Shiwan Zhao

Liting Zhang, Shiwan Zhao, Xuyang Zhao et al.

Latent reasoning has emerged as a promising alternative to discrete Chain-of-Thought (CoT) in large language models (LLMs), enabling more expressive reasoning by operating over continuous representations. However, the inherently deterministic nature of continuous representations limits policy exploration in reinforcement learning (RL). To address this, we propose TARPO (Token-Wise Latent-Explicit Reasoning via Action-Routing Policy Optimization), a pure RL framework that adaptively switches between discrete token generation and continuous latent reasoning at each step. TARPO introduces a lightweight action head router that observes the current hidden state and samples a routing decision from a binary mode-selection space, preserving the stochasticity of discrete token sampling from the vocabulary. The LLM backbone and router are jointly optimized end-to-end with a shared group-relative advantage signal. Extensive experiments across Qwen2.5 (from 1.5B to 7B) and Llama-3.1-8B backbones demonstrate that TARPO consistently outperforms existing explicit and latent reasoning RL baselines across diverse benchmarks. Further analysis shows that TARPO learns adaptive token-wise switching behaviors while maintaining stable training dynamics. Our code is available at https://github.com/NKU-LITI/TARPO-master.

Aobo Kong, Shiwan Zhao, Hao Chen et al.

Modern large language models (LLMs) exhibit a remarkable capacity for role-playing, enabling them to embody not only human characters but also non-human entities. This versatility allows them to simulate complex human-like interactions and behaviors within various contexts, as well as to emulate specific objects or systems. While these capabilities have enhanced user engagement and introduced novel modes of interaction, the influence of role-playing on LLMs' reasoning abilities remains underexplored. In this study, we introduce a strategically designed role-play prompting methodology and assess its performance under the zero-shot setting across twelve diverse reasoning benchmarks. Our empirical results illustrate that role-play prompting consistently surpasses the standard zero-shot approach across most datasets. Notably, in experiments conducted using ChatGPT, accuracy on AQuA rises from 53.5% to 63.8%, and on Last Letter from 23.8% to 84.2%.Upon further comparison with the Zero-Shot-CoT technique, which prompts the model to "think step by step", our study demonstrates that role-play prompting acts as a more effective trigger for the CoT process. This highlights its potential to augment the reasoning capabilities of LLMs. We release our code at https://github.com/NKU-HLT/Role-Play-Prompting.

Mengting Hu, Zhen Zhang, Shiwan Zhao et al. · tsinghua

As a main field of artificial intelligence, natural language processing (NLP) has achieved remarkable success via deep neural networks. Plenty of NLP tasks have been addressed in a unified manner, with various tasks being associated with each other through sharing the same paradigm. However, neural networks are black boxes and rely on probability computation. Making mistakes is inevitable. Therefore, estimating the reliability and trustworthiness (in other words, uncertainty) of neural networks becomes a key research direction, which plays a crucial role in reducing models' risks and making better decisions. Therefore, in this survey, we provide a comprehensive review of uncertainty-relevant works in the NLP field. Considering the data and paradigms characteristics, we first categorize the sources of uncertainty in natural language into three types, including input, system, and output. Then, we systemically review uncertainty quantification approaches and the main applications. Finally, we discuss the challenges of uncertainty estimation in NLP and discuss potential future directions, taking into account recent trends in the field. Though there have been a few surveys about uncertainty estimation, our work is the first to review uncertainty from the NLP perspective.

Yike Wu, Yu Zhao, Shiwan Zhao et al.

Despite the great progress of Visual Question Answering (VQA), current VQA models heavily rely on the superficial correlation between the question type and its corresponding frequent answers (i.e., language priors) to make predictions, without really understanding the input. In this work, we define the training instances with the same question type but different answers as \textit{superficially similar instances}, and attribute the language priors to the confusion of VQA model on such instances. To solve this problem, we propose a novel training framework that explicitly encourages the VQA model to distinguish between the superficially similar instances. Specifically, for each training instance, we first construct a set that contains its superficially similar counterparts. Then we exploit the proposed distinguishing module to increase the distance between the instance and its counterparts in the answer space. In this way, the VQA model is forced to further focus on the other parts of the input beyond the question type, which helps to overcome the language priors. Experimental results show that our method achieves the state-of-the-art performance on VQA-CP v2. Codes are available at \href{https://github.com/wyk-nku/Distinguishing-VQA.git}{Distinguishing-VQA}.

Hui Wang, Jinghua Zhao, Yifan Yang et al.

Generative speech technologies are progressing rapidly, but evaluating the perceptual quality of synthetic speech remains a core challenge. Existing methods typically rely on scalar scores or binary decisions, which lack interpretability and generalization across tasks and languages. We present SpeechLLM-as-Judges, a new paradigm for enabling large language models (LLMs) to conduct structured and explanation-based speech quality evaluation. To support this direction, we introduce SpeechEval, a large-scale dataset containing 32,207 multilingual speech clips and 128,754 annotations spanning four tasks: quality assessment, pairwise comparison, improvement suggestion, and deepfake detection. Based on this resource, we develop SQ-LLM, a speech-quality-aware LLM trained with chain-of-thought reasoning and reward optimization to improve capability. Experimental results show that SQ-LLM delivers strong performance across tasks and languages, revealing the potential of this paradigm for advancing speech quality evaluation. The relevant code, models, and data are publicly available at https://github.com/NKU-HLT/SpeechLLM-as-Judges.

Shiyao Wang, Shiwan Zhao, Jiaming Zhou et al.

