Research and Technology Forum 2025

Haoxi LI - PhD(CSE) student

Title: Causally Motivated Sycophancy Mitigation for Large Language Models

Author(s): Haoxi Li, Xueyang Tang, Jie ZHANG, Song Guo, Sikai Bai, Peiran Dong, Yue Yu

Abstract:

Incorporating user preferences into large language models (LLMs) can enhance the personalization and reliability of model outputs and facilitate the application of LLMs to real-world scenarios. However, leveraging user preferences can be a double-edged sword. Recent studies have found that improper utilization can incur sycophancy, where LLMs prioritize alignment with user preferences over the correctness of their outputs. To address sycophancy in LLMs, we analyze and model the problem through the lens of structured causal models (SCMs). We attribute sycophancy to LLMs' reliance on spurious correlations between user preferences and model outputs in this paper. Based on the proposed SCMs, we develop a novel framework, termed CAUSM, to mitigate sycophancy in LLMs by exploiting a significant causal signature. Specifically, we eliminate the spurious correlations embedded in the intermediate layers of LLMs through causally motivated head reweighting, and then calibrate the intra-head knowledge along the causal representation direction. Extensive experiments are conducted across diverse language tasks to demonstrate the superiority of our method over state-of-the-art competitors in mitigating sycophancy in LLMs.

Ziqi JIANG - PhD(CSE) student

Title: CLIPDrag: Combining Text-based and Drag-based Instructions for Image Editing

Author(s): Ziqi Jiang, Zhen Wang, Long Chen

Abstract:

Precise and flexible image editing remains a fundamental challenge in computer vision. Based on the modified areas, most editing methods can be divided into two main types: global editing and local editing. In this paper, we choose the two most common editing approaches (ie text-based editing and drag-based editing) and analyze their drawbacks. Specifically, text-based methods often fail to describe the desired modifications precisely, while drag-based methods suffer from ambiguity. To address these issues, we proposed CLIPDrag, a novel image editing method that is the first to combine text and drag signals for precise and ambiguity-free manipulations on diffusion models. To fully leverage these two signals, we treat text signals as global guidance and drag points as local information. Then we introduce a novel global-local motion supervision method to integrate text signals into existing drag-based methods by adapting a pre-trained language-vision model like CLIP. Furthermore, we also address the problem of slow convergence in CLIPDrag by presenting a fast point-tracking method that enforces drag points moving toward correct directions. Extensive experiments demonstrate that CLIPDrag outperforms existing single drag-based methods or text-based methods.

Wei CHEN - PhD(CSE) student

Title: CoMM: A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation

Author(s): Wei Chen, Lin Li, Yongqi Yang, Bin Wen, Fan Yang, Tingting Gao, Yu Wu, Long Chen

Abstract:

Interleaved image-text generation has emerged as a crucial multimodal task, aiming at creating sequences of interleaved visual and textual content given a query. Despite notable advancements in recent multimodal large language models (MLLMs), generating integrated image-text sequences that exhibit narrative coherence and entity and style consistency remains challenging due to poor training data quality. To address this gap, we introduce CoMM, a high-quality Coherent interleaved image-text MultiModal dataset designed to enhance the coherence, consistency, and alignment of generated multimodal content. Initially, CoMM harnesses raw data from diverse sources, focusing on instructional content and visual storytelling, establishing a foundation for coherent and consistent content. To further refine the data quality, we devise a multi-perspective filter strategy that leverages advanced pre-trained models to ensure the development of sentences, consistency of inserted images, and semantic alignment between them. Various quality evaluation metrics are designed to prove the high quality of the filtered dataset. Meanwhile, extensive few-shot experiments on various downstream tasks demonstrate CoMM's effectiveness in significantly enhancing the in-context learning capabilities of MLLMs. Moreover, we propose four new tasks to evaluate MLLMs' interleaved generation abilities, supported by a comprehensive evaluation framework. We believe CoMM opens a new avenue for advanced MLLMs with superior multimodal in-context learning and understanding ability.

