Academic Experience
Nanjing University of Posts and Telecommunications (NUPT)
Advisor: Lei Chen
- Weighted average: 89.1/100.
- GPA: 3.91/5.0; WES-evaluated GPA: 3.82/4.0.
- TOEFL: 103. GRE: 336.
- Research interests: evidential deep learning, uncertainty quantification, and reinforcement learning.
University of Cambridge
- Short-term academic training focused on modern AI foundations and applications.
Yale University
- Incoming master's student in the Data Science track within Biostatistics.
Industry Experience
Algorithm Intern · Baidu
Improved a production dual-encoder retrieval system by systematically analyzing key failure modes and contributing across post-training strategy design, data construction, and offline evaluation.
- Systematically analyzed failure modes including term omission, term mismatch, semantic misunderstanding, and relevant-but-not-answerable results to guide retrieval optimization.
- Developed a hard-negative enhancement framework based on mis-retrieved samples, together with negative loss reweighting and staged optimization, improving average Recall by ~10.5%.
- Built an LLM-driven data augmentation framework with two major categories—q+u→q and u→q—covering four sub-pipelines for query–URL constrained semantic transfer and content-driven retrieval scenarios, yielding a further ~3.2% improvement in average Recall.
- Introduced an offline LLM + rule-based teacher scoring mechanism for fine-grained supervision over candidate URL sets, and leveraged it for post-training and listwise optimization of the dual-encoder, driving an additional ~6.0% gain in average Recall.
- Increased overall average Recall from ~60 to 72.5, corresponding to a cumulative relative improvement of ~20.8%, and established a full optimization loop spanning failure analysis, data construction, training paradigm refinement, and offline evaluation.
Algorithm Intern · HONOR
Built Generative Engine Optimization (GEO) strategy and evaluation infrastructure for HONOR 400 and Magic V5, connecting LLM-driven content generation with measurable gains in recommendation and citation visibility.
- Owned GEO strategy deployment and experiment-platform construction for HONOR 400 and Magic V5. Abstracted LLM web-search behavior (e.g., DeepSeek) into a two-stage pipeline—candidate retrieval and re-ranking—and analyzed the key drivers behind citation and recommendation using BM25, semantic similarity, and related signals.
- Proposed and independently implemented a Multi-Agent Consistency Controller: a generation agent drafted promotional copy, while a review agent automatically detected factual errors and exaggerated claims, returned structured feedback, and enabled multi-round self-correction to reduce hallucination and compliance risk while preserving factual consistency and brand tone at scale.
- Designed a GAN-like three-stage workflow spanning generation, adversarial posterior evaluation, and prior scoring. Built a multi-agent system for large-scale copy generation; wrote headless-browser scripts to query LLMs, collect Top-50 reference lists, and track Top-10 citation coverage and Top-50 candidate coverage; and constructed a query–document dataset with LLM Top-5 candidates as positives and traditional SEO documents as negatives.
- Fine-tuned DeepSeek-8B with LoRA to learn a scoring function for LLM citation probability. In offline evaluation, the model achieved AUC ≈ 0.88 for predicting whether a document would be used as a high-weight reference, Top-5 hit rate above 80%, and Spearman correlation of ~0.7 between prior scores and real online exposure.
- Using unified prior and posterior metrics to iteratively refine generation and content-selection strategies, increased the share of HONOR 400 / Magic V5 content proactively recommended or cited by LLMs by ~3× (key-query recommendation rate from ~20% to 70%+), maintained overall visibility above 65%, reduced weekly publication volume from ~100 pieces to 5–6, and translated the efficiency gain into roughly RMB 4 million in annual marketing cost savings.
Algorithm Intern · Bosch
Developed data-driven automation for Ethernet gateway tooling and protocol-level validation.
- Designed rule-based automated decision algorithms and built a gateway intelligent system to reduce system error rates.
- Implemented a core parsing engine that auto-generates C++ protocol parsing code and an automated test framework.
- Completed Python implementations of RTP/RTSP/RTCP and studied IEEE1588/gPTP for time synchronization.
- Optimized a lightweight protobuf library and improved accuracy from 98.4% to 100%.
Algorithm Intern · Industrial and Commercial Bank of China (ICBC)
Built ML-based credit and risk assessment prototypes with structured data pipelines.
- Integrated multi-source customer data and standardized preprocessing workflows for consistency and availability.
- Designed an intelligent loan issuance assessment model reaching ~72% accuracy.
- Developed an LSTM–Transformer time-series model achieving ~89% AUC and early risk identification.
