Jie Zhou
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Jie Zhou has a total of 175 co-authors. Here are the top co-authors by number of shared publications:
- Fandong Meng
Tencent Inc., Beijing, China
15 shared publications - Ruobing Xie
Tencent, Beijing, China
8 shared publications - Jinchao Zhang
WeChat AI, Tencent Inc., Beijing, China
6 shared publications - Zhiyuan Liu
Tsinghua University
6 shared publications - Jinan Xu
Beijing Jiaotong University
5 shared publications - Leyu Lin
Tencent, Beijing, China
4 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 12969-12978
RAVA: Retrieval-Augmented Viewpoint Alignment for Subject-Driven Image Generation
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 12969-12978
Open AccessRAVA: Retrieval-Augmented Viewpoint Alignment for Subject-Driven Image Generation
- 26 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. PP (01628828) , 1-17
CoDA: Color Distribution Probing for Efficient and Generalizable AI-Generated Image Detection
AI-generated image detection faces a persistent trade-off between generalization and efficiency: lightweight artifact-based methods often degrade on unseen generators or domains, whereas more robust large-scale models are computationally expensive. Meanwhile, existing benchmarks mainly focus on cross-model evaluation in photorealistic settings, leaving cross-domain robustness underexplored. To address this gap, we introduce FakeForm, a large-scale benchmark with approximately 370,000 images across 62 diverse domains for both cross-model and cross-domain evaluation. Motivated by this broader setting, we revisit color-distribution probing as an efficient complementary cue for AI-generated image detection. We observe that, especially for photographic content, real photographs tend to exhibit smoother and more stable color patterns, whereas synthetic images often show characteristic color imbalances introduced by neural generation. Based on this observation, we propose CoDA, a compact 1.48M-parameter detector built on a Noise-Quantization Probe, together with a theoretical analysis linking probe responses to color non-uniformity. Experiments show that CoDA achieves state-of-the-art performance on standard benchmarks and the best results on the challenging cross-domain evaluation of FakeForm, while remaining highly competitive in cross-model photorealistic settings. These results suggest that persistent generative artifacts can provide a practical foundation for efficient and robust AI-generated image detection. The models and FakeForm benchmark will be made publicly available.
- 26 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. PP (01628828) , 1-17
CoDA: Color Distribution Probing for Efficient and Generalizable AI-Generated Image Detection
AI-generated image detection faces a persistent trade-off between generalization and efficiency: lightweight artifact-based methods often degrade on unseen generators or domains, whereas more robust large-scale models are computationally expensive. Meanwhile, existing benchmarks mainly focus on cross-model evaluation in photorealistic settings, leaving cross-domain robustness underexplored. To address this gap, we introduce FakeForm, a large-scale benchmark with approximately 370,000 images across 62 diverse domains for both cross-model and cross-domain evaluation. Motivated by this broader setting, we revisit color-distribution probing as an efficient complementary cue for AI-generated image detection. We observe that, especially for photographic content, real photographs tend to exhibit smoother and more stable color patterns, whereas synthetic images often show characteristic color imbalances introduced by neural generation. Based on this observation, we propose CoDA, a compact 1.48M-parameter detector built on a Noise-Quantization Probe, together with a theoretical analysis linking probe responses to color non-uniformity. Experiments show that CoDA achieves state-of-the-art performance on standard benchmarks and the best results on the challenging cross-domain evaluation of FakeForm, while remaining highly competitive in cross-model photorealistic settings. These results suggest that persistent generative artifacts can provide a practical foundation for efficient and robust AI-generated image detection. The models and FakeForm benchmark will be made publicly available.
- Open Access
- 2 March 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Information Systems
- Vol. 44 (3) , 1-35
Rethinking User Retention Modeling in Recommendation
Recommendations usually focus on immediate accuracy metrics like Click-Through Rate (CTR), ignoring user long-term metrics. User retention, which reflects the percentage of today’s users who will return to the system in the next few days, should be paid more attention to. However, most existing methods did not focus on user retention, since their complexity and uncertainty make it extremely hard to discover why a user will or will not return to a system. Recently, a few pioneers have optimized user retention, focusing solely on accuracy without delving into its underlying rationale. This is primarily due to the absence of explicit supervised signals. In this work, we design a Behavior-wise Contrastive Multi-Instance Learning (BCMIL) module, which jointly models clicked and impressed items to capture interpretable user retention. Specifically, we conduct in-depth analyses in real-world scenarios to discover implicit retention-related supervised signals. To model these signals, we design a Forward Supervised Signals Extractor (FSSE) that utilizes a heterogeneous graph, enhancing the reliability of user retention. To mitigate randomness and uncertainty, we propose a Backward Supervised Signals Stabilizer (BSSS) that utilizes overlooked label-part behaviors within each training window to retrospectively guide the training process. Offline and online evaluations of an industrial system verify the effectiveness of our methods.