Haodong WANG - PhD(CSE) student

Title: D2MoE: Dual Routing and Dynamic Scheduling for Efficient On-Device MoE-based LLM Serving

Author(s): Haodong Wang, Qihua Zhou, Hongzi Cong, Song Guo

Abstract:

The mixture of experts (MoE) model is a sparse variant of large language models (LLMs), designed to hold a better balance between intelligent capability and computational overhead. Despite its benefits, MoE is still too expensive to deploy on resource-constrained edge devices, especially with the demands of on-device inference services. Recent research efforts often apply model compression techniques, such as quantization, pruning and merging, to restrict MoE complexity. Unfortunately, due to their predefined static model optimization strategies, they cannot always achieve the desired quality-overhead trade-off when handling multiple requests, finally degrading the on-device quality of service. These limitations motivate us to propose the D2MoE, an algorithm-system co-design framework that matches diverse task requirements by dynamically allocating the most proper bit-width to each expert. Specifically, inspired by the nested structure of matryoshka dolls, we propose the matryoshka weight quantization (MWQ) to progressively compress expert weights in a bit-nested manner and reduce the required runtime memory. On top of it, we further optimize the I/O-computation pipeline and design a heuristic scheduling algorithm following our hottest-expert-bit-first (HEBF) principle, which maximizes the expert parallelism between I/O and computation queue under constrained memory budgets, thus significantly reducing the idle temporal bubbles waiting for the experts to load. Evaluations on real edge devices show that D2MoE improves the overall inference throughput by up to 1.39X and reduces the peak memory footprint by up to 53% over the latest on-device inference frameworks, while still preserving comparable serving accuracy as its INT8 counterparts.

Kai CHEN - PhD(CSE) student

Title: EMOVA: Empowering Language Models to See, Hear and Speak with Vivid Emotions

Author(s): Kai Chen, Yunhao Gou, Runhui Huang, Zhili Liu, Daxin Tan, Jing Xu, Chunwei Wang, Yi Zhu, Yihan Zeng, Kuo Yang, Dingdong Wang, Kun Xiang, Haoyuan Li, Haoli Bai, Jianhua Han, Xiaohui Li, Weike Jin, Nian Xie, Yu Zhang, James T. Kwok, Hengshuang Zhao, Xiaodan Liang, Dit-Yan Yeung, Xiao Chen, Zhenguo Li, Wei Zhang, Qun Liu, Jun Yao, Lanqing Hong, Lu Hou, Hang Xu

Abstract:

GPT-4o, an omni-modal model that enables vocal conversations with diverse emotions and tones, marks a milestone for omni-modal foundation models. However, empowering Large Language Models to perceive and generate images, texts, and speeches end-to-end with publicly available data remains challenging in the open-source community. Existing vision-language models rely on external tools for the speech processing, while speech-language models still suffer from limited or even without vision-understanding abilities. To address this gap, we propose EMOVA (EMotionally Omni-present Voice Assistant), to enable Large Language Models with end-to-end speech capabilities while maintaining the leading vision-language performance. With a semantic-acoustic disentangled speech tokenizer, we notice surprisingly that omni-modal alignment can further enhance vision-language and speech abilities compared with the corresponding bi-modal aligned counterparts. Moreover, a lightweight style module is proposed for flexible speech style controls (e.g., emotions and pitches). For the first time, EMOVA achieves state-of-the-art performance on both the vision-language and speech benchmarks, and meanwhile, supporting omni-modal spoken dialogue with vivid emotions.

Hei Ching IU - BSc(COSC) student

Title: Exploring Automated Cognitive Tests: Impact of CoT Prompting and Human Skills on LLM Performance

Author(s): IU Hei Ching, ZHANG Li

Abstract:

Conversational agents (CAs) possess significant potential to automate the early screening of neurocognitive disorders (NCDs) and enhance accessibility to these services, enabling timely and critical intervention. However, their effectiveness is highly dependent on the quality of interactions they establish with users. This study investigates prompting strategies aimed at improving the effectiveness of CAs in NCD screening. Specifically, it examines how different prompting structures—from complex prompts incorporating chain-of-thought (CoT) reasoning and human knowledge through scaffolding, to simpler end-to-end approaches—affect the quality and appropriateness of model-generated responses. The study further evaluates and compares the performance of two large language models, GPT- 4o and Qwen, across four distinct levels of prompting complexity. Additionally, it explores the impact of automatic speech recognition (ASR) errors on the models’ ability to accurately interpret user intentions and generate responses aligned with task requirements.

