same Chinese idiom at both coarse and fine levels, while utilizing a graph structure to capture the semantic relationships between idiom candidates. Finally, we use a cross-attention module to fuse multimodal features with graph features of candidate idioms to predict correct answers. The authoritativeness of MChIRC and the effectiveness of DCIGN are demonstrated through a variety of experiments, which provides a new benchmark for the multimodal Chinese idiom reading comprehension task.

Linyong Wang, Lianwei Wu, Shaoqi Song, Yaxiong Wang, Cuiyun Gao, Kang Wang

Large Language Models (LLMs) have permeated various Natural Language Processing (NLP) tasks. For the summarization tasks, LLMs can generate well-structured rationales, which consist of Essential Aspects (EA), Associated Sentences (AS) and Triple Entity Relations (TER). These rationales guide smaller models (≤1B) to produce better summaries. However, their high deployment costs (≥70B), such as substantial storage space and high computing requirements, limit their utilization in resource-constrained environments. Furthermore, effectively distilling these structured rationales from LLMs into Small Language Models (SLMs) models remains a challenge. To address this, we propose the LLM-based Structured Rationale-guided Multi-view Weak-gated Fusion framework (LSR-MWF). The framework initially employs LLMs to dig structural rationales from a document, considering multiple viewpoints such as EA, AS, and TER. Then, it develop a multi-step summary generation evaluation strategy to select high-quality structured rationales. Subsequently, it aligns with these rationales using additional modules organized in a hierarchical structure. Finally, the framework integrates the features output by these modules with original abstractive model through a weak-gated mechanism. Experimental results on two publicly available CNN/DailyMail and XSum datasets show that our method improves the performance of the abstractive model, outperforming baselines by 11.2％ and 5.8％, respectively. In addition, our method improves the interpretability of summary generation from the viewpoints of EA, AS and TER.

Linqin Wang, Yaping Liu, Zhengtao Yu, Shengxiang Gao, Cunli Mao, Yuxin Huang, Wenjun Wang, Ling Dong

With the rapid advancement of large language models (LLMs), discrete speech representations have become crucial for integrating speech into LLMs. Existing methods for speech representation discretization rely on a predefined codebook size and Euclidean distance-based quantization. However, 1) the size of codebook is a critical parameter that affects both codec performance and downstream task training efficiency. 2) The Euclidean distance-based quantization may lead to audio distortion when the size of the codebook is controlled within a reasonable range. In fact, in the field of information compression, structural information and entropy guidance are crucial, but previous methods have largely overlooked these factors. Therefore, we address the above issues from an information-theoretic perspective, we present SECodec, a novel speech representation codec based on structural entropy (SE) for building speech language models. Specifically, we first model speech as a graph, clustering the speech features nodes within the graph and extracting the corresponding codebook by hierarchically and disentangledly minimizing 2D SE. Then, to address the issue of audio distortion, we propose a new quantization method. This method still adheres to the 2D SE minimization principle, adaptively selecting the most suitable token corresponding to the cluster for each incoming original speech node. Furthermore, we develop a Structural Entropy-based Speech Language Model (SESLM) that leverages SECodec. Experimental results demonstrate that SECodec performs comparably to EnCodec in speech reconstruction, and SESLM surpasses VALL-E in zero-shot text-to-speech tasks.

Jieyi Wang, Yue Huang, Zeming Liu, Dexuan Xu, Chuan Wang, Xiaoming Shi, Ruiyuan Guan, Hongxing Wang, Weihua Yue, Yu Huang

Online psychological counseling dialogue systems are trending, offering a convenient and accessible alternative to traditional in-person therapy. However, existing psychological counseling dialogue systems mainly focus on basic empathetic dialogue or QA with minimal professional knowledge and without goal guidance. In many real-world counseling scenarios, clients often seek multi-type help, such as diagnosis, consultation, therapy, console, and common questions, but existing dialogue systems struggle to combine different dialogue types naturally. In this paper, we identify this challenge as how to construct mixed-type dialogue systems for psychological counseling that enable clients to clarify their goals before proceeding with counseling. To mitigate the challenge, we collect a mixed-type counseling dialogues corpus termed STAMPsy, covering five dialogue types, task-oriented dialogue for diagnosis, knowledge-grounded dialogue, conversational recommendation, empathetic dialogue, and question answering, over 5,000 conversations. Moreover, spatiotemporal-aware knowledge enables systems to have world awareness and has been proven to affect one's mental health. Therefore, we link dialogues in STAMPsy to spatiotemporal state and propose a spatiotemporal-aware mixed-type psychological counseling dataset. Additionally, we build baselines on STAMPsy and develop an iterative self-feedback psychological dialogue generation framework, named Self-STAMPsy. Results indicate that clarifying dialogue goals in advance and utilizing spatiotemporal states are effective.