Publications
Post-hoc Budgeted Uncertainty-Guided Topology Repair for Retinal Vessel Segmentation
A plug-and-play post-hoc repair framework that combines evidential uncertainty and topology-violation cues to perform budgeted local edits for structurally reliable retinal vessel segmentation.
Show abstract
Accurate retinal vessel segmentation is critical for quantitative ophthalmic analysis, yet modern segmenters can still produce structurally invalid masks such as broken vessels, spurious fragments, and small holes. We propose BUTR, a plug-and-play post-hoc framework that first constructs a compact region of interest by combining predictive uncertainty with explicit structure-violation signals under an edit budget, and then performs ROI-constrained logit refinement with a lightweight repair head. This design enables minimal, localized intervention rather than uncontrolled global post-processing. Experiments on DRIVE, CHASE_DB1, and STARE show that BUTR consistently improves topology-sensitive metrics while matching or improving standard overlap and boundary metrics.
Meta-Policy Controller for Dynamic Uncertainty Calibration in Evidential Deep Learning
A bi-level meta-policy framework that dynamically optimizes KL regularization and learnable Dirichlet priors for robust uncertainty calibration.
Show abstract
Traditional evidential deep learning methods rely on static hyperparameters for uncertainty calibration, limiting adaptability under dynamic training states and shifting data distributions. We propose the Meta-Policy Controller (MPC), a dynamic meta-learning framework that adjusts the KL divergence coefficient and class-specific Dirichlet prior strengths through bi-level optimization. In the inner loop, model parameters are updated with a dynamically configured loss function. In the outer loop, a policy network learns to balance multi-objective rewards that consider both predictive accuracy and uncertainty quality. Unlike fixed uniform priors, our learnable Dirichlet prior flexibly adapts to class distributions and training dynamics. Experiments show improved reliability and calibration across classification, medical segmentation, and OOD detection tasks.
Evidential U-KAN for Trustworthy Medical Image Segmentation
A progressive uncertainty-guided attention framework that strengthens boundary awareness and preserves semantic evidence for reliable medical segmentation.
Show abstract
Trustworthy medical image segmentation aims to deliver accurate and reliable results for clinical decision-making. Most existing methods adopt the evidential deep learning (EDL) paradigm due to its computational efficiency and theoretical robustness. However, EDL-based methods often neglect leveraging uncertainty maps rich in attention cues to refine ambiguous boundary segmentation. To address this, we propose a progressive evidence uncertainty-guided attention (PEUA) mechanism to guide the model to focus on hard regions. PEUA progressively refines attention using uncertainty maps while employing low-rank learning to denoise attention weights. We further introduce a semantic-preserving evidence learning (SAEL) strategy that retains critical semantics through a semantic-smooth evidence generator and fidelity-enhancing regularization. By embedding PEUA and SAEL into the state-of-the-art U-KAN, we propose Evidential U-KAN for trustworthy medical image segmentation. Extensive experiments across four datasets demonstrate superior accuracy and reliability over competing methods.
Projects
CP31 Ticket Snatcher
An automation toolkit designed to improve the success rate of convention ticket purchasing under heavy competition.
Show more
Built for real-world ticketing bottlenecks such as queue congestion and rapid sell-out scenarios.
Designed as a practical automation toolkit for anime convention ticket purchasing under extreme contention.
BUTR
A plug-and-play post-hoc framework that combines evidential uncertainty and structure-violation cues for budgeted local repair in retinal vessel segmentation.
Show more
Constructs a compact ROI under an edit budget by fusing uncertainty with topology-violation signals, then applies lightweight ROI-constrained logit refinement for minimal and targeted structural correction. Serves as the codebase accompanying the MICCAI 2026 under-review manuscript.
Meta-Policy Controller (MPC)
A bi-level meta-policy system that learns dynamic uncertainty calibration strategies for evidential models.
Show more
Uses a policy network to optimize KL coefficients and class-wise Dirichlet prior strengths. Aims to unify accuracy and uncertainty quality under a single adaptive optimization framework. Serves as the codebase accompanying the arXiv/CVPR under-review manuscript.
Evidential U-KAN
A trustworthy medical segmentation framework integrating progressive uncertainty-guided attention and semantic-preserving evidence learning.
Show more
Combines PEUA and SAEL to enhance boundary precision and preserve clinically meaningful semantics. Demonstrates strong accuracy–reliability trade-offs across multiple public polyp datasets. Serves as the codebase accompanying the arXiv preprint.
Life Updates
Mahjong Soul Rank-Up
Reached Saint rank in 4-player ranked matches.
StarCraft II Milestone
Achieved Master League as Zerg.
Fun Stuff
Random Anime Buddy
A tiny gacha that refreshes your mood. New buddy on every reload.
Loading...