Qiankun Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Qiankun Li has a total of 187 co-authors. Here are the top co-authors by number of shared publications:
- Junxin Chen
Dalian University of Technology, Dalian, China
9 shared publications - Huabao Chen
University of Science and Technology of China, China, China
7 shared publications - Xiaolong Huang
School of Artificial Intelligence, Chongqing University of Technology, Chongqing, China
7 shared publications - Qiupu Chen
Henan University
6 shared publications - Xin Ning
Institute of Semiconductors, Chinese Academy of Sciences, Beijing, China
6 shared publications - Yani Zhang
University of Science and Technology of China, China, China
6 shared publications
Latest Publications
- 19 November 2026
- Published by American Association for Cancer Research (AACR) in Cancer Research
Pretreatment Histopathology-Based Tumor Microenvironment Modeling Predicts Treatment-Specific Objective Responses in Non-small Cell Lung Cancer
Accurate prediction of treatment response remains challenging in cancer. To facilitate response prediction in non-small cell lung cancer (NSCLC), we developed PRISM, an interpretable multiscale graph-learning framework for estimating treatment-specific objective response from pretreatment H&E whole-slide images. PRISM integrated multiresolution tissue representations with a cell-level tumor microenvironment (TME) hypergraph, thereby treating higher-order microenvironmental organization as a learnable component of treatment-response prediction. Samples from 7,082 patients from four Chinese centers were screened. PRISM was evaluated across treatment specific cohorts receiving targeted therapy (T), EGFR-targeted therapy (T_EGFR), immune checkpoint inhibitor–based therapy (ICI), chemoimmunotherapy (CI), and chemotherapy (C) using patient-level internal cross-validation, independent multicenter external validation, and prospective silent validation. PRISM predicted early objective responses across prespecified systemic therapies in NSCLC. In comparison to 18 established multiple-instance learning and computational pathology baseline models, PRISM remained competitive, with numerical advantages in selected tasks. Multiscale attribution and exploratory H&E-derived TME analyses provided spatial context for model predictions. Overall, PRISM offers a morphology-based framework for estimating early response to prespecified systemic therapies and warrants further prospective multicenter evaluation.
- 19 November 2026
- Published by American Association for Cancer Research (AACR) in Cancer Research
Pretreatment Histopathology-Based Tumor Microenvironment Modeling Predicts Treatment-Specific Objective Responses in Non-small Cell Lung Cancer
Accurate prediction of treatment response remains challenging in cancer. To facilitate response prediction in non-small cell lung cancer (NSCLC), we developed PRISM, an interpretable multiscale graph-learning framework for estimating treatment-specific objective response from pretreatment H&E whole-slide images. PRISM integrated multiresolution tissue representations with a cell-level tumor microenvironment (TME) hypergraph, thereby treating higher-order microenvironmental organization as a learnable component of treatment-response prediction. Samples from 7,082 patients from four Chinese centers were screened. PRISM was evaluated across treatment specific cohorts receiving targeted therapy (T), EGFR-targeted therapy (T_EGFR), immune checkpoint inhibitor–based therapy (ICI), chemoimmunotherapy (CI), and chemotherapy (C) using patient-level internal cross-validation, independent multicenter external validation, and prospective silent validation. PRISM predicted early objective responses across prespecified systemic therapies in NSCLC. In comparison to 18 established multiple-instance learning and computational pathology baseline models, PRISM remained competitive, with numerical advantages in selected tasks. Multiscale attribution and exploratory H&E-derived TME analyses provided spatial context for model predictions. Overall, PRISM offers a morphology-based framework for estimating early response to prespecified systemic therapies and warrants further prospective multicenter evaluation.