Dysarthric speech recognition (DSR) presents a formidable challenge due to inherent inter-speaker variability, leading to severe performance degradation when applying DSR models to new dysarthric speakers. Traditional speaker adaptation methodologies typically involve fine-tuning models for each speaker, but this strategy is cost-prohibitive and inconvenient for disabled users, requiring substantial data collection. To address this issue, we introduce a prototype-based approach that markedly improves DSR performance for unseen dysarthric speakers without additional fine-tuning. Our method employs a feature extractor trained with HuBERT to produce per-word prototypes that encapsulate the characteristics of previously unseen speakers. These prototypes serve as the basis for classification. Additionally, we incorporate supervised contrastive learning to refine feature extraction. By enhancing representation quality, we further improve DSR performance, enabling effective personalized DSR. We release our code at https://github.com/NKU-HLT/PB-DSR.

Zhihu Wang, Shiwan Zhao, Yu Wang et al.

The Chain-of-Thought (CoT) paradigm has become a pivotal method for solving complex problems with large language models (LLMs). However, its application to domain-specific tasks remains challenging, as LLMs often fail to decompose tasks accurately or execute subtasks effectively. This paper introduces the Re-TASK framework, a novel theoretical model that revisits LLM tasks from capability, skill, and knowledge perspectives, drawing on the principles of Bloom's Taxonomy and Knowledge Space Theory. While CoT provides a workflow-centric perspective on tasks, Re-TASK introduces a Chain-of-Learning (CoL) paradigm that highlights task dependencies on specific capability items, further broken down into their constituent knowledge and skill components. To address CoT failures, we propose a Re-TASK prompting strategy, which strengthens task-relevant capabilities through targeted knowledge injection and skill adaptation. Experiments across diverse domains demonstrate the effectiveness of Re-TASK. In particular, we achieve improvements of 45.00% on Yi-1.5-9B and 24.50% on Llama3-Chinese-8B for legal tasks. These results highlight the potential of Re-TASK to significantly enhance LLM performance and its applicability in specialized domains. We release our code and data at https://github.com/Uylee/Re-TASK.

Mengting Hu, Yike Wu, Hang Gao et al.

Recently, aspect sentiment quad prediction (ASQP) has become a popular task in the field of aspect-level sentiment analysis. Previous work utilizes a predefined template to paraphrase the original sentence into a structure target sequence, which can be easily decoded as quadruplets of the form (aspect category, aspect term, opinion term, sentiment polarity). The template involves the four elements in a fixed order. However, we observe that this solution contradicts with the order-free property of the ASQP task, since there is no need to fix the template order as long as the quadruplet is extracted correctly. Inspired by the observation, we study the effects of template orders and find that some orders help the generative model achieve better performance. It is hypothesized that different orders provide various views of the quadruplet. Therefore, we propose a simple but effective method to identify the most proper orders, and further combine multiple proper templates as data augmentation to improve the ASQP task. Specifically, we use the pre-trained language model to select the orders with minimal entropy. By fine-tuning the pre-trained language model with these template orders, our approach improves the performance of quad prediction, and outperforms state-of-the-art methods significantly in low-resource settings.

Mengting Hu, Yinhao Bai, Yike Wu et al.

Recently, aspect sentiment quad prediction has received widespread attention in the field of aspect-based sentiment analysis. Existing studies extract quadruplets via pre-trained generative language models to paraphrase the original sentence into a templated target sequence. However, previous works only focus on what to generate but ignore what not to generate. We argue that considering the negative samples also leads to potential benefits. In this work, we propose a template-agnostic method to control the token-level generation, which boosts original learning and reduces mistakes simultaneously. Specifically, we introduce Monte Carlo dropout to understand the built-in uncertainty of pre-trained language models, acquiring the noises and errors. We further propose marginalized unlikelihood learning to suppress the uncertainty-aware mistake tokens. Finally, we introduce minimization entropy to balance the effects of marginalized unlikelihood learning. Extensive experiments on four public datasets demonstrate the effectiveness of our approach on various generation templates.

Jiaming Zhou, Xuxin Cheng, Shiwan Zhao et al.

Autoregressive (AR) large audio language models (LALMs) such as Qwen-2.5-Omni have achieved strong performance on audio understanding and interaction, but scaling them remains costly in data and computation, and strictly sequential decoding limits inference efficiency. Diffusion large language models (dLLMs) have recently been shown to make effective use of limited training data, and prior work on DIFFA indicates that replacing an AR backbone with a diffusion counterpart can substantially improve audio understanding under matched settings, albeit at a proof-of-concept scale without large-scale instruction tuning, preference alignment, or practical decoding schemes. We introduce DIFFA-2, a practical diffusion-based LALM for general audio understanding. DIFFA-2 upgrades the speech encoder, employs dual semantic and acoustic adapters, and is trained with a four-stage curriculum that combines semantic and acoustic alignment, large-scale supervised fine-tuning, and variance-reduced preference optimization, using only fully open-source corpora. Experiments on MMSU, MMAU, and MMAR show that DIFFA-2 consistently improves over DIFFA and is competitive to strong AR LALMs under practical training budgets, supporting diffusion-based modeling is a viable backbone for large-scale audio understanding. Our code is available at https://github.com/NKU-HLT/DIFFA.git.

Jiaming Zhou, Shiwan Zhao, Ning Jiang et al.