- 2 March 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Information Systems
- Vol. 44 (3) , 1-35
Open AccessRethinking User Retention Modeling in Recommendation
Recommendations usually focus on immediate accuracy metrics like Click-Through Rate (CTR), ignoring user long-term metrics. User retention, which reflects the percentage of today’s users who will return to the system in the next few days, should be paid more attention to. However, most existing methods did not focus on user retention, since their complexity and uncertainty make it extremely hard to discover why a user will or will not return to a system. Recently, a few pioneers have optimized user retention, focusing solely on accuracy without delving into its underlying rationale. This is primarily due to the absence of explicit supervised signals. In this work, we design a Behavior-wise Contrastive Multi-Instance Learning (BCMIL) module, which jointly models clicked and impressed items to capture interpretable user retention. Specifically, we conduct in-depth analyses in real-world scenarios to discover implicit retention-related supervised signals. To model these signals, we design a Forward Supervised Signals Extractor (FSSE) that utilizes a heterogeneous graph, enhancing the reliability of user retention. To mitigate randomness and uncertainty, we propose a Backward Supervised Signals Stabilizer (BSSS) that utilizes overlooked label-part behaviors within each training window to retrospectively guide the training process. Offline and online evaluations of an industrial system verify the effectiveness of our methods.
- Top Cited PapersOpen Access
- 19 January 2026
- Published by Springer Nature in Nature Health
- Vol. 1 (2) , 238-250
A community-codesigned LLM-powered chatbot for primary care: a randomized controlled trial
With a global shortage of primary healthcare physicians—particularly in resource-limited settings—large language models (LLMs) have the potential to support and enhance patients’ health awareness. Here we developed P&P Care (Population Medicine and Public Health), an LLM-powered primary care chatbot using a dual-track role-play codesign framework where community stakeholders and researchers simulated each another’s perspectives across four phases: contextual understanding; cocreation; testing and refinement; and implementation and evolution. The codesigned chatbot was integrated with e-learning modules and tested in a randomized controlled trial. The trial included 2,113 participants (1,052 women and 1,061 men) from urban and rural areas across 11 Chinese provinces who were randomly assigned to receive a consultation either with preparatory e-learning via the P&P Care or without. The study met its primary endpoint with the e-learning group showing significantly higher objective health awareness (mean score 2.95 ± 1.22) compared with the consultation-only group (mean score 2.34 ± 1.02; P < 0.001). Codesign offers a scalable solution for deploying LLMs in resource-limited settings. Chinese Clinical Trial Registry identifier: ChiCTR2500098101 .
- 19 January 2026
- Published by Springer Nature in Nature Health
- Vol. 1 (2) , 238-250
Top Cited PapersOpen AccessA community-codesigned LLM-powered chatbot for primary care: a randomized controlled trial
With a global shortage of primary healthcare physicians—particularly in resource-limited settings—large language models (LLMs) have the potential to support and enhance patients’ health awareness. Here we developed P&P Care (Population Medicine and Public Health), an LLM-powered primary care chatbot using a dual-track role-play codesign framework where community stakeholders and researchers simulated each another’s perspectives across four phases: contextual understanding; cocreation; testing and refinement; and implementation and evolution. The codesigned chatbot was integrated with e-learning modules and tested in a randomized controlled trial. The trial included 2,113 participants (1,052 women and 1,061 men) from urban and rural areas across 11 Chinese provinces who were randomly assigned to receive a consultation either with preparatory e-learning via the P&P Care or without. The study met its primary endpoint with the e-learning group showing significantly higher objective health awareness (mean score 2.95 ± 1.22) compared with the consultation-only group (mean score 2.34 ± 1.02; P < 0.001). Codesign offers a scalable solution for deploying LLMs in resource-limited settings. Chinese Clinical Trial Registry identifier: ChiCTR2500098101 .