- 15 November 2026
- Published by Springer Nature in International Journal of Computer Vision
- Vol. 134 (12)
CoFi-M$^{2}$AR: A Coarse-to-Fine Multimodal Framework for Micro-Action Recognition- 15 November 2026
- Published by Springer Nature in International Journal of Computer Vision
- Vol. 134 (12)
CoFi-M$^{2}$AR: A Coarse-to-Fine Multimodal Framework for Micro-Action Recognition- 5 October 2026
- Published by Wiley in Expert Systems
- Vol. 43 (11) , e70446
From Static to Adaptive Inference: A Survey on Test‐Time Learning
Modern artificial intelligence (AI) systems are usually deployed under a static inference paradigm in which, after training, a model processes test inputs through fixed forward computation. This assumption becomes limiting when models encounter distribution shifts, long contexts, interactive sessions, user‐specific patterns, or changing environments. Test‐Time Learning (TTL) offers a complementary perspective by allowing a deployed model to use information available during inference to adapt parameters, statistics, representations, states, memory, prompts, or context before or during prediction. This survey provides an introductory and unified account of TTL, beginning with the original Test‐Time Training (TTT) formulation, incorporating Test‐Time Adaptation (TTA), and broadening it into a general paradigm of adaptive inference. Existing methods are organized according to the adaptation target, representative applications are reviewed in computer vision, natural language processing and large language models (LLMs), speech and signal processing, medical and scientific AI, and robotics, and key challenges are discussed in efficiency, stability, objective alignment, safety, privacy, and evaluation standards. By connecting early self‐supervised test‐time updates with recent state‐ and memory‐oriented architectures such as TTT layers and Titans, this survey highlights TTL as a step towards adaptive foundation models whose inference process can continue to learn under controlled deployment conditions, practical constraints, and governance requirements.
- 5 October 2026
- Published by Wiley in Expert Systems
- Vol. 43 (11) , e70446
From Static to Adaptive Inference: A Survey on Test‐Time Learning
Modern artificial intelligence (AI) systems are usually deployed under a static inference paradigm in which, after training, a model processes test inputs through fixed forward computation. This assumption becomes limiting when models encounter distribution shifts, long contexts, interactive sessions, user‐specific patterns, or changing environments. Test‐Time Learning (TTL) offers a complementary perspective by allowing a deployed model to use information available during inference to adapt parameters, statistics, representations, states, memory, prompts, or context before or during prediction. This survey provides an introductory and unified account of TTL, beginning with the original Test‐Time Training (TTT) formulation, incorporating Test‐Time Adaptation (TTA), and broadening it into a general paradigm of adaptive inference. Existing methods are organized according to the adaptation target, representative applications are reviewed in computer vision, natural language processing and large language models (LLMs), speech and signal processing, medical and scientific AI, and robotics, and key challenges are discussed in efficiency, stability, objective alignment, safety, privacy, and evaluation standards. By connecting early self‐supervised test‐time updates with recent state‐ and memory‐oriented architectures such as TTT layers and Titans, this survey highlights TTL as a step towards adaptive foundation models whose inference process can continue to learn under controlled deployment conditions, practical constraints, and governance requirements.
- Open Access
- 30 September 2026
- Published by MDPI AG in Preprints.org
Recursive Self-Improvement in AI: A Survey
Recursive self-improvement (RSI) seeks to turn gains in an AI system into a stronger capacity to produce subsequent gains. Foundation models make parts of this process practical through generated training data, program revision and automated experimentation, but repeated optimization alone does not establish recursive improvement. This survey organizes the field around two overlapping goals: automated design of AI infrastructure, and repair and optimization of existing systems. We relate these goals to AutoML, meta-learning and program search, and formulate their objectives alongside criteria for task gains, retained capabilities and improved update procedures. A common account of update targets, feedback, inheritance and external control connects self-training, memory and skills, agent design, editable improvers and automated research. We synthesize methods by shared mechanisms and compare their evidence under independent testing, resource controls, transfer probes and multi-round evaluation. Existing studies show that systems can retain useful changes and, in limited settings, improve how they produce subsequent updates. They do not yet establish unrestricted or reliably accelerating self-improvement. We identify feedback quality, interacting updates and independent validation as central constraints, and develop testable directions for more reliable, transferable improvement loops.