End-to-end automatic speech recognition (ASR) usually suffers from performance degradation when applied to a new domain due to domain shift. Unsupervised domain adaptation (UDA) aims to improve the performance on the unlabeled target domain by transferring knowledge from the source to the target domain. To improve transferability, existing UDA approaches mainly focus on matching the distributions of the source and target domains globally and/or locally, while ignoring the model discriminability. In this paper, we propose a novel UDA approach for ASR via inter-domain MAtching and intra-domain DIscrimination (MADI), which improves the model transferability by fine-grained inter-domain matching and discriminability by intra-domain contrastive discrimination simultaneously. Evaluations on the Libri-Adapt dataset demonstrate the effectiveness of our approach. MADI reduces the relative word error rate (WER) on cross-device and cross-environment ASR by 17.7% and 22.8%, respectively.

Aobo Kong, Shiwan Zhao, Hao Chen et al.

Recent advancements in LLMs have showcased their remarkable role-playing capabilities, able to accurately simulate the dialogue styles and cognitive processes of various roles based on different instructions and contexts. Studies indicate that assigning LLMs the roles of experts, a strategy known as role-play prompting, can enhance their performance in the corresponding domains. However, the prompt needs to be manually designed for the given problem, requiring certain expertise and iterative modifications. To this end, we propose self-prompt tuning, making LLMs themselves generate role-play prompts through fine-tuning. Leveraging the LIMA dataset as our foundational corpus, we employ GPT-4 to annotate role-play prompts for each data points, resulting in the creation of the LIMA-Role dataset. We then fine-tune LLMs like Llama-2-7B and Mistral-7B on LIMA-Role. Consequently, the self-prompt tuned LLMs can automatically generate expert role prompts for any given question. We extensively evaluate self-prompt tuned LLMs on widely used NLP benchmarks and open-ended question test. Our empirical results illustrate that self-prompt tuned LLMs outperform standard instruction tuned baselines across most datasets. This highlights the great potential of utilizing fine-tuning to enable LLMs to self-prompt, thereby automating complex prompting strategies. We release the dataset, models, and code at this \href{https://anonymous.4open.science/r/Self-Prompt-Tuning-739E/}{url}.

Haoqin Sun, Shiwan Zhao, Shaokai Li et al.

Multimodal emotion recognition systems rely heavily on the full availability of modalities, suffering significant performance declines when modal data is incomplete. To tackle this issue, we present the Cross-Modal Alignment, Reconstruction, and Refinement (CM-ARR) framework, an innovative approach that sequentially engages in cross-modal alignment, reconstruction, and refinement phases to handle missing modalities and enhance emotion recognition. This framework utilizes unsupervised distribution-based contrastive learning to align heterogeneous modal distributions, reducing discrepancies and modeling semantic uncertainty effectively. The reconstruction phase applies normalizing flow models to transform these aligned distributions and recover missing modalities. The refinement phase employs supervised point-based contrastive learning to disrupt semantic correlations and accentuate emotional traits, thereby enriching the affective content of the reconstructed representations. Extensive experiments on the IEMOCAP and MSP-IMPROV datasets confirm the superior performance of CM-ARR under conditions of both missing and complete modalities. Notably, averaged across six scenarios of missing modalities, CM-ARR achieves absolute improvements of 2.11% in WAR and 2.12% in UAR on the IEMOCAP dataset, and 1.71% and 1.96% in WAR and UAR, respectively, on the MSP-IMPROV dataset.

Yu Wang, Shiwan Zhao, Zhihu Wang et al.

The Chain-of-Thought (CoT) paradigm has emerged as a critical approach for enhancing the reasoning capabilities of large language models (LLMs). However, despite their widespread adoption and success, CoT methods often exhibit instability due to their inability to consistently ensure the quality of generated reasoning paths, leading to sub-optimal reasoning performance. To address this challenge, we propose the \textbf{Strategic Chain-of-Thought} (SCoT), a novel methodology designed to refine LLM performance by integrating strategic knowledge prior to generating intermediate reasoning steps. SCoT employs a two-stage approach within a single prompt: first eliciting an effective problem-solving strategy, which is then used to guide the generation of high-quality CoT paths and final answers. Our experiments across eight challenging reasoning datasets demonstrate significant improvements, including a 21.05\% increase on the GSM8K dataset and 24.13\% on the Tracking\_Objects dataset, respectively, using the Llama3-8b model. Additionally, we extend the SCoT framework to develop a few-shot method with automatically matched demonstrations, yielding even stronger results. These findings underscore the efficacy of SCoT, highlighting its potential to substantially enhance LLM performance in complex reasoning tasks.

Hui Wang, Shiwan Zhao, Jiaming Zhou et al.

Mean Opinion Score (MOS) prediction has made significant progress in specific domains. However, the unstable performance of MOS prediction models across diverse samples presents ongoing challenges in the practical application of these systems. In this paper, we point out that the absence of uncertainty modeling is a significant limitation hindering MOS prediction systems from applying to the real and open world. We analyze the sources of uncertainty in the MOS prediction task and propose to establish an uncertainty-aware MOS prediction system that models aleatory uncertainty and epistemic uncertainty by heteroscedastic regression and Monte Carlo dropout separately. The experimental results show that the system captures uncertainty well and is capable of performing selective prediction and out-of-domain detection. Such capabilities significantly enhance the practical utility of MOS systems in diverse real and open-world environments.

Yang Chen, Yuhang Jia, Shiwan Zhao et al.

As text-based speech editing becomes increasingly prevalent, the demand for unrestricted free-text editing continues to grow. However, existing speech editing techniques encounter significant challenges, particularly in maintaining intelligibility and acoustic consistency when dealing with out-of-domain (OOD) text. In this paper, we introduce, DiffEditor, a novel speech editing model designed to enhance performance in OOD text scenarios through semantic enrichment and acoustic consistency. To improve the intelligibility of the edited speech, we enrich the semantic information of phoneme embeddings by integrating word embeddings extracted from a pretrained language model. Furthermore, we emphasize that interframe smoothing properties are critical for modeling acoustic consistency, and thus we propose a first-order loss function to promote smoother transitions at editing boundaries and enhance the overall fluency of the edited speech. Experimental results demonstrate that our model achieves state-of-the-art performance in both in-domain and OOD text scenarios.

