Shaojun Wang

Zhigen Li, Jianxiang Peng, Yanmeng Wang et al.

Dialogue agents powered by Large Language Models (LLMs) show superior performance in various tasks. Despite the better user understanding and human-like responses, their lack of controllability remains a key challenge, often leading to unfocused conversations or task failure. To address this, we introduce Standard Operating Procedure (SOP) to regulate dialogue flow. Specifically, we propose ChatSOP, a novel SOP-guided Monte Carlo Tree Search (MCTS) planning framework designed to enhance the controllability of LLM-driven dialogue agents. To enable this, we curate a dataset comprising SOP-annotated multi-scenario dialogues, generated using a semi-automated role-playing system with GPT-4o and validated through strict manual quality control. Additionally, we propose a novel method that integrates Chain of Thought reasoning with supervised fine-tuning for SOP prediction and utilizes SOP-guided Monte Carlo Tree Search for optimal action planning during dialogues. Experimental results demonstrate the effectiveness of our method, such as achieving a 27.95% improvement in action accuracy compared to baseline models based on GPT-3.5 and also showing notable gains for open-source models. Dataset and codes are publicly available.

Yanmeng Wang, Jun Bai, Ye Wang et al.

Dual-Encoders is a promising mechanism for answer retrieval in question answering (QA) systems. Currently most conventional Dual-Encoders learn the semantic representations of questions and answers merely through matching score. Researchers proposed to introduce the QA interaction features in scoring function but at the cost of low efficiency in inference stage. To keep independent encoding of questions and answers during inference stage, variational auto-encoder is further introduced to reconstruct answers (questions) from question (answer) embeddings as an auxiliary task to enhance QA interaction in representation learning in training stage. However, the needs of text generation and answer retrieval are different, which leads to hardness in training. In this work, we propose a framework to enhance the Dual-Encoders model with question answer cross-embeddings and a novel Geometry Alignment Mechanism (GAM) to align the geometry of embeddings from Dual-Encoders with that from Cross-Encoders. Extensive experimental results show that our framework significantly improves Dual-Encoders model and outperforms the state-of-the-art method on multiple answer retrieval datasets.

Nick J. C. Wang, Zongfeng Quan, Shaojun Wang et al.

The Conformer model is an excellent architecture for speech recognition modeling that effectively utilizes the hybrid losses of connectionist temporal classification (CTC) and attention to train model parameters. To improve the decoding efficiency of Conformer, we propose a novel connectionist temporal summarization (CTS) method that reduces the number of frames required for the attention decoder fed from the acoustic sequences generated by the encoder, thus reducing operations. However, to achieve such decoding improvements, we must fine-tune model parameters, as cross-attention observations are changed and thus require corresponding refinements. Our final experiments show that, with a beamwidth of 4, the LibriSpeech's decoding budget can be reduced by up to 20% and for FluentSpeech data it can be reduced by 11%, without losing ASR accuracy. An improvement in accuracy is even found for the LibriSpeech "test-other" set. The word error rate (WER) is reduced by 6\% relative at the beam width of 1 and by 3% relative at the beam width of 4.

Nick J. C. Wang, Shaojun Wang, Jing Xiao

SLU combines ASR and NLU capabilities to accomplish speech-to-intent understanding. In this paper, we compare different ways to combine ASR and NLU, in particular using a single Conformer model with different ways to use its components, to better understand the strengths and weaknesses of each approach. We find that it is not necessarily a choice between two-stage decoding and end-to-end systems which determines the best system for research or application. System optimization still entails carefully improving the performance of each component. It is difficult to prove that one direction is conclusively better than the other. In this paper, we also propose a novel connectionist temporal summarization (CTS) method to reduce the length of acoustic encoding sequences while improving the accuracy and processing speed of end-to-end models. This method achieves the same intent accuracy as the best two-stage SLU recognition with complicated and time-consuming decoding but does so at lower computational cost. This stacked end-to-end SLU system yields an intent accuracy of 93.97% for the SmartLights far-field set, 95.18% for the close-field set, and 99.71% for FluentSpeech.

Zhuochun Li, Yong Zhang, Ming Li et al.

