arXiv subjects

Zhanhui Kang

Publications and source records attributed to Zhanhui Kang.

At least 37 records · Page 2Linked to original sources

arXiv2025

Large Language Model Empowered Recommendation Meets All-domain Continual Pre-Training

Recent research efforts have investigated how to integrate Large Language Models (LLMs) into recommendation, capitalizing on their semantic comprehension and open-world knowledge for user behavior understanding. These approaches predominantly employ supervised fine-tuning on single-domain user interactions to adapt LLMs for specific recommendation tasks. However, they typically encounter dual challenges: the mismatch between general language representations and domain-specific preference patterns, as well as the limited adaptability to multi-domain recommendation scenarios. To bridge these gaps, we introduce CPRec -- an All-domain Continual Pre-Training framework for Recommendation -- designed to holistically align LLMs with universal user behaviors through the continual pre-training paradigm. Specifically, we first design a unified prompt template and organize users' multi-domain behaviors into domain-specific behavioral sequences and all-domain mixed behavioral sequences that emulate real-world user decision logic. To optimize behavioral knowledge infusion, we devise a Warmup-Stable-Annealing learning rate schedule tailored for the continual pre-training paradigm in recommendation to progressively enhance the LLM's capability in knowledge adaptation from open-world knowledge to universal recommendation tasks. To evaluate the effectiveness of our CPRec, we implement it on a large-scale dataset covering seven domains and conduct extensive experiments on five real-world datasets from two distinct platforms. Experimental results confirm that our continual pre-training paradigm significantly mitigates the semantic-behavioral discrepancy and achieves state-of-the-art performance in all recommendation scenarios. The source code will be released upon acceptance.

cs.IR↗

arXiv2025

Improving Multi-modal Recommender Systems by Denoising and Aligning Multi-modal Content and User Feedback

Multi-modal recommender systems (MRSs) are pivotal in diverse online web platforms and have garnered considerable attention in recent years. However, previous studies overlook the challenges of (1) noisy multi-modal content, (2) noisy user feedback, and (3) aligning multi-modal content with user feedback. In order to tackle these challenges, we propose Denoising and Aligning Multi-modal Recommender System (DA-MRS). To mitigate multi-modal noise, DA-MRS first constructs item-item graphs determined by consistent content similarity across modalities. To denoise user feedback, DA-MRS associates the probability of observed feedback with multi-modal content and devises a denoised BPR loss. Furthermore, DA-MRS implements Alignment guided by User preference to enhance task-specific item representation and Alignment guided by graded Item relations to provide finer-grained alignment. Extensive experiments verify that DA-MRS is a plug-and-play framework and achieves significant and consistent improvements across various datasets, backbone models, and noisy scenarios.

cs.IR↗

arXiv2025

Continuous Speech Tokenizer in Text To Speech

The fusion of speech and language in the era of large language models has garnered significant attention. Discrete speech token is often utilized in text-to-speech tasks for speech compression and portability, which is convenient for joint training with text and have good compression efficiency. However, we found that the discrete speech tokenizer still suffers from information loss. Therefore, we propose a simple yet effective continuous speech tokenizer named Cont-SPT, and a text-to-speech model based on continuous speech tokens. Our results show that the speech language model based on the continuous speech tokenizer has better continuity and higher estimated Mean Opinion Scores (MoS). This enhancement is attributed to better information preservation rate of the continuous speech tokenizer across both low and high frequencies in the frequency domain. The code and resources for Cont-SPT can be found in https://github.com/Yixing-Li/Continuous-Speech-Tokenizer

cs.SD↗

arXiv2025

Magnifier Prompt: Tackling Multimodal Hallucination via Extremely Simple Instructions