- Top Cited PapersOpen Access
- 19 January 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (3) , 934-942
An LLM chatbot to facilitate primary-to-specialist care transitions: a randomized controlled trial
Patient-facing large language models (LLMs) hold potential to streamline inefficient transitions from primary to specialist care. We developed the preassessment (PreA), an LLM chatbot co-designed with local stakeholders, to perform the general medical consultations for history-taking, preliminary diagnoses, and test ordering that would normally be performed by primary care providers and to generate referral reports for specialists. PreA was tested in a randomized controlled trial involving 111 specialists from 24 medical disciplines across two health centers, where 2,069 patients (1,141 women; 928 men) were randomly assigned to use PreA independently (PreA-only), use it with staff support (PreA-human), or not use it (No-PreA) before specialist consultation. The trial met its primary end points with the PreA-only group showing significantly reduced physician consultation duration (28.7% reduction; 3.14 ± 2.25 min) compared to the No-PreA group (4.41 ± 2.77 min; P < 0.001), alongside significant improvements in physician-perceived care coordination (mean scores 113.1% increase; 3.69 ± 0.90 versus 1.73 ± 0.95; P < 0.001) and patient-reported communication ease (mean scores 16.0% increase; 3.99 ± 0.62 versus 3.44 ± 0.97; P < 0.001). Equivalent outcomes between the PreA-only and PreA-human groups confirmed the autonomous operation capability. Co-designed PreA outperformed the same model with additional fine-tuning on local dialogues across clinical decision-making domains. Co-design with local stakeholders, compared to passive local data collecting, represents a more effective strategy for deploying LLMs to strengthen health systems and enhance patient-centered care in resource-limited settings. Chinese Clinical Trial Registry identifier: ChiCTR2400094159 .
- 19 January 2026
- Published by Springer Nature in Nature Medicine
- Vol. 32 (3) , 934-942
Top Cited PapersOpen AccessAn LLM chatbot to facilitate primary-to-specialist care transitions: a randomized controlled trial
Patient-facing large language models (LLMs) hold potential to streamline inefficient transitions from primary to specialist care. We developed the preassessment (PreA), an LLM chatbot co-designed with local stakeholders, to perform the general medical consultations for history-taking, preliminary diagnoses, and test ordering that would normally be performed by primary care providers and to generate referral reports for specialists. PreA was tested in a randomized controlled trial involving 111 specialists from 24 medical disciplines across two health centers, where 2,069 patients (1,141 women; 928 men) were randomly assigned to use PreA independently (PreA-only), use it with staff support (PreA-human), or not use it (No-PreA) before specialist consultation. The trial met its primary end points with the PreA-only group showing significantly reduced physician consultation duration (28.7% reduction; 3.14 ± 2.25 min) compared to the No-PreA group (4.41 ± 2.77 min; P < 0.001), alongside significant improvements in physician-perceived care coordination (mean scores 113.1% increase; 3.69 ± 0.90 versus 1.73 ± 0.95; P < 0.001) and patient-reported communication ease (mean scores 16.0% increase; 3.99 ± 0.62 versus 3.44 ± 0.97; P < 0.001). Equivalent outcomes between the PreA-only and PreA-human groups confirmed the autonomous operation capability. Co-designed PreA outperformed the same model with additional fine-tuning on local dialogues across clinical decision-making domains. Co-design with local stakeholders, compared to passive local data collecting, represents a more effective strategy for deploying LLMs to strengthen health systems and enhance patient-centered care in resource-limited settings. Chinese Clinical Trial Registry identifier: ChiCTR2400094159 .