Large language models (LLMs) are widely used as reference-free evaluators via prompting, but this "LLM-as-a-Judge" paradigm is costly, opaque, and sensitive to prompt design. In this work, we investigate whether smaller models can serve as efficient evaluators by leveraging internal representations instead of surface generation. We uncover a consistent empirical pattern: small LMs, despite with weak generative ability, encode rich evaluative signals in their hidden states. This motivates us to propose the Semantic Capacity Asymmetry Hypothesis: evaluation requires significantly less semantic capacity than generation and can be grounded in intermediate representations, suggesting that evaluation does not necessarily need to rely on large-scale generative models but can instead leverage latent features from smaller ones. Our findings motivate a paradigm shift from LLM-as-a-Judge to Representation-as-a-Judge, a decoding-free evaluation strategy that probes internal model structure rather than relying on prompted output. We instantiate this paradigm through INSPECTOR, a probing-based framework that predicts aspect-level evaluation scores from small model representations. Experiments on reasoning benchmarks (GSM8K, MATH, GPQA) show that INSPECTOR substantially outperforms prompting-based small LMs and closely approximates full LLM judges, while offering a more efficient, reliable, and interpretable alternative for scalable evaluation.

Yong Zhang, Yanwen Huang, Ning Cheng et al.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) with external context, but retrieved passages are often lengthy, noisy, or exceed input limits. Existing compression methods typically require supervised training of dedicated compression models, increasing cost and reducing portability. We propose Sentinel, a lightweight sentence-level compression framework that reframes context filtering as an attention-based understanding task. Rather than training a compression model, Sentinel probes decoder attention from an off-the-shelf 0.5B proxy LLM using a lightweight classifier to identify sentence relevance. Empirically, we find that query-context relevance estimation is consistent across model scales, with 0.5B proxies closely matching the behaviors of larger models. On the LongBench benchmark, Sentinel achieves up to 5$\times$ compression while matching the QA performance of 7B-scale compression systems. Our results suggest that probing native attention signals enables fast, effective, and question-aware context compression. Code available at: https://github.com/yzhangchuck/Sentinel.

Kainan Liu, Yong Zhang, Ning Cheng et al.

Parameter-Efficient Fine-Tuning (PEFT) methods, especially LoRA, are widely used for adapting pre-trained models to downstream tasks due to their computational and storage efficiency. However, in the context of LoRA and its variants, the potential of activation subspaces corresponding to tail eigenvectors remains substantially under-exploited, which may lead to suboptimal fine-tuning performance. In this work, we propose Astra (Activation-Space Tail-Eigenvector Low-Rank Adaptation), a novel PEFT method that leverages the tail eigenvectors of the model output activations-estimated from a small task-specific calibration set-to construct task-adaptive low-rank adapters. By constraining updates to the subspace spanned by these tail eigenvectors, Astra achieves faster convergence and improved downstream performance with a significantly reduced parameter budget. Extensive experiments across natural language understanding (NLU) and natural language generation (NLG) tasks demonstrate that Astra consistently outperforms existing PEFT baselines across 16 benchmarks and even surpasses full fine-tuning (FFT) in certain scenarios.

Zhigen Li, Yanmeng Wang, Rizhao Fan et al.

Paraphrase generation is a longstanding NLP task and achieves great success with the aid of large corpora. However, transferring a paraphrasing model to another domain encounters the problem of domain shifting especially when the data is sparse. At the same time, widely using large pre-trained language models (PLMs) faces the overfitting problem when training on scarce labeled data. To mitigate these two issues, we propose, LAPA, an effective adapter for PLMs optimized by meta-learning. LAPA has three-stage training on three types of related resources to solve this problem: 1. pre-training PLMs on unsupervised corpora, 2. inserting an adapter layer and meta-training on source domain labeled data, and 3. fine-tuning adapters on a small amount of target domain labeled data. This method enables paraphrase generation models to learn basic language knowledge first, then learn the paraphrasing task itself later, and finally adapt to the target task. Our experimental results demonstrate that LAPA achieves state-of-the-art in supervised, unsupervised, and low-resource settings on three benchmark datasets. With only 2\% of trainable parameters and 1\% labeled data of the target task, our approach can achieve a competitive performance with previous work.