Hallucinations in multimodal large language models (MLLMs) hinder their practical applications. To address this, we propose a Magnifier Prompt (MagPrompt), a simple yet effective method to tackle hallucinations in MLLMs via extremely simple instructions. MagPrompt is based on the following two key principles, which guide the design of various effective prompts, demonstrating robustness: (1) MLLMs should focus more on the image. (2) When there are conflicts between the image and the model's inner knowledge, MLLMs should prioritize the image. MagPrompt is training-free and can be applied to open-source and closed-source models, such as GPT-4o and Gemini-pro. It performs well across many datasets and its effectiveness is comparable or even better than more complex methods like VCD. Furthermore, our prompt design principles and experimental analyses provide valuable insights into multimodal hallucination.

cs.CL↗

arXiv2025

Language Models "Grok" to Copy

We examine the pre-training dynamics of language models, focusing on their ability to copy text from preceding context--a fundamental skill for various LLM applications, including in-context learning (ICL) and retrieval-augmented generation (RAG). We propose a novel perspective that Transformer-based language models develop copying abilities similarly to grokking, which refers to sudden generalization on test set long after the model fit to the training set. Our experiments yield three arguments: (1) The pre-training loss decreases rapidly, while the context copying ability of models initially lags and then abruptly saturates. (2) The speed of developing copying ability is independent of the number of tokens trained, similarly to how grokking speed is unaffected by dataset size as long as the data distribution is preserved. (3) Induction heads, the attention heads responsible for copying, form from shallow to deep layers during training, mirroring the development of circuits in deeper layers during grokking. We contend that the connection between grokking and context copying can provide valuable insights for more effective language model training, ultimately improving in-context performance. For example, we demonstrated that techniques that enhance grokking, such as regularization, either accelerate or enhance the development of context copying.

cs.CL↗

arXiv2025

More Expressive Attention with Negative Weights

We propose a novel attention mechanism, named Cog Attention, that enables attention weights to be negative for enhanced expressiveness, which stems from two key factors: (1) Cog Attention enhances parameter flexibility. For example, unlike traditional softmax attention heads that use a static output-value (OV) matrix to delete or copy inputs that the heads attend to, Cog Attention naturally learns to use the sign of dynamic query-key (QK) inner products to represent these operations. This enables Cog Attention to perform multiple operations simultaneously within a single head. Meanwhile, Cog Attention's OV matrix can focus more on refinement or modification. (2) Cog Attention enhances the model's robustness against representational collapse by preventing the ``over-squashing'' of earlier tokens into later positions. We develop Transformer-like models which use Cog Attention as attention modules, including decoder-only models at various scales for language modeling and U-ViT diffusion models for image generation. Experiments show that models using Cog Attention exhibit superior performance compared to those employing traditional softmax attention modules. Our approach suggests a promising research direction for rethinking and breaking the entrenched constraints of traditional softmax attention, such as the requirement for non-negative weights.

cs.CL↗

arXiv2024

Mitigating Hallucination in Multimodal Large Language Model via Hallucination-targeted Direct Preference Optimization

Multimodal Large Language Models (MLLMs) are known to hallucinate, which limits their practical applications. Recent works have attempted to apply Direct Preference Optimization (DPO) to enhance the performance of MLLMs, but have shown inconsistent improvements in mitigating hallucinations. To address this issue more effectively, we introduce Hallucination-targeted Direct Preference Optimization (HDPO) to reduce hallucinations in MLLMs. Unlike previous approaches, our method tackles hallucinations from their diverse forms and causes. Specifically, we develop three types of preference pair data targeting the following causes of MLLM hallucinations: (1) insufficient visual capabilities, (2) long context generation, and (3) multimodal conflicts. Experimental results demonstrate that our method achieves superior performance across multiple hallucination evaluation datasets, surpassing most state-of-the-art (SOTA) methods and highlighting the potential of our approach. Ablation studies and in-depth analyses further confirm the effectiveness of our method and suggest the potential for further improvements through scaling up.