- Open Access
- 27 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 6654-6662
ArtFRD: A Fisher-Rao Mixture Metric for Generative Model Aesthetic Evaluation
- 27 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 6654-6662
Open AccessArtFRD: A Fisher-Rao Mixture Metric for Generative Model Aesthetic Evaluation
- Top Cited Papers
- 5 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 33, 3776-3786
An Empirical Study of Catastrophic Forgetting in Large Language Models During Continual Fine-Tuning
Catastrophic forgetting (CF) is a phenomenon that occurs in machine learning when a model forgets previously learned information while acquiring new knowledge for achieving satisfactory performance in downstream tasks. As large language models (LLMs) have demonstrated remarkable performance, it is intriguing to investigate whether CF exists during the continual instruction tuning of LLMs. This study empirically evaluates the forgetting phenomenon in LLMs' knowledge during continual instruction tuning from the perspectives of domain knowledge, reasoning, and reading comprehension. The experiments reveal that catastrophic forgetting is generally observed in LLMs ranging from 1b to 7b parameters. Surprisingly, as the model scale increases, the severity of forgetting intensifies in such a model scale range, which may result from the much more significant initial performance in the larger LLM. The finding is also observed by the experiment of Qwen-2.5-Inst from 3B to 14B. Comparing the decoder-only model BLOOMZ with the encoder-decoder model mT0, BLOOMZ exhibits less forgetting and retains more knowledge. Interestingly, we also observe that LLMs can mitigate language biases, such as gender bias, during continual fine-tuning. Furthermore, our findings indicate that general instruction tuning can help alleviate the forgetting phenomenon in LLMs during subsequent fine-tuning.
- 5 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 33, 3776-3786
Top Cited PapersAn Empirical Study of Catastrophic Forgetting in Large Language Models During Continual Fine-Tuning
Catastrophic forgetting (CF) is a phenomenon that occurs in machine learning when a model forgets previously learned information while acquiring new knowledge for achieving satisfactory performance in downstream tasks. As large language models (LLMs) have demonstrated remarkable performance, it is intriguing to investigate whether CF exists during the continual instruction tuning of LLMs. This study empirically evaluates the forgetting phenomenon in LLMs' knowledge during continual instruction tuning from the perspectives of domain knowledge, reasoning, and reading comprehension. The experiments reveal that catastrophic forgetting is generally observed in LLMs ranging from 1b to 7b parameters. Surprisingly, as the model scale increases, the severity of forgetting intensifies in such a model scale range, which may result from the much more significant initial performance in the larger LLM. The finding is also observed by the experiment of Qwen-2.5-Inst from 3B to 14B. Comparing the decoder-only model BLOOMZ with the encoder-decoder model mT0, BLOOMZ exhibits less forgetting and retains more knowledge. Interestingly, we also observe that LLMs can mitigate language biases, such as gender bias, during continual fine-tuning. Furthermore, our findings indicate that general instruction tuning can help alleviate the forgetting phenomenon in LLMs during subsequent fine-tuning.
- Open Access
- 20 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 224-235
Personalized Language Model Learning on Text Data Without User Identifiers
- 20 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 224-235
Open AccessPersonalized Language Model Learning on Text Data Without User Identifiers
- Open Access
Continuous Visual Autoregressive Generation via Score Maximization
Conventional wisdom suggests that autoregressive models are used to process discrete data. When applied to continuous modalities such as visual data, Visual AutoRegressive modeling (VAR) typically resorts to quantization-based approaches to cast the data into a discrete space, which can introduce significant information loss. To tackle this issue, we introduce a Continuous VAR framework that enables direct visual autoregressive generation without vector quantization. The underlying theoretical foundation is strictly proper scoring rules, which provide powerful statistical tools capable of evaluating how well a generative model approximates the true distribution. Within this framework, all we need is to select a strictly proper score and set it as the training objective to optimize. We primarily explore a class of training objectives based on the energy score, which is likelihood-free and thus overcomes the difficulty of making probabilistic predictions in the continuous space. Previous efforts on continuous autoregressive generation, such as GIVT and diffusion loss, can also be derived from our framework using other strictly proper scores. Source code: https://github.com/shaochenze/EAR.
Open AccessContinuous Visual Autoregressive Generation via Score Maximization
Conventional wisdom suggests that autoregressive models are used to process discrete data. When applied to continuous modalities such as visual data, Visual AutoRegressive modeling (VAR) typically resorts to quantization-based approaches to cast the data into a discrete space, which can introduce significant information loss. To tackle this issue, we introduce a Continuous VAR framework that enables direct visual autoregressive generation without vector quantization. The underlying theoretical foundation is strictly proper scoring rules, which provide powerful statistical tools capable of evaluating how well a generative model approximates the true distribution. Within this framework, all we need is to select a strictly proper score and set it as the training objective to optimize. We primarily explore a class of training objectives based on the energy score, which is likelihood-free and thus overcomes the difficulty of making probabilistic predictions in the continuous space. Previous efforts on continuous autoregressive generation, such as GIVT and diffusion loss, can also be derived from our framework using other strictly proper scores. Source code: https://github.com/shaochenze/EAR.