Aobo Kong, Wentao Ma, Shiwan Zhao et al.

Social agents powered by large language models (LLMs) can simulate human social behaviors but fall short in handling complex social dialogues. Direct Preference Optimization (DPO) has proven effective in aligning LLM behavior with human preferences across various agent tasks. However, standard DPO focuses solely on individual turns, which limits its effectiveness in multi-turn social interactions. Several DPO-based multi-turn alignment methods with session-level data have shown potential in addressing this problem.While these methods consider multiple turns across entire sessions, they are often overly coarse-grained, introducing training noise, and lack robust theoretical support. To resolve these limitations, we propose Segment-Level Direct Preference Optimization (SDPO), which dynamically select key segments within interactions to optimize multi-turn agent behavior. SDPO minimizes training noise and is grounded in a rigorous theoretical framework. Evaluations on the SOTOPIA benchmark demonstrate that SDPO-tuned agents consistently outperform both existing DPO-based methods and proprietary LLMs like GPT-4o, underscoring SDPO's potential to advance the social intelligence of LLM-based agents. We release our code and data at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/SDPO.

Yujie Guo, Jiaming Zhou, Yuhang Jia et al.

End-to-end multi-talker automatic speech recognition (MTASR) faces significant challenges in accurately transcribing overlapping speech. A critical bottleneck is that speaker-specific acoustic characteristics, which are essential for distinguishing overlapping speech, are often diluted in deep network layers. To address this, we propose the Global-Local Aware Dynamic Mixture-of-Experts (GLAD) architecture. GLAD introduces a novel routing mechanism that dynamically fuses speaker-aware global context with fine-grained local acoustic details to adaptively guide expert selection. Experiments on the LibriSpeechMix and CH109 datasets demonstrate that GLAD significantly outperforms existing Serialized Output Training (SOT)-based MTASR approaches, exhibiting exceptional robustness in challenging, high-overlap scenarios. To the best of our knowledge, this is the first work to apply a global-local fusion MoE strategy to MTASR.

Liting Zhang, Shiwan Zhao, Aobo Kong et al.

Keyphrase extraction is a fundamental task in natural language processing. However, existing unsupervised prompt-based methods for Large Language Models (LLMs) often rely on single-stage inference pipelines with uniform prompting, regardless of document length or LLM backbone. Such one-size-fits-all designs hinder the full exploitation of LLMs' reasoning and generation capabilities, especially given the complexity of keyphrase extraction across diverse scenarios. To address these challenges, we propose MAPEX, the first framework that introduces multi-agent collaboration into keyphrase extraction. MAPEX coordinates LLM-based agents through modules for expert recruitment, candidate extraction, topic guidance, knowledge augmentation, and post-processing. A dual-path strategy dynamically adapts to document length: knowledge-driven extraction for short texts and topic-guided extraction for long texts. Extensive experiments on six benchmark datasets across three different LLMs demonstrate its strong generalization and universality, outperforming the state-of-the-art unsupervised method by 2.44% and standard LLM baselines by 4.01% in F1@5 on average. Code is available at https://github.com/NKU-LITI/MAPEX.

Shiwan Zhao, Xuyang Zhao, Jiaming Zhou et al.

Supervised fine-tuning (SFT) of large language models can be viewed as an off-policy learning problem, where expert demonstrations come from a fixed behavior policy while training aims to optimize a target policy. Importance sampling is the standard tool for correcting this distribution mismatch, but large policy gaps lead to skewed weights, high variance, and unstable optimization. Existing methods mitigate this issue with KL penalties or clipping, which passively restrict updates rather than actively reducing the gap. We propose a simple yet effective data rewriting framework that proactively shrinks the policy gap before training. For each problem, correct model-generated solutions are kept as on-policy data, while incorrect ones are rewritten through guided re-solving, falling back to expert demonstrations only when needed. This aligns the training distribution with the target policy, reducing variance and improving stability. To handle residual mismatch after rewriting, we additionally apply importance sampling during training, forming a two-stage approach that combines data-level alignment with lightweight optimization-level correction. Experiments on five mathematical reasoning benchmarks show consistent and significant gains over both vanilla SFT and the state-of-the-art Dynamic Fine-Tuning (DFT) approach. Data and code will be released at https://github.com/NKU-HLT/Off-Policy-SFT.

Longxuan Wei, Yubo Zhang, Zijiao Zhang et al.

Large language models achieve strong reasoning performance, yet existing decoding strategies either explore blindly (random sampling) or redundantly (independent multi-sampling). We propose Entropy-Tree, a tree-based decoding method that exploits entropy as a signal for branching decisions--expanding the search tree only at positions where the model exhibits genuine uncertainty. Entropy-Tree shows superior accuracy and calibration in reasoning tasks: it achieves better pass@k than Multi-chain across multiple models and datasets, and its predictive entropy demonstrates better AUROC compared to several traditional metrics. Entropy-Tree unifies efficient structured exploration and reliable uncertainty estimation within a single decoding procedure.

Hui Wang, Shujie Liu, Lingwei Meng et al.