- 30 September 2026
- Published by MDPI AG in Preprints.org
Open AccessRecursive Self-Improvement in AI: A Survey
Recursive self-improvement (RSI) seeks to turn gains in an AI system into a stronger capacity to produce subsequent gains. Foundation models make parts of this process practical through generated training data, program revision and automated experimentation, but repeated optimization alone does not establish recursive improvement. This survey organizes the field around two overlapping goals: automated design of AI infrastructure, and repair and optimization of existing systems. We relate these goals to AutoML, meta-learning and program search, and formulate their objectives alongside criteria for task gains, retained capabilities and improved update procedures. A common account of update targets, feedback, inheritance and external control connects self-training, memory and skills, agent design, editable improvers and automated research. We synthesize methods by shared mechanisms and compare their evidence under independent testing, resource controls, transfer probes and multi-round evaluation. Existing studies show that systems can retain useful changes and, in limited settings, improve how they produce subsequent updates. They do not yet establish unrestricted or reliably accelerating self-improvement. We identify feedback quality, interacting updates and independent validation as central constraints, and develop testable directions for more reliable, transferable improvement loops.
- 14 September 2026
- Published by Springer Nature
- p. 588-606
Reasoning with Memory: A Temporal Granularity-Adaptive Framework for Training-Free Long Video Understanding
- 14 September 2026
- Published by Springer Nature
- p. 588-606
Reasoning with Memory: A Temporal Granularity-Adaptive Framework for Training-Free Long Video Understanding
- 11 September 2026
- Published by Springer Nature
- p. 621-639
HIVE: Understanding Post-hallucination Reasoning in Vision Language Models
- 11 September 2026
- Published by Springer Nature
- p. 621-639
HIVE: Understanding Post-hallucination Reasoning in Vision Language Models
- Open Access
- 11 September 2026
- Published by MDPI AG in Preprints.org
Trustworthy Medical Multimodal Large Language Models: Taxonomy, Evaluation, and Benchmarks
The integration of Multimodal Large Language Models (MLLMs) into healthcare has the potential to drive a considerable advancement toward many AI use cases in medicine, offering transformative capabilities across hierarchical levels of clinical granularity, ranging from microscopic tissue analysis over organ level imaging and individual patient modeling to population-level health surveillance. However, such integration into clinical practice is constrained by a multifaceted “trustworthiness gap”: the mismatch between the raw capabilities demonstrated by Medical MLLMs in controlled experiments and their reliability, fairness, safety, privacy protection, explainability, and regulatory accountability required for clinical deployment in the real world. This gap requires systematic examination of both model behavior and clinical risk across the full lifecycle of Medical MLLM development and use. In this paper, we survey the existing literature on methods, evaluation, and benchmarks of trustworthiness research in Medical MLLMs. We first establish a multi-scale clinical landscape across tissue, organ, individual, and population levels to categorize current models, providing critical insights into the interplay between data heterogeneity and clinical task requirements. Subsequently, we propose a holistic, six-dimensional taxonomy of trustworthiness, comprising truthfulness, robustness, fairness, safety, privacy, and explainability. Using this taxonomy, we critically analyze existing literature to elicit recurring Medical MLLM trustworthiness failure modes and suggest appropriate remediation strategies. Furthermore, our survey enumerates open challenges in evaluating trustworthiness by means of traditional automated metrics, frontier “LLM-as-a-Judge” methods and expert-centric assessment protocols. As a response to these challenges we identify emerging research directions including dynamic and workflow-oriented evaluations for interactive Medical MLLMs. We aim for this work to serve as a systematic guide for researchers and practitioners aiming to develop the next generation of trustworthy medical AI by achieving clinically acceptable reliability through technological innovations. Project Link: https://github.com/junyuanM/Trustworthy-Medical-MLLMs-Survey.
