Xiaoyu Tan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Xiaoyu Tan has a total of 106 co-authors. Here are the top co-authors by number of shared publications:
- Xihe Qiu
Shanghai University of Engineering Science
42 shared publications - Haoyu Wang
Sino-German College (SGC), University of Shanghai for Science and Technology, Shanghai, China
15 shared publications - Chao Qu
Fudan University, Shanghai, China
6 shared publications - Shaojie Shi
Artificial Intelligence Innovation and Incubation Institute, Fudan University, Shanghai, China
6 shared publications - Yinghui Xu
Fudan University, Shanghai, China
6 shared publications - Leijun Cheng
School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai, China
5 shared publications
Latest Publications
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Fuzzy Systems
- Vol. 34 (11) , 3745-3758
BTBR: A Bayesian-Theory-Driven Probabilistic–Fuzzy Framework for Implicit Bias Removal in Large Language Models
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Fuzzy Systems
- Vol. 34 (11) , 3745-3758
BTBR: A Bayesian-Theory-Driven Probabilistic–Fuzzy Framework for Implicit Bias Removal in Large Language Models
- Open Access
- 19 September 2026
- Published by Elsevier in SSRN Electronic Journal
SRUG: A Fusion-Driven Generator Network for Limited-Data Medical Image Translation
Magnetic Resonance Imaging (MRI) provides detailed tissue information, but its clinical application is often constrained by long acquisition time, cost, and limited availability of complete multi-sequence scans. Image-to-image translation provides a strategy for synthesizing missing MRI sequences from acquired ones. Although generator-based image translation has been widely studied in medical imaging, its architecture design has not been fully explored for limited-data MRI synthesis. In this study, we propose SRUG, a supervised standalone fusion-driven generator network designed to improve structural preservation and image fidelity under limited paired-data conditions. SRUG is trained with paired source-target supervision using reconstruction and structural losses, without a discriminator or adversarial objective. Rather than simply combining existing modules, the proposed generator reorganizes feature encoding, channel-wise reweighting, and dense multi-scale decoding into a unified representation-learning process. This design enables the model to emphasize anatomically relevant features while maintaining effective information flow across spatial scales. Experiments on multiple intra-modality MRI translation tasks show that the proposed method achieves competitive image quality and structural consistency using approximately 300 patient-level paired 2.5D samples, where each sample corresponds to one patient/case. Additional evaluation on IXI demonstrates applicability to another public dataset and modality pair, while zero-shot testing on BraTS 2019 provides preliminary evidence of cross-dataset transferability. These findings indicate that SRUG is a data-efficient supervised generator network for MRI sequence synthesis, while further multi-center validation and clinical assessment remain necessary before practical deployment. The SRUG implementation is publicly available at https://github.com/RisingRich/SRUG.
- 19 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessSRUG: A Fusion-Driven Generator Network for Limited-Data Medical Image Translation
Magnetic Resonance Imaging (MRI) provides detailed tissue information, but its clinical application is often constrained by long acquisition time, cost, and limited availability of complete multi-sequence scans. Image-to-image translation provides a strategy for synthesizing missing MRI sequences from acquired ones. Although generator-based image translation has been widely studied in medical imaging, its architecture design has not been fully explored for limited-data MRI synthesis. In this study, we propose SRUG, a supervised standalone fusion-driven generator network designed to improve structural preservation and image fidelity under limited paired-data conditions. SRUG is trained with paired source-target supervision using reconstruction and structural losses, without a discriminator or adversarial objective. Rather than simply combining existing modules, the proposed generator reorganizes feature encoding, channel-wise reweighting, and dense multi-scale decoding into a unified representation-learning process. This design enables the model to emphasize anatomically relevant features while maintaining effective information flow across spatial scales. Experiments on multiple intra-modality MRI translation tasks show that the proposed method achieves competitive image quality and structural consistency using approximately 300 patient-level paired 2.5D samples, where each sample corresponds to one patient/case. Additional evaluation on IXI demonstrates applicability to another public dataset and modality pair, while zero-shot testing on BraTS 2019 provides preliminary evidence of cross-dataset transferability. These findings indicate that SRUG is a data-efficient supervised generator network for MRI sequence synthesis, while further multi-center validation and clinical assessment remain necessary before practical deployment. The SRUG implementation is publicly available at https://github.com/RisingRich/SRUG.