To advance continuous-valued token modeling and temporal-coherence enforcement, we propose FELLE, an autoregressive model that integrates language modeling with token-wise flow matching. By leveraging the autoregressive nature of language models and the generative efficacy of flow matching, FELLE effectively predicts continuous-valued tokens (mel-spectrograms). For each continuous-valued token, FELLE modifies the general prior distribution in flow matching by incorporating information from the previous step, improving coherence and stability. Furthermore, to enhance synthesis quality, FELLE introduces a coarse-to-fine flow-matching mechanism, generating continuous-valued tokens hierarchically, conditioned on the language model's output. Experimental results demonstrate the potential of incorporating flow-matching techniques in autoregressive mel-spectrogram modeling, leading to significant improvements in TTS generation quality, as shown in https://aka.ms/felle.

Xuechen Wang, Shiwan Zhao, Haoqin Sun et al.

Multimodal emotion recognition (MER), leveraging speech and text, has emerged as a pivotal domain within human-computer interaction, demanding sophisticated methods for effective multimodal integration. The challenge of aligning features across these modalities is significant, with most existing approaches adopting a singular alignment strategy. Such a narrow focus not only limits model performance but also fails to address the complexity and ambiguity inherent in emotional expressions. In response, this paper introduces a Multi-Granularity Cross-Modal Alignment (MGCMA) framework, distinguished by its comprehensive approach encompassing distribution-based, instance-based, and token-based alignment modules. This framework enables a multi-level perception of emotional information across modalities. Our experiments on IEMOCAP demonstrate that our proposed method outperforms current state-of-the-art techniques.

Shiwan Zhao, Zhihu Wang, Xuyang Zhao et al.

Post-training has become central to turning pretrained large language models (LLMs) into aligned and deployable systems. Recent progress spans supervised fine-tuning (SFT), preference optimization, reinforcement learning (RL), process supervision, verifier-guided methods, distillation, and multi-stage pipelines. Yet these methods are often discussed in fragmented ways, organized by labels or objective families rather than by the behavioral bottlenecks they address. This survey argues that LLM post-training is best understood as structured intervention on model behavior. We organize the field first by trajectory provenance, which defines two primary learning regimes: off-policy learning on externally supplied trajectories, and on-policy learning on learner-generated rollouts. We then interpret methods through two recurring roles -- effective support expansion, which makes useful behaviors more reachable, and policy reshaping, which improves behavior within already reachable regions -- together with a complementary systems-level role, behavioral consolidation, which preserves, transfers, and amortizes behavior across stages and model transitions. This perspective yields a unified reading of major paradigms. SFT may serve either support expansion or policy reshaping, whereas preference-based methods are usually off-policy reshaping. On-policy RL often improves behavior on learner-generated states, though under stronger guidance it can also make hard-to-reach reasoning paths reachable. Distillation is often best understood as consolidation rather than only compression, and hybrid pipelines emerge as coordinated multi-stage compositions. Overall, the framework helps diagnose post-training bottlenecks and reason about stage composition, suggesting that progress in LLM post-training increasingly depends on coordinated system design rather than any single dominant objective.

Xuyang Zhao, Shiwan Zhao, Hualong Yu et al.

Multi-agent systems (MAS) powered by large language models (LLMs) hold significant promise for solving complex decision-making tasks. However, the core process of collaborative decision-making (CDM) within these systems remains underexplored. Existing approaches often rely on either ``dictatorial" strategies that are vulnerable to the cognitive biases of a single agent, or ``voting-based" methods that fail to fully harness collective intelligence. To address these limitations, we propose \textbf{AgentCDM}, a structured framework for enhancing collaborative decision-making in LLM-based multi-agent systems. Drawing inspiration from the Analysis of Competing Hypotheses (ACH) in cognitive science, AgentCDM introduces a structured reasoning paradigm that systematically mitigates cognitive biases and shifts decision-making from passive answer selection to active hypothesis evaluation and construction. To internalize this reasoning process, we develop a two-stage training paradigm: the first stage uses explicit ACH-inspired scaffolding to guide the model through structured reasoning, while the second stage progressively removes this scaffolding to encourage autonomous generalization. Experiments on multiple benchmark datasets demonstrate that AgentCDM achieves state-of-the-art performance and exhibits strong generalization, validating its effectiveness in improving the quality and robustness of collaborative decisions in MAS.

Yu Wang, Shiwan Zhao, Zhihu Wang et al.

The integration of external knowledge through Retrieval-Augmented Generation (RAG) has become foundational in enhancing large language models (LLMs) for knowledge-intensive tasks. However, existing RAG paradigms often overlook the cognitive step of applying knowledge, leaving a gap between retrieved facts and task-specific reasoning. In this work, we introduce RAG+, a principled and modular extension that explicitly incorporates application-aware reasoning into the RAG pipeline. RAG+ constructs a dual corpus consisting of knowledge and aligned application examples, created either manually or automatically, and retrieves both jointly during inference. This design enables LLMs not only to access relevant information but also to apply it within structured, goal-oriented reasoning processes. Experiments across mathematical, legal, and medical domains, conducted on multiple models, demonstrate that RAG+ consistently outperforms standard RAG variants, achieving average improvements of 3-5%, and peak gains up to 13.5% in complex scenarios. By bridging retrieval with actionable application, RAG+ advances a more cognitively grounded framework for knowledge integration, representing a step toward more interpretable and capable LLMs.

Jiaming Zhou, Yujie Guo, Shiwan Zhao et al.