- 11 September 2026
- Published by MDPI AG in Preprints.org
Open AccessTrustworthy Medical Multimodal Large Language Models: Taxonomy, Evaluation, and Benchmarks
The integration of Multimodal Large Language Models (MLLMs) into healthcare has the potential to drive a considerable advancement toward many AI use cases in medicine, offering transformative capabilities across hierarchical levels of clinical granularity, ranging from microscopic tissue analysis over organ level imaging and individual patient modeling to population-level health surveillance. However, such integration into clinical practice is constrained by a multifaceted “trustworthiness gap”: the mismatch between the raw capabilities demonstrated by Medical MLLMs in controlled experiments and their reliability, fairness, safety, privacy protection, explainability, and regulatory accountability required for clinical deployment in the real world. This gap requires systematic examination of both model behavior and clinical risk across the full lifecycle of Medical MLLM development and use. In this paper, we survey the existing literature on methods, evaluation, and benchmarks of trustworthiness research in Medical MLLMs. We first establish a multi-scale clinical landscape across tissue, organ, individual, and population levels to categorize current models, providing critical insights into the interplay between data heterogeneity and clinical task requirements. Subsequently, we propose a holistic, six-dimensional taxonomy of trustworthiness, comprising truthfulness, robustness, fairness, safety, privacy, and explainability. Using this taxonomy, we critically analyze existing literature to elicit recurring Medical MLLM trustworthiness failure modes and suggest appropriate remediation strategies. Furthermore, our survey enumerates open challenges in evaluating trustworthiness by means of traditional automated metrics, frontier “LLM-as-a-Judge” methods and expert-centric assessment protocols. As a response to these challenges we identify emerging research directions including dynamic and workflow-oriented evaluations for interactive Medical MLLMs. We aim for this work to serve as a systematic guide for researchers and practitioners aiming to develop the next generation of trustworthy medical AI by achieving clinically acceptable reliability through technological innovations. Project Link: https://github.com/junyuanM/Trustworthy-Medical-MLLMs-Survey.
- Open Access
- 3 September 2026
- Published by Wiley in Expert Systems
- Vol. 43 (10) , e70383
From SAM 1 to SAM 3: Benchmarking Zero‐Shot Cross‐Domain Medical Image Segmentation
Segmentation models have demonstrated significant potential in medical image segmentation. However, there is currently a lack of systematic, cross‐generation comparative evaluations to assess whether the iterations from SAM1 to SAM 3 can effectively enhance the clinical applicability of zero‐shot segmentation. To address this issue, this paper systematically compares the zero‐shot segmentation performance of the SAM family across four representative medical imaging datasets which cover different imaging modalities. The evaluation framework encompasses multi‐dimensional metrics and introduces logarithmic‐domain HD95 analysis to capture and quantify boundary spillover behaviour. This paper demonstrates that the evolution from SAM1 to SAM 3 represents a qualitative leap rather than a linear accumulation; SAM 3 achieves a several‐fold increase in Dice score over SAM1 on most datasets, whilst effectively suppressing boundary spillover into background; conversely, SAM 2 shows no significant advantage over SAM1 in our benchmarks. Concurrently, the zero‐shot capabilities of the models exhibit a distinct modality gradient. Only in a few modalities can SAM 3 maintain acceptable segmentation results, while in all other modalities the entire SAM family consistently fails. In summary, SAM 3 demonstrates zero‐shot clinical applicability in specific modalities. Its semantic prompts are able to segment medical images more easily and accurately than visual prompts.