cs.CL↗

arXiv2024

Exploring Forgetting in Large Language Model Pre-Training

Catastrophic forgetting remains a formidable obstacle to building an omniscient model in large language models (LLMs). Despite the pioneering research on task-level forgetting in LLM fine-tuning, there is scant focus on forgetting during pre-training. We systematically explored the existence and measurement of forgetting in pre-training, questioning traditional metrics such as perplexity (PPL) and introducing new metrics to better detect entity memory retention. Based on our revised assessment of forgetting metrics, we explored low-cost, straightforward methods to mitigate forgetting during the pre-training phase. Further, we carefully analyzed the learning curves, offering insights into the dynamics of forgetting. Extensive evaluations and analyses on forgetting of pre-training could facilitate future research on LLMs.

cs.CL↗

arXiv2024

Lossless KV Cache Compression to 2%

Large language models have revolutionized data processing in numerous domains, with their ability to handle extended context reasoning receiving notable recognition. To speed up inference, maintaining a key-value (KV) cache memory is essential. Nonetheless, the growing demands for KV cache memory create significant hurdles for efficient implementation. This work introduces a novel architecture, Cross-Layer Latent Attention (CLLA), aimed at compressing the KV cache to less than 2% of its original size while maintaining comparable performance levels. CLLA integrates multiple aspects of KV cache compression, including attention head/dimension reduction, layer sharing, and quantization techniques, into a cohesive framework. Our extensive experiments demonstrate that CLLA achieves lossless performance on most tasks while utilizing minimal KV cache, marking a significant advancement in practical KV cache compression.

cs.CL↗

arXiv2024

RosePO: Aligning LLM-based Recommenders with Human Values

Recently, there has been a growing interest in leveraging Large Language Models (LLMs) for recommendation systems, which usually adapt a pre-trained LLM to the recommendation scenario through supervised fine-tuning (SFT). However, both the pre-training and SFT stages fail to explicitly model the comparative relationships of a user's preferences on different items. To construct a "helpful and harmless" LLM-based recommender, we propose a general framework -- Recommendation with smoothing personalized Preference Optimization (RosePO), which better aligns with customized human values during the post-training stage. Specifically, in addition to the input and chosen response that naturally align with SFT data, we design a rejected sampling strategy tailored for enhancing helpfulness, along with two strategies aimed at mitigating biases to promote harmlessness. To ensure robustness against uncertain labels present in automatically constructed preference data, we introduce a personalized smoothing factor predicted by a preference oracle into the optimization objective. Evaluation on three real-world datasets demonstrates the effectiveness of our method, showcasing not only improved recommendation performance but also mitigation of semantic hallucination and popularity bias.

cs.IR↗

arXiv2024

Towards Empathetic Conversational Recommender Systems

Conversational recommender systems (CRSs) are able to elicit user preferences through multi-turn dialogues. They typically incorporate external knowledge and pre-trained language models to capture the dialogue context. Most CRS approaches, trained on benchmark datasets, assume that the standard items and responses in these benchmarks are optimal. However, they overlook that users may express negative emotions with the standard items and may not feel emotionally engaged by the standard responses. This issue leads to a tendency to replicate the logic of recommenders in the dataset instead of aligning with user needs. To remedy this misalignment, we introduce empathy within a CRS. With empathy we refer to a system's ability to capture and express emotions. We propose an empathetic conversational recommender (ECR) framework. ECR contains two main modules: emotion-aware item recommendation and emotion-aligned response generation. Specifically, we employ user emotions to refine user preference modeling for accurate recommendations. To generate human-like emotional responses, ECR applies retrieval-augmented prompts to fine-tune a pre-trained language model aligning with emotions and mitigating hallucination. To address the challenge of insufficient supervision labels, we enlarge our empathetic data using emotion labels annotated by large language models and emotional reviews collected from external resources. We propose novel evaluation metrics to capture user satisfaction in real-world CRS scenarios. Our experiments on the ReDial dataset validate the efficacy of our framework in enhancing recommendation accuracy and improving user satisfaction.

