Xiao Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Xiao Zhang has a total of 16 co-authors. Here are the top co-authors by number of shared publications:
- Huimin Ma
University of Science and Technology Beijing, Beijing, China
4 shared publications - Tianyu Hu
University of Science and Technology Beijing
4 shared publications - Juntao Lyu
School of Computer and Communication Engineering, University of Science and Technology Beijing, Beijing, China
2 shared publications - Kangsheng Wang
University of Science and Technology Beijing
2 shared publications - Fang Zhou
University of Science and Technology Beijing
1 shared publication - Huansheng Ning
University of Science and Technology Beijing
1 shared publication
Latest Publications
- Open Access
- 22 April 2026
- Published by Springer Nature in Nature Communications
- Vol. 17
Autonomous driving system based on dual process theory and deliberate practice theory
Autonomous driving, despite significant progress, is still not widely applied in open, unconstrained environments, primarily owing to deficiencies in hazard perception, few-shot generalization, corner case generalization, and evaluation metrics, resulting in reliability concerns. To address these challenges, we propose CogniDrive, a framework based on dual-process and deliberate practice theories, leveraging contextual reasoning of the Large Language Model (LLM) to enhance driving systems robustness and generalization. Inspired by dual-process theory, CogniDrive comprises two cognition modes: InstinctNav for rapid, intuitive decision-making and ReflectPlan for reflective reasoning. Enhanced by a thought model and experience embedding for LLM, InstinctNav combines behavioral cloning and retrieval augmented generation to enhance few-shot learning efficiency based on deliberate practice theory. ReflectPlan processes and internalizes reward signals embedded in language tokens within the prompt, derived from a self-reflection mechanism, to enable continuous improvement and generalization. To detect hazards in corner cases precluded by limited training data distribution, a vision language model is integrated for comprehensive environmental understanding through multimodal self-reflection. We further propose an evaluation framework that complements traditional metrics by emphasizing safety, comfort, and energy efficiency, and demonstrate state-of-the-art performance through extensive open-loop and closed-loop experiments.
- 22 April 2026
- Published by Springer Nature in Nature Communications
- Vol. 17
Open AccessAutonomous driving system based on dual process theory and deliberate practice theory
Autonomous driving, despite significant progress, is still not widely applied in open, unconstrained environments, primarily owing to deficiencies in hazard perception, few-shot generalization, corner case generalization, and evaluation metrics, resulting in reliability concerns. To address these challenges, we propose CogniDrive, a framework based on dual-process and deliberate practice theories, leveraging contextual reasoning of the Large Language Model (LLM) to enhance driving systems robustness and generalization. Inspired by dual-process theory, CogniDrive comprises two cognition modes: InstinctNav for rapid, intuitive decision-making and ReflectPlan for reflective reasoning. Enhanced by a thought model and experience embedding for LLM, InstinctNav combines behavioral cloning and retrieval augmented generation to enhance few-shot learning efficiency based on deliberate practice theory. ReflectPlan processes and internalizes reward signals embedded in language tokens within the prompt, derived from a self-reflection mechanism, to enable continuous improvement and generalization. To detect hazards in corner cases precluded by limited training data distribution, a vision language model is integrated for comprehensive environmental understanding through multimodal self-reflection. We further propose an evaluation framework that complements traditional metrics by emphasizing safety, comfort, and energy efficiency, and demonstrate state-of-the-art performance through extensive open-loop and closed-loop experiments.
- 1 November 2025
- Published by Springer Nature
- p. 212-223
Adversarial Iterative Pre-enactment Framework for Air Combat Based on Mental Simulation Theory
- 1 November 2025
- Published by Springer Nature
- p. 212-223
Adversarial Iterative Pre-enactment Framework for Air Combat Based on Mental Simulation Theory
- Open Access
- 22 April 2025
- Published by Springer Nature in Research Square
Efficient Knowledge Transfer in Multi-Task Learning through Task-Adaptive Low-Rank Representation
Pre-trained language models (PLMs) demonstrate remarkable intelligence but struggle with emerging tasks unseen during training in real-world applications. Training separate models for each new task is usually impractical. Multi-task learning (MTL) addresses this challenge by transferring shared knowledge from source tasks to target tasks. As an dominant parameter-efficient fine-tuning method, prompt tuning (PT) enhances MTL by introducing an adaptable vector that captures task-specific knowledge, which acts as a prefix to the original prompt that preserves shared knowledge, while keeping PLM parameters frozen. However, PT struggles to effectively capture the heterogeneity of task-specific knowledge due to its limited representational capacity. To address this challenge, we propose Task-Adaptive Low-Rank Representation (TA-LoRA), an MTL method built on PT, employing the low-rank representation to model task heterogeneity and a fast-slow weights mechanism where the slow weight encodes shared knowledge, while the fast weight captures task-specific nuances, avoiding the mixing of shared and task-specific knowledge, caused by training lowrank representations from scratch. Moreover, a zero-initialized attention mechanism is introduced to minimize the disruption of immature low-rank components on original prompts during warm-up epochs. Experiments on 16 tasks demonstrate that TALoRA achieves state-of-the-art performance in full-data and fewshot settings while maintaining superior parameter efficiency.