- 3 September 2026
- Published by Wiley in Expert Systems
- Vol. 43 (10) , e70383
Open AccessFrom SAM 1 to SAM 3: Benchmarking Zero‐Shot Cross‐Domain Medical Image Segmentation
Segmentation models have demonstrated significant potential in medical image segmentation. However, there is currently a lack of systematic, cross‐generation comparative evaluations to assess whether the iterations from SAM1 to SAM 3 can effectively enhance the clinical applicability of zero‐shot segmentation. To address this issue, this paper systematically compares the zero‐shot segmentation performance of the SAM family across four representative medical imaging datasets which cover different imaging modalities. The evaluation framework encompasses multi‐dimensional metrics and introduces logarithmic‐domain HD95 analysis to capture and quantify boundary spillover behaviour. This paper demonstrates that the evolution from SAM1 to SAM 3 represents a qualitative leap rather than a linear accumulation; SAM 3 achieves a several‐fold increase in Dice score over SAM1 on most datasets, whilst effectively suppressing boundary spillover into background; conversely, SAM 2 shows no significant advantage over SAM1 in our benchmarks. Concurrently, the zero‐shot capabilities of the models exhibit a distinct modality gradient. Only in a few modalities can SAM 3 maintain acceptable segmentation results, while in all other modalities the entire SAM family consistently fails. In summary, SAM 3 demonstrates zero‐shot clinical applicability in specific modalities. Its semantic prompts are able to segment medical images more easily and accurately than visual prompts.
- Top Cited Papers
- 1 September 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (9) , 10221-10234
Neural Network Optimization Reimagined: Decoupled Techniques for Scratch and Fine-Tuning
With the accumulation of resources in the era of big data and the rise of pre-trained models in deep learning, optimizing neural networks for various tasks often involves different strategies for fine-tuning pre-trained models versus training from scratch. However, existing optimizers primarily focus on reducing the loss function by updating model parameters, without fully addressing the unique demands of these two major paradigms. In this paper, we propose DualOpt, a novel approach that decouples optimization techniques specifically tailored for these distinct training scenarios. For training from scratch, we introduce real-time layer-wise weight decay, designed to enhance both convergence and generalization by aligning with the characteristics of weight updates and network architecture. For more importantly fine-tuning, we integrate weight rollback with the optimizer, incorporating a rollback term into each weight update step. This ensures consistency in the weight distribution between upstream and downstream models, effectively mitigating knowledge forgetting and improving fine-tuning performance. Additionally, we extend the layer-wise weight decay to dynamically adjust the rollback levels across layers, adapting to the varying demands of different downstream tasks. Extensive experiments across diverse tasks, including image classification, object detection, semantic segmentation, and instance segmentation, demonstrate the broad applicability and state-of-the-art performance of DualOpt. Code is available at https://github.com/qklee-lz/OLOR-AAAI-2024.
- 1 September 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (9) , 10221-10234
Top Cited PapersNeural Network Optimization Reimagined: Decoupled Techniques for Scratch and Fine-Tuning
With the accumulation of resources in the era of big data and the rise of pre-trained models in deep learning, optimizing neural networks for various tasks often involves different strategies for fine-tuning pre-trained models versus training from scratch. However, existing optimizers primarily focus on reducing the loss function by updating model parameters, without fully addressing the unique demands of these two major paradigms. In this paper, we propose DualOpt, a novel approach that decouples optimization techniques specifically tailored for these distinct training scenarios. For training from scratch, we introduce real-time layer-wise weight decay, designed to enhance both convergence and generalization by aligning with the characteristics of weight updates and network architecture. For more importantly fine-tuning, we integrate weight rollback with the optimizer, incorporating a rollback term into each weight update step. This ensures consistency in the weight distribution between upstream and downstream models, effectively mitigating knowledge forgetting and improving fine-tuning performance. Additionally, we extend the layer-wise weight decay to dynamically adjust the rollback levels across layers, adapting to the varying demands of different downstream tasks. Extensive experiments across diverse tasks, including image classification, object detection, semantic segmentation, and instance segmentation, demonstrate the broad applicability and state-of-the-art performance of DualOpt. Code is available at https://github.com/qklee-lz/OLOR-AAAI-2024.
- Open Access
- 20 August 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 333
PARF-Net: Phase-adaptive robust fusion for liver tumor segmentation with missing-phase CECT
- 20 August 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 333
Open AccessPARF-Net: Phase-adaptive robust fusion for liver tumor segmentation with missing-phase CECT