cs.IR↗

arXiv2024

HMoE: Heterogeneous Mixture of Experts for Language Modeling

Mixture of Experts (MoE) offers remarkable performance and computational efficiency by selectively activating subsets of model parameters. Traditionally, MoE models use homogeneous experts, each with identical capacity. However, varying complexity in input data necessitates experts with diverse capabilities, while homogeneous MoE hinders effective expert specialization and efficient parameter utilization. In this study, we propose a novel Heterogeneous Mixture of Experts (HMoE), where experts differ in size and thus possess diverse capacities. This heterogeneity allows for more specialized experts to handle varying token complexities more effectively. To address the imbalance in expert activation, we propose a novel training objective that encourages the frequent activation of smaller experts, enhancing computational efficiency and parameter utilization. Extensive experiments demonstrate that HMoE achieves lower loss with fewer activated parameters and outperforms conventional homogeneous MoE models on various pre-training evaluation benchmarks. Codes will be released upon acceptance.

cs.CL↗

arXiv2024

The Elephant in the Room: Rethinking the Usage of Pre-trained Language Model in Sequential Recommendation

Sequential recommendation (SR) has seen significant advancements with the help of Pre-trained Language Models (PLMs). Some PLM-based SR models directly use PLM to encode user historical behavior's text sequences to learn user representations, while there is seldom an in-depth exploration of the capability and suitability of PLM in behavior sequence modeling. In this work, we first conduct extensive model analyses between PLMs and PLM-based SR models, discovering great underutilization and parameter redundancy of PLMs in behavior sequence modeling. Inspired by this, we explore different lightweight usages of PLMs in SR, aiming to maximally stimulate the ability of PLMs for SR while satisfying the efficiency and usability demands of practical systems. We discover that adopting behavior-tuned PLMs for item initializations of conventional ID-based SR models is the most economical framework of PLM-based SR, which would not bring in any additional inference cost but could achieve a dramatic performance boost compared with the original version. Extensive experiments on five datasets show that our simple and universal framework leads to significant improvement compared to classical SR and SOTA PLM-based SR models without additional inference costs. Our code can be found in https://github.com/777pomingzi/Rethinking-PLM-in-RS.

cs.IR↗

arXiv2024

Diverse and Fine-Grained Instruction-Following Ability Exploration with Synthetic Data

Instruction-following is particularly crucial for large language models (LLMs) to support diverse user requests. While existing work has made progress in aligning LLMs with human preferences, evaluating their capabilities on instruction following remains a challenge due to complexity and diversity of real-world user instructions. While existing evaluation methods focus on general skills, they suffer from two main shortcomings, i.e., lack of fine-grained task-level evaluation and reliance on singular instruction expression. To address these problems, this paper introduces DINGO, a fine-grained and diverse instruction-following evaluation dataset that has two main advantages: (1) DINGO is based on a manual annotated, fine-grained and multi-level category tree with 130 nodes derived from real-world user requests; (2) DINGO includes diverse instructions, generated by both GPT-4 and human experts. Through extensive experiments, we demonstrate that DINGO can not only provide more challenging and comprehensive evaluation for LLMs, but also provide task-level fine-grained directions to further improve LLMs.

cs.AI↗

arXiv2024

DFGNN: Dual-frequency Graph Neural Network for Sign-aware Feedback

The graph-based recommendation has achieved great success in recent years. However, most existing graph-based recommendations focus on capturing user preference based on positive edges/feedback, while ignoring negative edges/feedback (e.g., dislike, low rating) that widely exist in real-world recommender systems. How to utilize negative feedback in graph-based recommendations still remains underexplored. In this study, we first conducted a comprehensive experimental analysis and found that (1) existing graph neural networks are not well-suited for modeling negative feedback, which acts as a high-frequency signal in a user-item graph. (2) The graph-based recommendation suffers from the representation degeneration problem. Based on the two observations, we propose a novel model that models positive and negative feedback from a frequency filter perspective called Dual-frequency Graph Neural Network for Sign-aware Recommendation (DFGNN). Specifically, in DFGNN, the designed dual-frequency graph filter (DGF) captures both low-frequency and high-frequency signals that contain positive and negative feedback. Furthermore, the proposed signed graph regularization is applied to maintain the user/item embedding uniform in the embedding space to alleviate the representation degeneration problem. Additionally, we conduct extensive experiments on real-world datasets and demonstrate the effectiveness of the proposed model. Codes of our model will be released upon acceptance.

