Artificial intelligence · Research & engineering

Yipeng Chen.

I study how intelligent systems
learn, remember, and use evidence.

I’m a master’s student in Artificial Intelligence at Nanyang Technological University, with a background in Physics from UCL. My work spans continual learning for perception, retrieval agents, and the evaluation of language models.

Explore my work GitHub

01 / Selected work

Research & projects

Journal article

IEEE Sensors Journal 2026

Continual Learning for Automotive
Radar Semantic Segmentation

Yipeng Chen, Ziwei Zhang, Jun Liu

How can a perception model learn new object categories while retaining what it already knows? This work studies that question in sparse automotive radar point clouds, using a RadarScenes benchmark with six classes across three incremental tasks.

A frozen previous model guides the current model through knowledge distillation. Comparisons across four architectures examine the balance between learning new classes and retaining old ones.

Read the paper Code

Independent project

Language model agents 2026

Personal RAG

Retrieval, evidence, and learning when to stop.

I built a local agent that combines keyword and semantic search, reranking, memory, and multi-step tool use. A separate training and evaluation pipeline studies how supervised fine-tuning and reinforcement learning affect its retrieval decisions.

Latest experiment · 5 September 2026

Supervised training increased exact-answer accuracy from 43.9% to 48.0%. The subsequent reinforcement-learning run reduced retrieval calls, but accuracy fell to 45.6%.

Same 1,000 HotpotQA development questions; one training and evaluation seed. These results measure a specific experiment, not general performance.

Code & evaluation Experiment report

Research project

Language model evaluation 2025

Learning from imbalanced reviews

I compared supervised learning, learning from examples in the prompt, and efficient fine-tuning for three-way Chinese sentiment classification. The study examines class imbalance, ambiguous neutral reviews, and how the choice of examples affects model reliability.

Chinese RoBERTa · Qwen3 · Fixed data splits · Paired statistical comparisons

Ongoing exploration

From perception to world models

I’m exploring how action-conditioned world models can be evaluated through controlled simulator comparisons, alongside agent-based models of collective foraging. These are exploratory projects; spatial world modelling is a direction I want to develop further.

02 / Experience

Building systems in practice

ByteDance

AI Agent Intern

July 2026 – present

I work on the reliability and efficiency of long-running AI agent conversations, including context compression, token-count caching, and asynchronous preparation of conversation summaries.

I also examine how automated presentation evaluation agrees with human judgement, and what evidence an evaluator needs to make useful decisions.

03 / Background

Physics to artificial intelligence

My physics background shapes the way I approach AI: start with a clear question, build a controlled experiment, and check what the evidence actually supports.

I’m interested in systems that keep learning over time, retrieve useful knowledge, and connect perception with action. Across my projects, I focus on reproducibility and on understanding where a method succeeds or fails.

2025 – present

Nanyang Technological University

Master’s degree · Artificial Intelligence (PSI)

2022 – 2025

University College London

Bachelor’s degree · Physics

Upper Second-Class Honours

04 / Contact

Let’s talk research.

For research conversations and collaboration enquiries.

yipeng003@e.ntu.edu.sg

GitHub LinkedIn