- 20 August 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 333
A workflow-aligned multimodal framework for interpretable breast cancer diagnosis and molecular subtyping
- 20 August 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 333
A workflow-aligned multimodal framework for interpretable breast cancer diagnosis and molecular subtyping
- 3 July 2026
- Published by Elsevier in Pattern Recognition Letters
- Vol. 207, 227-233
MLLoRA: Leveraging meta-learning and LoRA for efficient multi-task fine-tuning in LLM-driven medical applications
- 3 July 2026
- Published by Elsevier in Pattern Recognition Letters
- Vol. 207, 227-233
MLLoRA: Leveraging meta-learning and LoRA for efficient multi-task fine-tuning in LLM-driven medical applications
- Open Access
- 20 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1378-1389
AURORA: Automated Training Framework of Universal Process Reward Models via Ensemble Prompting and Reverse Verification
- 20 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1378-1389
Open AccessAURORA: Automated Training Framework of Universal Process Reward Models via Ensemble Prompting and Reverse Verification
- Open Access
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 5198-5209
Curiosity Driven Knowledge Retrieval for Mobile Agents
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 5198-5209
Open AccessCuriosity Driven Knowledge Retrieval for Mobile Agents
- 16 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cognitive and Developmental Systems
- Vol. PP (23798920) , 1-11
CLAIM: Mitigating Catastrophic Forgetting in Continual Instruction Fine-tuning Large Language Models
In recent times, Large Language Models (LLMs) have garnered substantial interest from the research community, underpinned by their widespread applicability in numerous language-oriented tasks and their notable performance enhancement through instruction fine-tuning techniques, grounded in human preference. However, enhancing the instruction-following proficiency of LLMs in specific tasks necessitates users undertaking supervised fine-tuning on already fine-tuned LLMs, leading to a substantial loss in generality and inducing catastrophic forgetting. This phenomenon significantly undermines the models’ generalization capacity across diverse tasks and intensifies hallucination in LLMs’ generation. Given the proprietary nature of high-quality instruction-following data, there’s a pressing need for methods that combat both catastrophic forgetting and hallucination without relying on prior fine-tuning datasets. We introduce CLAIM: Continual instruction fine-tuning Large language models with Augmented Instruction self-training and Model averaging, a novel framework that leverages augmented instruction self-training and model averaging to counter these challenges. Empirical experiment results confirm CLAIM’s ability to significantly curtail catastrophic forgetting and hallucinations during continual fine-tuning without accessing previously used instruction data.
- 16 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cognitive and Developmental Systems
- Vol. PP (23798920) , 1-11
CLAIM: Mitigating Catastrophic Forgetting in Continual Instruction Fine-tuning Large Language Models
In recent times, Large Language Models (LLMs) have garnered substantial interest from the research community, underpinned by their widespread applicability in numerous language-oriented tasks and their notable performance enhancement through instruction fine-tuning techniques, grounded in human preference. However, enhancing the instruction-following proficiency of LLMs in specific tasks necessitates users undertaking supervised fine-tuning on already fine-tuned LLMs, leading to a substantial loss in generality and inducing catastrophic forgetting. This phenomenon significantly undermines the models’ generalization capacity across diverse tasks and intensifies hallucination in LLMs’ generation. Given the proprietary nature of high-quality instruction-following data, there’s a pressing need for methods that combat both catastrophic forgetting and hallucination without relying on prior fine-tuning datasets. We introduce CLAIM: Continual instruction fine-tuning Large language models with Augmented Instruction self-training and Model averaging, a novel framework that leverages augmented instruction self-training and model averaging to counter these challenges. Empirical experiment results confirm CLAIM’s ability to significantly curtail catastrophic forgetting and hallucinations during continual fine-tuning without accessing previously used instruction data.