Code-switching (CS), the alternation between two or more languages within a single conversation, presents significant challenges for automatic speech recognition (ASR) systems. Existing Mandarin-English code-switching datasets often suffer from limitations in size, spontaneity, and the lack of full-length dialogue recordings with transcriptions, hindering the development of robust ASR models for real-world conversational scenarios. This paper introduces CS-Dialogue, a novel large-scale Mandarin-English code-switching speech dataset comprising 104 hours of spontaneous conversations from 200 speakers. Unlike previous datasets, CS-Dialogue provides full-length dialogue recordings with complete transcriptions, capturing naturalistic code-switching patterns in continuous speech. We describe the data collection and annotation processes, present detailed statistics of the dataset, and establish benchmark ASR performance using state-of-the-art models. Our experiments, using Transformer, Conformer, and Branchformer, demonstrate the challenges of code-switching ASR, and show that existing pre-trained models such as Whisper still have the space to improve. The CS-Dialogue dataset will be made freely available for all academic purposes.

Enzhi Wang, Qicheng Li, Shiwan Zhao et al.

In recent years, large language models (LLMs) have achieved remarkable advancements in multimodal processing, including end-to-end speech-based language models that enable natural interactions and perform specific tasks in task-oriented dialogue (TOD) systems. However, existing TOD datasets are predominantly text-based, lacking real speech signals that are essential for evaluating the robustness of speech-based LLMs. Moreover, existing speech TOD datasets are primarily English and lack critical aspects such as speech disfluencies and speaker variations. To address these gaps, we introduce RealTalk-CN, the first Chinese multi-turn, multi-domain speech-text dual-modal TOD dataset, comprising 5.4k dialogues (60K utterances, 150 hours) with paired speech-text annotations. RealTalk-CN captures diverse dialogue scenarios with annotated spontaneous speech disfluencies, ensuring comprehensive coverage of real-world complexities in speech dialogue. In addition, we propose a novel cross-modal chat task that authentically simulates real-world user interactions, allowing dynamic switching between speech and text modalities. Our evaluation covers robustness to speech disfluencies, sensitivity to speaker characteristics, and cross-domain performance. Extensive experiments validate the effectiveness of RealTalk-CN, establishing a strong foundation for Chinese speech-based LLMs research.

Jiaming Zhou, Shiwan Zhao, Hui Wang et al.

The kNN-CTC model has proven to be effective for monolingual automatic speech recognition (ASR). However, its direct application to multilingual scenarios like code-switching, presents challenges. Although there is potential for performance improvement, a kNN-CTC model utilizing a single bilingual datastore can inadvertently introduce undesirable noise from the alternative language. To address this, we propose a novel kNN-CTC-based code-switching ASR (CS-ASR) framework that employs dual monolingual datastores and a gated datastore selection mechanism to reduce noise interference. Our method selects the appropriate datastore for decoding each frame, ensuring the injection of language-specific information into the ASR process. We apply this framework to cutting-edge CTC-based models, developing an advanced CS-ASR system. Extensive experiments demonstrate the remarkable effectiveness of our gated datastore mechanism in enhancing the performance of zero-shot Chinese-English CS-ASR.

Zhen Zhang, Mengting Hu, Shiwan Zhao et al.

Most named entity recognition (NER) systems focus on improving model performance, ignoring the need to quantify model uncertainty, which is critical to the reliability of NER systems in open environments. Evidential deep learning (EDL) has recently been proposed as a promising solution to explicitly model predictive uncertainty for classification tasks. However, directly applying EDL to NER applications faces two challenges, i.e., the problems of sparse entities and OOV/OOD entities in NER tasks. To address these challenges, we propose a trustworthy NER framework named E-NER by introducing two uncertainty-guided loss terms to the conventional EDL, along with a series of uncertainty-guided training strategies. Experiments show that E-NER can be applied to multiple NER paradigms to obtain accurate uncertainty estimation. Furthermore, compared to state-of-the-art baselines, the proposed method achieves a better OOV/OOD detection performance and better generalization ability on OOV entities.

Mengting Hu, Honglei Guo, Shiwan Zhao et al.

A mind-map is a diagram that represents the central concept and key ideas in a hierarchical way. Converting plain text into a mind-map will reveal its key semantic structure and be easier to understand. Given a document, the existing automatic mind-map generation method extracts the relationships of every sentence pair to generate the directed semantic graph for this document. The computation complexity increases exponentially with the length of the document. Moreover, it is difficult to capture the overall semantics. To deal with the above challenges, we propose an efficient mind-map generation network that converts a document into a graph via sequence-to-graph. To guarantee a meaningful mind-map, we design a graph refinement module to adjust the relation graph in a reinforcement learning manner. Extensive experimental results demonstrate that the proposed approach is more effective and efficient than the existing methods. The inference time is reduced by thousands of times compared with the existing methods. The case studies verify that the generated mind-maps better reveal the underlying semantic structures of the document.

Mengting Hu, Shiwan Zhao, Honglei Guo et al.

Aspect category detection (ACD) in sentiment analysis aims to identify the aspect categories mentioned in a sentence. In this paper, we formulate ACD in the few-shot learning scenario. However, existing few-shot learning approaches mainly focus on single-label predictions. These methods can not work well for the ACD task since a sentence may contain multiple aspect categories. Therefore, we propose a multi-label few-shot learning method based on the prototypical network. To alleviate the noise, we design two effective attention mechanisms. The support-set attention aims to extract better prototypes by removing irrelevant aspects. The query-set attention computes multiple prototype-specific representations for each query instance, which are then used to compute accurate distances with the corresponding prototypes. To achieve multi-label inference, we further learn a dynamic threshold per instance by a policy network. Extensive experimental results on three datasets demonstrate that the proposed method significantly outperforms strong baselines.

Yiqin Yu, Xu Min, Shiwan Zhao et al.

