Research Assistant · Advised by Prof. Lianhui Qin

Code4Scene: Evaluating Coding Agents in 3D Environments

• Co-first author; led the evaluation framework. Designed and implemented six verifier modules covering candidate integrity, physical safety, semantic verification, detailed and overall alignment, and ground-truth comparison.
• Developed task-specific criteria: requirement-level checks and whole-scene alignment for construction; separate measures of target recovery and preservation of unrelated content for editing.
• Built the verification workflow, combining Unreal Engine state measurements, support and collision checks, camera planning, targeted visual evidence, VLM judging, and score aggregation, with parallel evaluation and human verification.
• Evaluated 14 model configurations across construction and indoor/outdoor editing. Joint analysis of scores and trajectories exposed weak spatial composition and unintended scene changes in 35.8% of complete target recoveries on public editing tasks.
• Audited execution trajectories and identified five answer-exposure pathways involving answer metadata, binary ground-truth map decoding, UE/MCP backends, cross-case inputs, and original asset-pack scenes.

Project Lead, MemTensor · Co-advised by Prof. Muning Wen, APEX Lab

Reinforcement Learning for Reasoning under Training–Inference Mismatch

• Independently lead PPO-based post-training research on training–inference mismatch, using high-precision feedback to guide low-bit reasoning policies.
• Developed Quantization-Counterfactual Advantage (QCA): rank BF16 alternatives against low-bit actions at the same prefix, guided by positive critic-estimated benefits.
• Developed Advantage-Weighted Consistency (AWC), a teacher-to-policy KL objective that uses positive advantages to prioritize BF16 feedback during PPO updates.
• Implemented QCA and AWC in verl/vLLM; evaluate their effects on DAPO, MATH, and GSM8K against matched BF16 PPO and QaRL baselines.

Research Contributor

• Designed the MicroDicom chest-X-ray annotation adjudication task using nine NIH CXR DICOM cases, structured outputs, and hidden case-level references.
• Implemented file/schema validation and IoU-based localization checks for verifiable medical-software agent evaluation.

Research Contributor · Advised by Prof. Siheng Chen

• Extended the text-only MA-RAG baseline to process text and images for G²-Reader; integrated and debugged LightRAG for controlled comparisons.
• Completed three full runs over 2,271 VisDoMBench samples across five domains, using Qwen3-VL-32B-Instruct and a fixed top-5 retrieval budget.

Summer Research Intern · Advised by Prof. Mingkai Dong

Course Project · Advanced Data Structure, Shanghai Jiao Tong University

• Implemented a C++ LSM-tree store with skip-list memtables, SSTable compaction, and Bloom filters; added embedding-based search, HNSW indexing, and persistence.
• Parallelized top-k search, reducing mean query latency from 76.34 ms to 7.94 ms on 49,998 precomputed vectors (1 vs. 28 threads, 128 queries).