Cheng JIN - PhD(CSE) student

Title: Generalizable Cervical Cancer Screening via Large-scale Pretraining and Test-Time Adaptation

Author(s): Hao Jiang, Cheng Jin, Huangjing Lin, Yanning Zhou, Xi Wang, Jiabo Ma, Li Ding, Jun Hou, Runsheng Liu, Zhizhong Chai, Luyang Luo, Huijuan Shi, Yinling Qian, Qiong Wang, Changzhong Li, Anjia Han, Ronald Cheong Kin Chan, Hao Chen

Abstract:

Cervical cancer is a leading malignancy in female reproductive system. While AI-assisted cytology offers a cost-effective and non-invasive screening solution, current systems struggle with generalizability in complex clinical scenarios. To address this issue, we introduced Smart-CCS, a generalizable Cervical Cancer Screening paradigm based on pretraining and adaptation to create robust and generalizable screening systems. To develop and validate Smart-CCS, we curated a large-scale, multi-center dataset named CCS-127K, which comprises a total of 127,471 cervical cytology whole-slide images collected from 48 medical centers. By leveraging large-scale self-supervised pretraining, our CCS models are equipped with strong generalization capability, potentially generalizing across diverse scenarios. Then, we incorporated test-time adaptation to specifically optimize the trained CCS model for complex clinical settings, which adapts and refines predictions, improving real-world applicability. For evaluation, Smart-CCS was tested in both retrospective and prospective centers, achieving AUCs of 0.965 (retrospective internal centers), 0.950 (retrospective external centers). and up to 0.986 in the prospective centers, with high sensitivity validated by histology. Interpretability analysis confirms model alignment with clinical practice. These results highlight Smart-CCS as a promising solution for enhancing the reliability and scalability of AI-assisted cervical cancer screening in real-world clinical settings.

Kangyu YUAN - PhD(CSE) student

Title: "I Love the Internet Again": Exploring the Interaction Inception of "TikTok Refugees" Flocking into RedNote

Author(s): Kangyu Yuan, Li Zhang, Hanfang Lyu, Ziqi Pan, Yuanhao Zhang, Junze Li, Bingcan Guo, Jiaxiong Hu, Qingyu Guo, Xiaojuan Ma

Abstract:

The U.S. government's announcement to ban TikTok in January 2025 led an influx of TikTok users to migrate to RedNote, a Chinese user dominant social media platform. These migrants, self-identified as "TikTok Refugees," engage directly with Chinese natives, overcoming cross-cultural communication barriers. This platform migration provides a unique opportunity to examine the communication behaviors and strategies employed when users from the Western culture integrate into communities primarily composed of users from the Eastern culture. In this study, we analyze 3,510 RedNote posts (mainly posted from America (47%) and China (39.3%)) to characterize how TikTok migrants blend into the platform. Through open coding, we identified 12 cross-cultural communication strategies manifested in newcomers' posts. We also discuss the critical role of friendly content quality in facilitating successful interactions and the challenges of maintaining sustainable engagement. Finally, we offer design implications for social media platforms to enhance cross-cultural communication.

Yanghao WANG - PhD(CSE) student

Title: Inversion Circle Interpolation: Diffusion-based Image Augmentation for Data-scarce Classification

Author(s): Yanghao Wang, Long Chen

Abstract:

Data Augmentation (DA), i.e., synthesizing faithful and diverse samples to expand the original training set, is a prevalent and effective strategy to improve the performance of various data-scarce tasks. With the powerful image generation ability, diffusion-based DA has shown strong performance gains on different image classification benchmarks. In this paper, we analyze today’s diffusion-based DA methods, and argue that they cannot take account of both faithfulness and diversity, which are two critical keys for generating high quality samples and boosting classification performance. To this end, we propose a novel Diffusion-based DA method: Diff-II. Specifically, it consists of three steps: 1) Category concepts learning: Learning concept embeddings for each category. 2) Inversion interpolation: Calculating the inversion for each image, and conducting circle interpolation for two randomly sampled inversions from the same category. 3) Two-stage denoising: Using different prompts to generate synthesized images in a coarse-to-fine manner. Extensive experiments on various data-scarce image classification tasks (e.g., few-shot, long-tailed, and outof-distribution classification) have demonstrated its effectiveness over state-of-the-art diffusion-based DA methods.