Chong Su, Yingbin Fu, Zheyuan Hu et al. · cambridge

We introduce CHOrD, a novel framework for scalable synthesis of 3D indoor scenes, designed to create house-scale, collision-free, and hierarchically structured indoor digital twins. In contrast to existing methods that directly synthesize the scene layout as a scene graph or object list, CHOrD incorporates a 2D image-based intermediate layout representation, enabling effective prevention of collision artifacts by successfully capturing them as out-of-distribution (OOD) scenarios during generation. Furthermore, unlike existing methods, CHOrD is capable of generating scene layouts that adhere to complex floor plans with multi-modal controls, enabling the creation of coherent, house-wide layouts robust to both geometric and semantic variations in room structures. Additionally, we propose a novel dataset with expanded coverage of household items and room configurations, as well as significantly improved data quality. CHOrD demonstrates state-of-the-art performance on both the 3D-FRONT and our proposed datasets, delivering photorealistic, spatially coherent indoor scene synthesis adaptable to arbitrary floor plan variations.

Yanwen Huang, Yong Zhang, Ning Cheng et al.

Large language models (LLMs) often exhibit Context Faithfulness Hallucinations, where outputs deviate from retrieved information due to incomplete context integration. Our analysis reveals a strong correlation between token-level uncertainty and hallucinations. We hypothesize that attention mechanisms inherently encode context utilization signals, supported by probing analysis. Based on these insights, we propose Dynamic Attention-Guided Context Decoding (DAGCD), a lightweight framework that leverages attention distributions and uncertainty signals in a single-pass decoding. Experiments on open-book QA datasets demonstrate DAGCD's effectiveness, yielding significant improvements in faithfulness and robustness while preserving computational efficiency.

Yong Zhang, Bingyuan Zhang, Zhitao Li et al.

The rapid advancement of large language models (LLMs) has significantly enhanced their reasoning abilities, enabling increasingly complex tasks. However, these capabilities often diminish in smaller, more computationally efficient models like GPT-2. Recent research shows that reasoning distillation can help small models acquire reasoning capabilities, but most existing methods focus primarily on improving teacher-generated reasoning paths. Our observations reveal that small models can generate high-quality reasoning paths during sampling, even without chain-of-thought prompting, though these paths are often latent due to their low probability under standard decoding strategies. To address this, we propose Self-Enhanced Reasoning Training (SERT), which activates and leverages latent reasoning capabilities in small models through self-training on filtered, self-generated reasoning paths under zero-shot conditions. Experiments using OpenAI's GPT-3.5 as the teacher model and GPT-2 models as the student models demonstrate that SERT enhances the reasoning abilities of small models, improving their performance in reasoning distillation.

Kainan Liu, Yong Zhang, Ning Cheng et al.

Recent studies have demonstrated that many layers are functionally redundant in large language models (LLMs), enabling model compression by removing these layers to reduce inference cost. While such approaches can improve efficiency, indiscriminate layer pruning often results in significant performance degradation. In this paper, we propose GRASP (Gradient-based Retention of Adaptive Singular Parameters), a novel compression framework that mitigates this issue by preserving sensitivity-aware singular values. Unlike direct layer pruning, GRASP leverages gradient-based attribution on a small calibration dataset to adaptively identify and retain critical singular components. By replacing redundant layers with only a minimal set of parameters, GRASP achieves efficient compression while maintaining strong performance with minimal overhead. Experiments across multiple LLMs show that GRASP consistently outperforms existing compression methods, achieving 90% of the original model's performance under a 20% compression ratio.

Chenfeng Miao, Shuang Liang, Zhencheng Liu et al.

In this work, we address the Text-to-Speech (TTS) task by proposing a non-autoregressive architecture called EfficientTTS. Unlike the dominant non-autoregressive TTS models, which are trained with the need of external aligners, EfficientTTS optimizes all its parameters with a stable, end-to-end training procedure, while allowing for synthesizing high quality speech in a fast and efficient manner. EfficientTTS is motivated by a new monotonic alignment modeling approach (also introduced in this work), which specifies monotonic constraints to the sequence alignment with almost no increase of computation. By combining EfficientTTS with different feed-forward network structures, we develop a family of TTS models, including both text-to-melspectrogram and text-to-waveform networks. We experimentally show that the proposed models significantly outperform counterpart models such as Tacotron 2 and Glow-TTS in terms of speech quality, training efficiency and synthesis speed, while still producing the speeches of strong robustness and great diversity. In addition, we demonstrate that proposed approach can be easily extended to autoregressive models such as Tacotron 2.