cs.IR↗

arXiv2024

ID-centric Pre-training for Recommendation

Classical sequential recommendation models generally adopt ID embeddings to store knowledge learned from user historical behaviors and represent items. However, these unique IDs are challenging to be transferred to new domains. With the thriving of pre-trained language model (PLM), some pioneer works adopt PLM for pre-trained recommendation, where modality information (e.g., text) is considered universal across domains via PLM. Unfortunately, the behavioral information in ID embeddings is still verified to be dominating in PLM-based recommendation models compared to modality information and thus limits these models' performance. In this work, we propose a novel ID-centric recommendation pre-training paradigm (IDP), which directly transfers informative ID embeddings learned in pre-training domains to item representations in new domains. Specifically, in pre-training stage, besides the ID-based sequential model for recommendation, we also build a Cross-domain ID-matcher (CDIM) learned by both behavioral and modality information. In the tuning stage, modality information of new domain items is regarded as a cross-domain bridge built by CDIM. We first leverage the textual information of downstream domain items to retrieve behaviorally and semantically similar items from pre-training domains using CDIM. Next, these retrieved pre-trained ID embeddings, rather than certain textual embeddings, are directly adopted to generate downstream new items' embeddings. Through extensive experiments on real-world datasets, both in cold and warm settings, we demonstrate that our proposed model significantly outperforms all baselines. Codes will be released upon acceptance.

cs.IR↗

arXiv2024

E-Sparse: Boosting the Large Language Model Inference through Entropy-based N:M Sparsity

Traditional pruning methods are known to be challenging to work in Large Language Models (LLMs) for Generative AI because of their unaffordable training process and large computational demands. For the first time, we introduce the information entropy of hidden state features into a pruning metric design, namely E-Sparse, to improve the accuracy of N:M sparsity on LLM. E-Sparse employs the information richness to leverage the channel importance, and further incorporates several novel techniques to put it into effect: (1) it introduces information entropy to enhance the significance of parameter weights and input feature norms as a novel pruning metric, and performs N:M sparsity without modifying the remaining weights. (2) it designs global naive shuffle and local block shuffle to quickly optimize the information distribution and adequately cope with the impact of N:M sparsity on LLMs' accuracy. E-Sparse is implemented as a Sparse-GEMM on FasterTransformer and runs on NVIDIA Ampere GPUs. Extensive experiments on the LLaMA family and OPT models show that E-Sparse can significantly speed up the model inference over the dense model (up to 1.53X) and obtain significant memory saving (up to 43.52%), with acceptable accuracy loss.

cs.LG↗

arXiv2024

EasyQuant: An Efficient Data-free Quantization Algorithm for LLMs

Large language models (LLMs) have proven to be very superior to conventional methods in various tasks. However, their expensive computations and high memory requirements are prohibitive for deployment. Model quantization is an effective method for reducing this overhead. The problem is that in most previous works, the quantized model was calibrated using few samples from the training data, which might affect the generalization of the quantized LLMs to unknown cases and tasks. Hence in this work, we explore an important question: Can we design a data-independent quantization method for LLMs to guarantee its generalization performance? In this work, we propose EasyQuant, a training-free and data-independent weight-only quantization algorithm for LLMs. Our observation indicates that two factors: outliers in the weight and quantization ranges, are essential for reducing the quantization error. Therefore, in EasyQuant, we leave the outliers (less than 1%) unchanged and optimize the quantization range to reduce the reconstruction error. With these methods, we surprisingly find that EasyQuant achieves comparable performance to the original model. Since EasyQuant does not depend on any training data, the generalization performance of quantized LLMs is safely guaranteed. Moreover, EasyQuant can be implemented in parallel so that the quantized model could be attained in a few minutes even for LLMs over 100B. To our best knowledge, we are the first work that achieves almost lossless quantization performance for LLMs under a data-independent setting and our algorithm runs over 10 times faster than the data-dependent methods.

cs.AI↗