Yingxue XU - PhD(CSE) student

Title: Large Foundation Models Empower Full-stack Solutions for Digital and Intelligent Pathology

Author(s): Xu Yingxue, Ma Jiabo, Wang Yihui

Abstract:

Pathological diagnosis in clinical practice heavily relies on scrutinizing slides manually, which is time-consuming and prone to subjective variability. Current digital pathology systems typically employ task-specific models (e.g., for metastasis detection or survival analysis) trained independently. However, their performance and generalizability are limited by the scarcity of annotated data for individual tasks. Therefore, we propose to leverage large pathology foundation models pretrained on over 190 million pathological images to empower oncological tasks, delivering the full-stack solution to whole-slide image (WSI) analysis for clinical decision support. The system, SmartPath, is developed to unify diverse oncological tasks into a single analysis system across lesion segmentation, auxiliary diagnosis, treatment response prediction, prognostic risk stratification and biomarker detection, while providing real-time interactions with pathology copilot and automating structured report generation. Currently, three major cancers are supported: lung, breast, and gastrointestinal cancers, covering nearly 100 oncological tasks. The system streamlines pathology workflows by automating multi-step analytical processes and standardizing reporting workflow, demonstrating scalable integration into clinical environments. This holds promise in addressing the critical need for full-stack AI solutions in digital and intelligent pathology, moving forward to the comprehensive demands of modern precision oncology.

Qianli LIU - PhD(CSE) student

Title: MELL: Memory-Efficient Large Language Model Serving via Multi-GPU KV Cache Managemen

Author(s): Qianli Liu, Zicong Hong, Peng Li, Fahao Chen and Song Guo

Abstract:

Serving large language models (LLMs) for massive users is challenged by the significant memory footprint of the transient state, known as the key-value (KV) cache, which scales with sequence length and number of requests. Instead of renting or buying more expensive GPUs, the load imbalance of the KV cache across GPUs, coupled with recent advances in inter-GPU communication, provides an opportunity to serve more requests via request migration. However, high migration overhead and unpredictable request patterns make it challenging. Therefore, this paper proposes MELL, a memory-efficient LLM serving system via multi-GPU KV cache management. It saves the number of GPUs needed in the system by considering the dynamic KV cache load and the costly request migration. Specifically, we first develop an adaptive request migration mechanism to balance the computational and communication overheads and adapt to diverse resource conditions. Then, we design an online algorithm tailored to a multi-LLM request and multi-GPU scheduling problem with migration enabled. It aims to minimise the required GPUs while limiting the number of migrations. Finally, we implement a prototype of MELL and demonstrate that it reduces the number of GPUs by 31% and increases the GPU utilization by 43% at most compared to existing LLM serving systems.

Kashun SHUM - PhD(CSE) student

Title: Predictive Data Selection: The Data That Predicts Is the Data That Teaches

Author(s): Kashun Shum, Yuzhen Huang, Hongjian Zou, Qi Ding, Yixuan Liao, Xiaoxin Chen, Qian Liu, Junxian He

Abstract:

Language model pretraining involves training on extensive corpora, where data quality plays a pivotal role. In this work, we aim to directly estimate the contribution of data during pretraining and select pretraining data in an efficient manner. Specifically, we draw inspiration from recent findings showing that compression efficiency (i.e., the normalized loss) of diverse models on certain text correlates strongly with their downstream performance, when the text domain aligns with the downstream benchmarks(Huang et al., 2024). Building on this observation, we hypothesize that data on which model losses are predictive of downstream abilities also contribute effectively to learning. To leverage this insight, we introduce predictive data selection (PreSelect), a lightweight and efficient data selection method that requires training and deploying only a fastText-based scorer. Through comprehensive experiments with 1B and 3B parameter models, we demonstrate that models trained on 30B tokens selected with PreSelect surpass the performance of the vanilla baseline trained on 300B tokens, achieving a 10x reduction in compute requirements. Furthermore, PreSelect significantly outperforms other competitive data selection baselines, such as DCLM and FineWeb-Edu on a scale of 3B models trained on 100B tokens. We open-source our trained data selection scorer along with the curated datasets at this https URL.