- 22 April 2025
- Published by Springer Nature in Research Square
Open AccessEfficient Knowledge Transfer in Multi-Task Learning through Task-Adaptive Low-Rank Representation
Pre-trained language models (PLMs) demonstrate remarkable intelligence but struggle with emerging tasks unseen during training in real-world applications. Training separate models for each new task is usually impractical. Multi-task learning (MTL) addresses this challenge by transferring shared knowledge from source tasks to target tasks. As an dominant parameter-efficient fine-tuning method, prompt tuning (PT) enhances MTL by introducing an adaptable vector that captures task-specific knowledge, which acts as a prefix to the original prompt that preserves shared knowledge, while keeping PLM parameters frozen. However, PT struggles to effectively capture the heterogeneity of task-specific knowledge due to its limited representational capacity. To address this challenge, we propose Task-Adaptive Low-Rank Representation (TA-LoRA), an MTL method built on PT, employing the low-rank representation to model task heterogeneity and a fast-slow weights mechanism where the slow weight encodes shared knowledge, while the fast weight captures task-specific nuances, avoiding the mixing of shared and task-specific knowledge, caused by training lowrank representations from scratch. Moreover, a zero-initialized attention mechanism is introduced to minimize the disruption of immature low-rank components on original prompts during warm-up epochs. Experiments on 16 tasks demonstrate that TALoRA achieves state-of-the-art performance in full-data and fewshot settings while maintaining superior parameter efficiency.
- Open Access
- 25 March 2025
- Published by Springer Nature in Research Square
CSCE: Boosting LLM Reasoning by Simultaneous Enhancing of Causal Significance and Consistency
Chain-based reasoning methods like chain of thought (CoT) play a rising role in solving reasoning tasks for large language models (LLMs). However, the causal hallucinations between a step of reasoning and corresponding state transitions are becoming a significant obstacle to advancing LLMs' reasoning capabilities, especially in long-range reasoning tasks. This paper proposes a non-chain-based reasoning framework for simultaneous consideration of causal significance and consistency, i.e., the Causal Significance and Consistency Enhancer (CSCE). We customize LLM's loss function utilizing treatment effect assessments to enhance its reasoning ability from two aspects: causal significance and consistency. This ensures that the model captures essential causal relationships and maintains robust and consistent performance across various scenarios. Additionally, we transform the reasoning process from the cascading multiple one-step reasoning commonly used in Chain-Based methods, like CoT, to a causal-enhanced method that outputs the entire reasoning process in one go, further improving the model's reasoning efficiency. Extensive experiments show that our method improves both the reasoning success rate and speed. These improvements further demonstrate that non-chain-based methods can also aid LLMs in completing reasoning tasks.
- 25 March 2025
- Published by Springer Nature in Research Square
Open AccessCSCE: Boosting LLM Reasoning by Simultaneous Enhancing of Causal Significance and Consistency
Chain-based reasoning methods like chain of thought (CoT) play a rising role in solving reasoning tasks for large language models (LLMs). However, the causal hallucinations between a step of reasoning and corresponding state transitions are becoming a significant obstacle to advancing LLMs' reasoning capabilities, especially in long-range reasoning tasks. This paper proposes a non-chain-based reasoning framework for simultaneous consideration of causal significance and consistency, i.e., the Causal Significance and Consistency Enhancer (CSCE). We customize LLM's loss function utilizing treatment effect assessments to enhance its reasoning ability from two aspects: causal significance and consistency. This ensures that the model captures essential causal relationships and maintains robust and consistent performance across various scenarios. Additionally, we transform the reasoning process from the cascading multiple one-step reasoning commonly used in Chain-Based methods, like CoT, to a causal-enhanced method that outputs the entire reasoning process in one go, further improving the model's reasoning efficiency. Extensive experiments show that our method improves both the reasoning success rate and speed. These improvements further demonstrate that non-chain-based methods can also aid LLMs in completing reasoning tasks.
- 22 February 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Computational Biology and Bioinformatics
- Vol. 21 (4) , 1051-1059
CASL: Capturing Activity Semantics Through Location Information for Enhanced Activity Recognition
Using portable tools to monitor and identify daily activities has increasingly become a focus of digital healthcare, especially for elderly care. One of the difficulties in this area is the excessive reliance on labeled activity data for corresponding recognition modeling. Labeled activity data is expensive to collect. To address this challenge, we propose an effective and robust semi-supervised active learning method, called CASL, which combines the mainstream semi-supervised learning method with a mechanism of expert collaboration. CASL takes a user's trajectory as the only input. In addition, CASL uses expert collaboration to judge the valuable samples of a model to further enhance its performance. CASL relies on very few semantic activities, outperforms all baseline activity recognition methods, and is close to the performance of supervised learning methods. On the adlnormal dataset with 200 semantic activities data, CASL achieved an accuracy of 89.07%, supervised learning has 91.77%. Our ablation study validated the components in our CASL using a query strategy and a data fusion approach.
- 22 February 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Computational Biology and Bioinformatics
- Vol. 21 (4) , 1051-1059
CASL: Capturing Activity Semantics Through Location Information for Enhanced Activity Recognition
Using portable tools to monitor and identify daily activities has increasingly become a focus of digital healthcare, especially for elderly care. One of the difficulties in this area is the excessive reliance on labeled activity data for corresponding recognition modeling. Labeled activity data is expensive to collect. To address this challenge, we propose an effective and robust semi-supervised active learning method, called CASL, which combines the mainstream semi-supervised learning method with a mechanism of expert collaboration. CASL takes a user's trajectory as the only input. In addition, CASL uses expert collaboration to judge the valuable samples of a model to further enhance its performance. CASL relies on very few semantic activities, outperforms all baseline activity recognition methods, and is close to the performance of supervised learning methods. On the adlnormal dataset with 200 semantic activities data, CASL achieved an accuracy of 89.07%, supervised learning has 91.77%. Our ablation study validated the components in our CASL using a query strategy and a data fusion approach.