In real world applications like healthcare, it is usually difficult to build a machine learning prediction model that works universally well across different institutions. At the same time, the available model is often proprietary, i.e., neither the model parameter nor the data set used for model training is accessible. In consequence, leveraging the knowledge hidden in the available model (aka. the hypothesis) and adapting it to a local data set becomes extremely challenging. Motivated by this situation, in this paper we aim to address such a specific case within the hypothesis transfer learning framework, in which 1) the source hypothesis is a black-box model and 2) the source domain data is unavailable. In particular, we introduce a novel algorithm called dynamic knowledge distillation for hypothesis transfer learning (dkdHTL). In this method, we use knowledge distillation with instance-wise weighting mechanism to adaptively transfer the "dark" knowledge from the source hypothesis to the target domain.The weighting coefficients of the distillation loss and the standard loss are determined by the consistency between the predicted probability of the source hypothesis and the target ground-truth label.Empirical results on both transfer learning benchmark datasets and a healthcare dataset demonstrate the effectiveness of our method.

Zhihang Yuan, Bingzhe Wu, Zheng Liang et al.

Recently, dynamic inference has emerged as a promising way to reduce the computational cost of deep convolutional neural network (CNN). In contrast to static methods (e.g. weight pruning), dynamic inference adaptively adjusts the inference process according to each input sample, which can considerably reduce the computational cost on "easy" samples while maintaining the overall model performance. In this paper, we introduce a general framework, S2DNAS, which can transform various static CNN models to support dynamic inference via neural architecture search. To this end, based on a given CNN model, we first generate a CNN architecture space in which each architecture is a multi-stage CNN generated from the given model using some predefined transformations. Then, we propose a reinforcement learning based approach to automatically search for the optimal CNN architecture in the generated space. At last, with the searched multi-stage network, we can perform dynamic inference by adaptively choosing a stage to evaluate for each sample. Unlike previous works that introduce irregular computations or complex controllers in the inference or re-design a CNN model from scratch, our method can generalize to most of the popular CNN architectures and the searched dynamic network can be directly deployed using existing deep learning frameworks in various hardware devices.

Bingzhe Wu, Chaochao Chen, Shiwan Zhao et al.

Bayesian deep learning is recently regarded as an intrinsic way to characterize the weight uncertainty of deep neural networks~(DNNs). Stochastic Gradient Langevin Dynamics~(SGLD) is an effective method to enable Bayesian deep learning on large-scale datasets. Previous theoretical studies have shown various appealing properties of SGLD, ranging from the convergence properties to the generalization bounds. In this paper, we study the properties of SGLD from a novel perspective of membership privacy protection (i.e., preventing the membership attack). The membership attack, which aims to determine whether a specific sample is used for training a given DNN model, has emerged as a common threat against deep learning algorithms. To this end, we build a theoretical framework to analyze the information leakage (w.r.t. the training dataset) of a model trained using SGLD. Based on this framework, we demonstrate that SGLD can prevent the information leakage of the training dataset to a certain extent. Moreover, our theoretical analysis can be naturally extended to other types of Stochastic Gradient Markov Chain Monte Carlo (SG-MCMC) methods. Empirical results on different datasets and models verify our theoretical findings and suggest that the SGLD algorithm can not only reduce the information leakage but also improve the generalization ability of the DNN models in real-world applications.

Mengting Hu, Shiwan Zhao, Honglei Guo et al.

Aspect-based sentiment analysis (ABSA) is to predict the sentiment polarity towards a particular aspect in a sentence. Recently, this task has been widely addressed by the neural attention mechanism, which computes attention weights to softly select words for generating aspect-specific sentence representations. The attention is expected to concentrate on opinion words for accurate sentiment prediction. However, attention is prone to be distracted by noisy or misleading words, or opinion words from other aspects. In this paper, we propose an alternative hard-selection approach, which determines the start and end positions of the opinion snippet, and selects the words between these two positions for sentiment prediction. Specifically, we learn deep associations between the sentence and aspect, and the long-term dependencies within the sentence by leveraging the pre-trained BERT model. We further detect the opinion snippet by self-critical reinforcement learning. Especially, experimental results demonstrate the effectiveness of our method and prove that our hard-selection approach outperforms soft-selection approaches when handling multi-aspect sentences.

Mengting Hu, Yike Wu, Shiwan Zhao et al.

Cross-domain sentiment classification has drawn much attention in recent years. Most existing approaches focus on learning domain-invariant representations in both the source and target domains, while few of them pay attention to the domain-specific information. Despite the non-transferability of the domain-specific information, simultaneously learning domain-dependent representations can facilitate the learning of domain-invariant representations. In this paper, we focus on aspect-level cross-domain sentiment classification, and propose to distill the domain-invariant sentiment features with the help of an orthogonal domain-dependent task, i.e. aspect detection, which is built on the aspects varying widely in different domains. We conduct extensive experiments on three public datasets and the experimental results demonstrate the effectiveness of our method.

Bingzhe Wu, Shiwan Zhao, ChaoChao Chen et al.

In this paper, we aim to understand the generalization properties of generative adversarial networks (GANs) from a new perspective of privacy protection. Theoretically, we prove that a differentially private learning algorithm used for training the GAN does not overfit to a certain degree, i.e., the generalization gap can be bounded. Moreover, some recent works, such as the Bayesian GAN, can be re-interpreted based on our theoretical insight from privacy protection. Quantitatively, to evaluate the information leakage of well-trained GAN models, we perform various membership attacks on these models. The results show that previous Lipschitz regularization techniques are effective in not only reducing the generalization gap but also alleviating the information leakage of the training dataset.

Yike Wu, Shiwan Zhao, Jia Chen et al.