Juyoung BAE - PhD(CSE) student

Title: Revolutionizing Dental Prosthetics: Generative AI in Biomimetic Design

Author(s): Juyoung Bae, Moo-hyun Son, Hao Chen

Abstract:

Partial edentulism significantly impacts oral functionality and aesthetics. Dental implant prostheses are the standard intervention, restoring function and appearance. Accurate prosthesis design is crucial, requiring factors such as occlusal alignment and adjacent teeth condition. This demands high level of expertise and a comprehensive understanding of each patient’s oral anatomy, leading to increased time and costs for customized treatments. Recent deep learning advancements show potential in generating missing teeth from intraoral scan (IOS) data. However, existing models are limited to generating a single tooth type, lack flexibility, and clinical applicability, and fail to thoroughly evaluate the usability of generated prosthesis designs. In this work, we introduce a novel generative AI model that adaptively creates one or multiple implant crowns based on patient-specific intraoral scan data, with dramatically reduced cost and manual effort. Our method incorporates user-driven control over crown size and placement, allowing practitioners to override and exclude compromised neighboring teeth where necessary. Through extensive expert evaluations and quantitative inter-rater analyses, we demonstrate that our solution achieves clinically viable accuracy, anatomical realism, and time-efficient workflows, advancing the state-of-the-art in AI-powered dental prosthesis design.

Haoyue ZHANG - PhD(CSE) student

Title: ToFe: Lagged Token Freezing and Reusing for Efficient Transformer Inference

Author(s): Haoyue Zhang, Jie Zhang, Song Guo

Abstract:

Token reduction, which discards less important tokens during forward propagation, has been proposed to enhance the efficiency of transformer-based model inference. However, existing methods handle unimportant tokens irreversibly, preventing their reuse in subsequent blocks. Considering that transformers focus on different information between blocks, tokens reduced in early blocks might be useful later. Furthermore, to adapt transformer models for resource-constrained devices, it is crucial to strike a balance between model performance and computational overhead. To address these challenges, in this paper, we introduce a novel Token Freezing and Reusing (ToFe) framework, where we identify important tokens at each stage and temporarily freeze the unimportant ones, allowing their lagged reuse at a later stage. Specifically, we design a prediction module for token identification and an approximate module for recovery of the frozen tokens. By jointly optimizing with the backbone through computation budget-aware end-to-end training, ToFe can adaptively process the necessary tokens at each block, thereby reducing computational cost while maintaining performance. Extensive experiments demonstrate that ToFe reduces the computational cost of LV-ViT model by 50% with less than 2% drop in Top-1 accuracy, achieving a better trade-off between performance and complexity compared to state-of-the-art methods.

Bingnan CHEN - PhD(CSE) student

Title: Yannakakis+: Practical Acyclic Query Evaluation with Theoretical Guarantees

Author(s): Qichen Wang, Bingnan Chen, Binyang Dai, Ke Yi, Feifei Li, Liang Lin

Abstract:

Acyclic conjunctive queries form the backbone of most analytical workloads, and have been extensively studied in the literature from both theoretical and practical angles. However, there is still a large divide between theory and practice. While the 40-year-old Yannakakis algorithm has strong theoretical running time guarantees, it has not been adopted in real systems due to its high hidden constant factor. In this paper, we strive to close this gap by proposing Yannakakis+, an improved version of the Yannakakis algorithm, which is more practically efficient while preserving its theoretical guarantees. Our experiments demonstrate that Yannakakis+ consistently outperforms the original Yannakakis algorithm by 2x to 5x across a wide range of queries and datasets.

Another nice feature of our new algorithm is that it generates a traditional DAG query plan consisting of standard relational operators, allowing Yannakakis+ to be easily plugged into any standard SQL engine. Our system prototype currently supports four different SQL engines (DuckDB, PostgreSQL, SparkSQL, and AnalyticDB from Alibaba Cloud), and our experiments show that Yannakakis+ is able to deliver better performance than their native query plans on 160 out of the 162 queries tested, with an average speedup of 2.41x and a maximum speedup of 47,059x.