Ziqian Bi

Ziqian Bi

Auto-generated profile, based on publicly available information

Ziqian Bi

Purdue University

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications

19 December 2025 in TechRxiv
Verified
Preprint
Open access
YMYang Mo
Yang Mo
EGEnze Ge
Enze Ge
XHXudong Han
Xudong Han
JSJunhao Song
Junhao Song
ZBZiqian Bi
Ziqian Bi
BHBinhua Huang
Binhua Huang
Chain-of-Thought (CoT) reasoning has emerged as a transformative paradigm for enhancing the reasoning capabilities of Large Language Models (LLMs). This survey provides a comprehensive analysis of CoT reasoning, spanning from foundational prompting techniques to the recent emergence of Large Reasoning Models (LRMs) such as OpenAI o1 and DeepSeek-R1. We present a unified taxonomy that categorizes CoT methods into four dimensions: (1) prompting strategies including zero-shot, few-shot, and self-consistency approaches; (2) structural extensions such as Tree-of-Thought, Graph-of-Thought, and Forestof-Thought; (3) training paradigms encompassing supervised fine-tuning, reinforcement learning, and preference optimization; and (4) emerging paradigms including latent reasoning and multi-modal CoT. Through extensive analysis of over 150 papers, we identify key characteristics that distinguish Long CoT reasoning, including deep multi-step inference, heuristicguided search, and self-verification mechanisms. We systematically compare performance across 25+ benchmarks spanning mathematical, commonsense, symbolic, logical, and multi-modal reasoning tasks. Our analysis reveals critical challenges including reasoning faithfulness, computational efficiency, and the "illusion of thinking" phenomenon. We conclude by outlining promising research directions including efficient reasoning, multi-modal integration, and the development of more robust evaluation frameworks. This survey serves as a comprehensive resource for researchers and practitioners seeking to understand and advance the frontier of reasoning in LLMs.

19 December 2025 in TechRxiv
Verified
Preprint
Open access
SXShengkai Xu
Shengkai Xu
HKHsiang Lun Kao
Hsiang Lun Kao
TXTianxiang Xu
Tianxiang Xu
HZHonghui Zhang
Honghui Zhang
JWJunqiao Wang
Junqiao Wang
YOYuqi Ouyang
Yuqi Ouyang
Polyp detectors trained on clean datasets often underperform in realworld endoscopy, where illumination changes, motion blur, and occlusions degrade image quality. Existing approaches struggle with the domain gap between controlled laboratory conditions and clinical practice, where adverse imaging conditions are prevalent. In this work, we propose ADAPTIVEDETECTOR, a novel two-stage detector-verifier framework comprising a YOLOv11 detector with a vision-language model (VLM) verifier. The detector adaptively adjusts per-frame confidence thresholds under VLM guidance, while the verifier is fine-tuned with Group Relative Policy Optimization (GRPO) using an asymmetric, cost-sensitive reward function specifically designed to discourage missed detections-a critical clinical requirement. To enable realistic assessment under challenging conditions, we construct a comprehensive synthetic testbed by systematically degrading clean datasets with adverse conditions commonly encountered in clinical practice, providing a rigorous benchmark for zero-shot evaluation. Extensive zero-shot evaluation on synthetically degraded CVC-ClinicDB [1] and Kvasir-SEG [2] images demonstrates that our approach improves recall by 14 to 22 percentage points over YOLO alone, while precision remains within 0.7 points below to 1.7 points above the baseline. This combination of adaptive thresholding and cost-sensitive reinforcement learning achieves clinically aligned, open-world polyp detection with substantially fewer false negatives, thereby reducing the risk of missed precancerous polyps and improving patient outcomes.

9 January 2026 in TechRxiv
Verified
Preprint
Open access
ZBZiqian Bi
Ziqian Bi
XLXiaoming Li
Xiaoming Li
EGEnze Ge
Enze Ge
JSJunhao Song
Junhao Song
YLYiming Lei
Yiming Lei
BHBinhua Huang
Binhua Huang
This survey paper provides a comprehensive overview of the development strategies for agentic AI and autonomous agents, with a focus on Multi-Agent Reinforcement Learning (MARL) techniques. First, it examines the fundamentals of reinforcement learning and deep learning, emphasising their pivotal role in achieving groundbreaking results in diverse domains. Second, the scope of MARL is examined, including its applications in cyber defense, robotics, and autonomous vehicles, as well as its potential to enhance decision-making in dynamic and interactive environments. Third, it also delves into the challenges and research directions in MARL, including trustworthiness, safety, and robustness, and discusses the importance of integrating human factors into autonomous decision-making processes. Finally, it highlights the significance of data-centric frameworks for self-adaptive agents and actor-critic methods in MARL, while also addressing open challenges and future directions, such as scalability, non-stationarity, and explainability. The main objective of this survey paper is to provide a systematic understanding of the current state of agentic AI and autonomous agents development. Most significantly, this work contributes a comprehensive overview of MARL techniques, an examination of the ethical considerations and societal impacts of autonomous agents, and a discussion of the prospects and challenges in the field. We hope this paper can guide researchers, practitioners, and policymakers in the development and deployment of agentic AI and autonomous agents, with a focus on ensuring responsible and beneficial outcomes.

29 December 2025 in TechRxiv
Verified
Preprint
Open access
YZYichao Zhang
Yichao Zhang
JYJunjie Yang
Junjie Yang
ZBZiqian Bi
Ziqian Bi
JSJunhao Song
Junhao Song
BHBinhua Huang
Binhua Huang
Image segmentation forms the backbone of modern computer vision, yet its evolution from classical techniques to deep learning paradigms reveals both remarkable progress and persistent challenges. This work navigates the landscape of segmentation methodologies, examining how traditional approaches like thresholding and edge detection have given way to sophisticated neural architectures, while retaining relevance in specific contexts. We explore applications spanning medical diagnostics, autonomous navigation, and remote sensing, where segmentation quality directly impacts real-world outcomes. Our analysis reveals that despite advances in foundation models and transformer architectures, critical gaps remain. Evaluation metrics struggle to capture segmentation nuances, particularly for small structures and boundary precision. Weakly supervised methods promise reduced annotation burden but face stability issues. Through examining over 300 recent publications, we identify emerging patterns: the tension between model complexity and deployment efficiency, the challenge of domain adaptation, and the need for more interpretable solutions. Rather than simply cataloging techniques, we synthesize connections between seemingly disparate approaches, offering insights into why certain methods excel in specific domains. This perspective aims to guide researchers toward addressing the field's most pressing challenges while recognizing the continued value of classical methods alongside modern innovations.