Improving the captioning performance on low-resource languages by leveraging English caption datasets has received increasing research interest in recent years. Existing works mainly fall into two categories: translation-based and alignment-based approaches. In this paper, we propose to combine the merits of both approaches in one unified architecture. Specifically, we use a pre-trained English caption model to generate high-quality English captions, and then take both the image and generated English captions to generate low-resource language captions. We improve the captioning performance by adding the cycle consistency constraint on the cycle of image regions, English words, and low-resource language words. Moreover, our architecture has a flexible design which enables it to benefit from large monolingual English caption datasets. Experimental results demonstrate that our approach outperforms the state-of-the-art methods on common evaluation metrics. The attention visualization also shows that the proposed approach really improves the fine-grained alignment between words and image regions.

Bingzhe Wu, Shiwan Zhao, Guangyu Sun et al.

Recently, deep convolutional neural networks (CNNs) have achieved great success in pathological image classification. However, due to the limited number of labeled pathological images, there are still two challenges to be addressed: (1) overfitting: the performance of a CNN model is undermined by the overfitting due to its huge amounts of parameters and the insufficiency of labeled training data. (2) privacy leakage: the model trained using a conventional method may involuntarily reveal the private information of the patients in the training dataset. The smaller the dataset, the worse the privacy leakage. To tackle the above two challenges, we introduce a novel stochastic gradient descent (SGD) scheme, named patient privacy preserving SGD (P3SGD), which performs the model update of the SGD in the patient level via a large-step update built upon each patient's data. Specifically, to protect privacy and regularize the CNN model, we propose to inject the well-designed noise into the updates. Moreover, we equip our P3SGD with an elaborated strategy to adaptively control the scale of the injected noise. To validate the effectiveness of P3SGD, we perform extensive experiments on a real-world clinical dataset and quantitatively demonstrate the superior ability of P3SGD in reducing the risk of overfitting. We also provide a rigorous analysis of the privacy cost under differential privacy. Additionally, we find that the models trained with P3SGD are resistant to the model-inversion attack compared with those trained using non-private SGD.

Mengting Hu, Shiwan Zhao, Li Zhang et al.

Aspect level sentiment classification is a fine-grained sentiment analysis task. To detect the sentiment towards a particular aspect in a sentence, previous studies have developed various attention-based methods for generating aspect-specific sentence representations. However, the attention may inherently introduce noise and downgrade the performance. In this paper, we propose constrained attention networks (CAN), a simple yet effective solution, to regularize the attention for multi-aspect sentiment analysis, which alleviates the drawback of the attention mechanism. Specifically, we introduce orthogonal regularization on multiple aspects and sparse regularization on each single aspect. Experimental results on two public datasets demonstrate the effectiveness of our approach. We further extend our approach to multi-task settings and outperform the state-of-the-art methods.

Jing Mei, Shiwan Zhao, Feng Jin et al.

In healthcare, applying deep learning models to electronic health records (EHRs) has drawn considerable attention. EHR data consist of a sequence of medical visits, i.e. a multivariate time series of diagnosis, medications, physical examinations, lab tests, etc. This sequential nature makes EHR well matching the power of Recurrent Neural Network (RNN). In this paper, we propose "Deep Diabetologist" - using RNNs for EHR sequential data modelling, to provide the personalized hyperglycemia medication prediction for diabetic patients. Particularly, we develop a hierarchical RNN to capture the heterogeneous sequential information in the EHR data. Our experimental results demonstrate the improved performance, compared with a baseline classifier using logistic regression. Moreover, hierarchical RNN models outperform basic ones, providing deeper data insights for clinical decision support.

Xiaolu Zhang, Shiwan Zhao, Lingxi Xie

The Wireless Capsule Endoscopy (WCE) is becoming a popular way of screening gastrointestinal system diseases and cancer. However, the time-consuming process in inspecting WCE data limits its applications and increases the cost of examinations. This paper considers WCE-based gastric ulcer detection, in which the major challenge is to detect the lesions in a local region. We propose an approach named infinite curriculum learning, which generalizes curriculum learning to an infinite sampling space by approximately measuring the difficulty of each patch by its scale. This allows us to adapt our model from local patches to global images gradually, leading to a consistent accuracy gain. Experiments are performed on a large dataset with more than 3 million WCE images. Our approach achieves a binary classification accuracy of 87%, and is able to detect some lesions mis-annotated by the physicians. In a real-world application, our approach can reduce the workload of a physician by 90%-98% in gastric ulcer screening.

Bingzhe Wu, Xiaolu Zhang, Shiwan Zhao et al.

Pathological glomerulus classification plays a key role in the diagnosis of nephropathy. As the difference between different subcategories is subtle, doctors often refer to slides from different staining methods to make decisions. However, creating correspondence across various stains is labor-intensive, bringing major difficulties in collecting data and training a vision-based algorithm to assist nephropathy diagnosis. This paper provides an alternative solution for integrating multi-stained visual cues for glomerulus classification. Our approach, named generator-to-classifier (G2C), is a two-stage framework. Given an input image from a specified stain, several generators are first applied to estimate its appearances in other staining methods, and a classifier follows to combine visual cues from different stains for prediction (whether it is pathological, or which type of pathology it has). We optimize these two stages in a joint manner. To provide a reasonable initialization, we pre-train the generators in an unlabeled reference set under an unpaired image-to-image translation task, and then fine-tune them together with the classifier. We conduct experiments on a glomerulus type classification dataset collected by ourselves (there are no publicly available datasets for this purpose). Although joint optimization slightly harms the authenticity of the generated patches, it boosts classification performance, suggesting more effective visual cues are extracted in an automatic way. We also transfer our model to a public dataset for breast cancer classification, and outperform the state-of-the-arts significantly.