- Open Access
PRISM: Festina Lente Proactivity -- Risk-Sensitive, Uncertainty-Aware Deliberation for Proactive Agents
Proactive agents must decide not only what to say but also whether and when to intervene. Many current systems rely on brittle heuristics or indiscriminate long reasoning, which offers little control over the benefit-burden tradeoff. We formulate the problem as cost-sensitive selective intervention and present PRISM, a novel framework that couples a decision-theoretic gate with a dual-process reasoning architecture. At inference time, the agent intervenes only when a calibrated probability of user acceptance exceeds a threshold derived from asymmetric costs of missed help and false alarms. Inspired by festina lente (Latin: "make haste slowly"), we gate by an acceptance-calibrated, cost-derived threshold and invoke a resource-intensive Slow mode with counterfactual checks only near the decision boundary, concentrating computation on ambiguous and high-stakes cases. Training uses gate-aligned, schema-locked distillation: a teacher running the full PRISM pipeline provides dense, executable supervision on unlabeled interaction traces, while the student learns a response policy that is explicitly decoupled from the intervention gate to enable tunable and auditable control. On ProactiveBench, PRISM reduces false alarms by 22.78% and improves F1 by 20.14% over strong baselines. These results show that principled decision-theoretic gating, paired with selective slow reasoning and aligned distillation, yields proactive agents that are precise, computationally efficient, and controllable. To facilitate reproducibility, we release our code, models, and resources at https://prism-festinalente.github.io/; all experiments use the open-source ProactiveBench benchmark.
Open AccessPRISM: Festina Lente Proactivity -- Risk-Sensitive, Uncertainty-Aware Deliberation for Proactive Agents
Proactive agents must decide not only what to say but also whether and when to intervene. Many current systems rely on brittle heuristics or indiscriminate long reasoning, which offers little control over the benefit-burden tradeoff. We formulate the problem as cost-sensitive selective intervention and present PRISM, a novel framework that couples a decision-theoretic gate with a dual-process reasoning architecture. At inference time, the agent intervenes only when a calibrated probability of user acceptance exceeds a threshold derived from asymmetric costs of missed help and false alarms. Inspired by festina lente (Latin: "make haste slowly"), we gate by an acceptance-calibrated, cost-derived threshold and invoke a resource-intensive Slow mode with counterfactual checks only near the decision boundary, concentrating computation on ambiguous and high-stakes cases. Training uses gate-aligned, schema-locked distillation: a teacher running the full PRISM pipeline provides dense, executable supervision on unlabeled interaction traces, while the student learns a response policy that is explicitly decoupled from the intervention gate to enable tunable and auditable control. On ProactiveBench, PRISM reduces false alarms by 22.78% and improves F1 by 20.14% over strong baselines. These results show that principled decision-theoretic gating, paired with selective slow reasoning and aligned distillation, yields proactive agents that are precise, computationally efficient, and controllable. To facilitate reproducibility, we release our code, models, and resources at https://prism-festinalente.github.io/; all experiments use the open-source ProactiveBench benchmark.
- Open Access
SRU-Pix2Pix: A Fusion-Driven Generator Network for Medical Image Translation with Few-Shot Learning
Magnetic Resonance Imaging (MRI) provides detailed tissue information, but its clinical application is limited by long acquisition time, high cost, and restricted resolution. Image translation has recently gained attention as a strategy to address these limitations. Although Pix2Pix has been widely applied in medical image translation, its potential has not been fully explored. In this study, we propose an enhanced Pix2Pix framework that integrates Squeeze-and-Excitation Residual Networks (SEResNet) and U-Net++ to improve image generation quality and structural fidelity. SEResNet strengthens critical feature representation through channel attention, while U-Net++ enhances multi-scale feature fusion. A simplified PatchGAN discriminator further stabilizes training and refines local anatomical realism. Experimental results demonstrate that under few-shot conditions with fewer than 500 images, the proposed method achieves consistent structural fidelity and superior image quality across multiple intra-modality MRI translation tasks, showing strong generalization ability. These results suggest an effective extension of Pix2Pix for medical image translation.
Open AccessSRU-Pix2Pix: A Fusion-Driven Generator Network for Medical Image Translation with Few-Shot Learning
Magnetic Resonance Imaging (MRI) provides detailed tissue information, but its clinical application is limited by long acquisition time, high cost, and restricted resolution. Image translation has recently gained attention as a strategy to address these limitations. Although Pix2Pix has been widely applied in medical image translation, its potential has not been fully explored. In this study, we propose an enhanced Pix2Pix framework that integrates Squeeze-and-Excitation Residual Networks (SEResNet) and U-Net++ to improve image generation quality and structural fidelity. SEResNet strengthens critical feature representation through channel attention, while U-Net++ enhances multi-scale feature fusion. A simplified PatchGAN discriminator further stabilizes training and refines local anatomical realism. Experimental results demonstrate that under few-shot conditions with fewer than 500 images, the proposed method achieves consistent structural fidelity and superior image quality across multiple intra-modality MRI translation tasks, showing strong generalization ability. These results suggest an effective extension of Pix2Pix for medical image translation.