cv

Master

Nanjing University

Artificial Intelligence

Bachelor

Nanjing University

Computer Science

Lookahead Tree-Based Rollouts for Enhanced Trajectory-Level Exploration in Reinforcement Learning with Verifiable Rewards

ICLR 2026

Introduce a new tree-based sampling strategy to enhance group rollout diversity through branching on high-uncertainty tokens, simulating and pruning branches that exhibits prolonged similarity.

RealBench: A Chinese Multi-image Understanding Benchmark Close to Real-world Scenarios

EMNLP 2025 (findings)

Construct a new multi-image benchmark for multimodal LLMs from real user inputs on social platform.

Maximizing the Effectiveness of Larger BERT Models for Compression

ACL 2025 (main)

Propose a new KD method for BERT model by maximizing linear difference between selected layers, enabling better performance with larger teachers.

GePBench: Evaluating Fundamental Geometric Perception for Multimodal Large Language Models

Submitted to ICML 2026

Propose a novel large-scale multimodal benchmark on geometric perception, identifying the limitations of current models and exploring the impact of geometric perception on high level tasks.

AnyPrefer: An Automatic Framework for Preference Data Synthesis

ICLR 2025

Design a framework that can automatically generate preference data for multiple scenarios without requiring manual annotation.

AlignGPT: Multi-modal Large Language Models with Adaptive Alignment Capability

Preprint

Create distinct alignment vectors for differently aligned text-image pairs during pretraining, and allocate them to various subtasks in finetuning and inference.

EFUF: Efficient Fine-grained Unlearning Framework for Mitigating Hallucinations in Multimodal Large Language Models

EMNLP’ 2024

Leverage external expert knowledge to reinforce the alignment between language and vision, thereby reducing multimodal hallucinations with no manually annotated data and minial computational resources.

DRIN: Dynamic Relation Interactive Network for Multimodal Entity Linking

ACMMM’ 2023

Explicitly model four types of alignment between multimodal mentions and entities and uses a dynamic Graph Convolutional Network to automatically select appropriate alignment relations for different input samples.

Learning from Different text-image Pairs: A Relation-enhanced Graph Convolutional Network for Multimodal NER

ACMMM’ 2022

Utilize Graph Neural Networks to capture external matching relationships across different text-image pairs.

National Scholarship

Ministry of Education of the People's Republic of China

Tencent Scholarship

Tencent

Outstanding Student Model of Nanjing University

Nanjing University

Outstanding Graduate of Nanjing University

Nanjing University

Tencent Scholarship

Tencent

Internship: Algorithm Engineer

Xiaohongshu Co., Ltd.

(In progress) Researching AI search summarization.

Teaching Assistant: Compilers - Principles, Techniques, and Tools

School of Artificial Intelligence, Nanjing University

Assisted in lectures, graded assignments and final exams.

Research Intern led by Xiang Ren (Online)

University of Southern California

Designed tree-based group sampling strategy for RLVR.

Research Intern led by Huaxiu Yao (Online)

University of North Carolina, Chapel Hill

Implemented data synthesis algorithm with feedback on target vision-language model (AnyPrefer).

Internship: Software R&D Engineer

INFLY Tech (Shanghai) Co., Ltd.

Implementing preference alignment algorithms RLHF/PPO and its variations DPO, RRHF; Training a BLOOM model with billion-level parameters using the Deepspeed and Megatron-LM frameworks, experimenting different algorithms.

Internship: Software R&D Engineer

Huawei Technologies Co., Ltd.

Integrating the open-source SOTA model Tacotron2 with a proprietary model optimized for handling Chinese spoken language pitch and rhythm.