Zan Shen, Jiang Qian, Bojin Zhuang et al.

One-Shot methods have evolved into one of the most popular methods in Neural Architecture Search (NAS) due to weight sharing and single training of a supernet. However, existing methods generally suffer from two issues: predetermined number of channels in each layer which is suboptimal; and model averaging effects and poor ranking correlation caused by weight coupling and continuously expanding search space. To explicitly address these issues, in this paper, a Broadening-and-Shrinking One-Shot NAS (BS-NAS) framework is proposed, in which `broadening' refers to broadening the search space with a spring block enabling search for numbers of channels during training of the supernet; while `shrinking' refers to a novel shrinking strategy gradually turning off those underperforming operations. The above innovations broaden the search space for wider representation and then shrink it by gradually removing underperforming operations, followed by an evolutionary algorithm to efficiently search for the optimal architecture. Extensive experiments on ImageNet illustrate the effectiveness of the proposed BS-NAS as well as the state-of-the-art performance.

Liming Deng, Jie Wang, Hangming Liang et al.

Owing to its unique literal and aesthetical characteristics, automatic generation of Chinese poetry is still challenging in Artificial Intelligence, which can hardly be straightforwardly realized by end-to-end methods. In this paper, we propose a novel iterative polishing framework for highly qualified Chinese poetry generation. In the first stage, an encoder-decoder structure is utilized to generate a poem draft. Afterwards, our proposed Quality-Aware Masked Language Model (QAMLM) is employed to polish the draft towards higher quality in terms of linguistics and literalness. Based on a multi-task learning scheme, QA-MLM is able to determine whether polishing is needed based on the poem draft. Furthermore, QAMLM is able to localize improper characters of the poem draft and substitute with newly predicted ones accordingly. Benefited from the masked language model structure, QAMLM incorporates global context information into the polishing process, which can obtain more appropriate polishing results than the unidirectional sequential decoding. Moreover, the iterative polishing process will be terminated automatically when QA-MLM regards the processed poem as a qualified one. Both human and automatic evaluation have been conducted, and the results demonstrate that our approach is effective to improve the performance of encoder-decoder structure.

Tian Xia, Shaodan Zhai, Shaojun Wang

Margin infused relaxed algorithms (MIRAs) dominate model tuning in statistical machine translation in the case of large scale features, but also they are famous for the complexity in implementation. We introduce a new method, which regards an N-best list as a permutation and minimizes the Plackett-Luce loss of ground-truth permutations. Experiments with large-scale features demonstrate that, the new method is more robust than MERT; though it is only matchable with MIRAs, it has a comparatively advantage, easier to implement.

Tian Xia, Shaodan Zhai, Shaojun Wang

List-wise based learning to rank methods are generally supposed to have better performance than point- and pair-wise based. However, in real-world applications, state-of-the-art systems are not from list-wise based camp. In this paper, we propose a new non-linear algorithm in the list-wise based framework called ListMLE, which uses the Plackett-Luce (PL) loss. Our experiments are conducted on the two largest publicly available real-world datasets, Yahoo challenge 2010 and Microsoft 30K. This is the first time in the single model level for a list-wise based system to match or overpass state-of-the-art systems in real-world datasets.

Tian Xia, Shaodan Zhai, Shaojun Wang

In learning to rank area, industry-level applications have been dominated by gradient boosting framework, which fits a tree using least square error principle. While in classification area, another tree fitting principle, weighted least square error, has been widely used, such as LogitBoost and its variants. However, there is a lack of analysis on the relationship between the two principles in the scenario of learning to rank. We propose a new principle named least objective loss based error that enables us to analyze the issue above as well as several important learning to rank models. We also implement two typical and strong systems and conduct our experiments in two real-world datasets. Experimental results show that our proposed method brings moderate improvements over least square error principle.

Wenhao Bian, Jie Wang, Bojin Zhuang et al.

In recent years, deep learning technique has received intense attention owing to its great success in image recognition. A tendency of adaption of deep learning in various information processing fields has formed, including music information retrieval (MIR). In this paper, we conduct a comprehensive study on music audio classification with improved convolutional neural networks (CNNs). To the best of our knowledge, this the first work to apply Densely Connected Convolutional Networks (DenseNet) to music audio tagging, which has been demonstrated to perform better than Residual neural network (ResNet). Additionally, two specific data augmentation approaches of time overlapping and pitch shifting have been proposed to address the deficiency of labelled data in the MIR. Moreover, an ensemble learning of stacking is employed based on SVM. We believe that the proposed combination of strong representation of DenseNet and data augmentation can be adapted to other audio processing tasks.

Xu Lu, Jie Wang, Bojin Zhuang et al.

This paper presents a novel, syllable-structured Chinese lyrics generation model given a piece of original melody. Most previously reported lyrics generation models fail to include the relationship between lyrics and melody. In this work, we propose to interpret lyrics-melody alignments as syllable structural information and use a multi-channel sequence-to-sequence model with considering both phrasal structures and semantics. Two different RNN encoders are applied, one of which is for encoding syllable structures while the other for semantic encoding with contextual sentences or input keywords. Moreover, a large Chinese lyrics corpus for model training is leveraged. With automatic and human evaluations, results demonstrate the effectiveness of our proposed lyrics generation model. To the best of our knowledge, there is few previous reports on lyrics generation considering both music and linguistic perspectives.

Haoshen Fan, Jie Wang, Bojin Zhuang et al.

In this paper, we comprehensively study on context-aware generation of Chinese song lyrics. Conventional text generative models generate a sequence or sentence word by word, failing to consider the contextual relationship between sentences. Taking account into the characteristics of lyrics, a hierarchical attention based Seq2Seq (Sequence-to-Sequence) model is proposed for Chinese lyrics generation. With encoding of word-level and sentence-level contextual information, this model promotes the topic relevance and consistency of generation. A large Chinese lyrics corpus is also leveraged for model training. Eventually, results of automatic and human evaluations demonstrate that our model is able to compose complete Chinese lyrics with one united topic constraint.

Haoshen Fan, Jie Wang, Bojin Zhuang et al.

As one of the quintessence of Chinese traditional culture, couplet compromises two syntactically symmetric clauses equal in length, namely, an antecedent and subsequent clause. Moreover, corresponding characters and phrases at the same position of the two clauses are paired with each other under certain constraints of semantic and/or syntactic relatedness. Automatic couplet generation is recognized as a challenging problem even in the Artificial Intelligence field. In this paper, we comprehensively study on automatic generation of acrostic couplet with the first characters defined by users. The complete couplet generation is mainly divided into three stages, that is, antecedent clause generation pipeline, subsequent clause generation pipeline and clause re-ranker. To realize semantic and/or syntactic relatedness between two clauses, attention-based Sequence-to-Sequence (S2S) neural network is employed. Moreover, to provide diverse couplet candidates for re-ranking, a cluster-based beam search approach is incorporated into the S2S network. Both BLEU metrics and human judgments have demonstrated the effectiveness of our proposed method. Eventually, a mini-program based on this generation system is developed and deployed on Wechat for real users.

Zhongliang Li, Raymond Kulhanek, Shaojun Wang et al.

Recurrent neural language models are the state-of-the-art models for language modeling. When the vocabulary size is large, the space taken to store the model parameters becomes the bottleneck for the use of recurrent neural language models. In this paper, we introduce a simple space compression method that randomly shares the structured parameters at both the input and output embedding layers of the recurrent neural language models to significantly reduce the size of model parameters, but still compactly represent the original input and output embedding layers. The method is easy to implement and tune. Experiments on several data sets show that the new method can get similar perplexity and BLEU score results while only using a very tiny fraction of parameters.

Shaojun Wang, Dale Schuurmans, Fuchun Peng et al.

We present a new statistical learning paradigm for Boltzmann machines based on a new inference principle we have proposed: the latent maximum entropy principle (LME). LME is different both from Jaynes maximum entropy principle and from standard maximum likelihood estimation.We demonstrate the LME principle BY deriving new algorithms for Boltzmann machine parameter estimation, and show how robust and fast new variant of the EM algorithm can be developed.Our experiments show that estimation based on LME generally yields better results than maximum likelihood estimation, particularly when inferring hidden units from small amounts of data.