HippoCamp: Benchmarking Contextual Agents on Personal Computers
Can an agent make sense of a life’s worth of files? A device-scale benchmark with 2,000+ real files, 581 questions, and step-wise failure diagnosis.
FIRST AUTHOR · RESEARCH LEAD
Hi, I’m Zhe.
Robots and agents
for
everyday life.
I work on versatile robots and multimodal agents, focusing on long-horizon tasks through policy learning, memory, and realistic evaluation.
M.Sc. in AI · NTU MMLabEmbodied AI Intern · SenseTime
Hello, there.
Drag to turn. Say hi to make a friend.
Research focusEmbodied intelligenceLong-horizon memoryMultimodal agents
hello from Singapore ☀
A LITTLE ABOUT ME
I’m a master’s student in Artificial Intelligence at Nanyang Technological University, working with Prof. Chen Change Loy and Prof. Ziwei Liu at MMLab. I also work on embodied AI at SenseTime.
My work connects memory, policy learning, and realistic evaluation—from personal computers to simulated homes and real robots. Previously, I graduated with First Class Honours from CUHK and worked with Prof. Michael R. Lyu on LLM memorization and privacy.
PAPERS & BENCHMARKS
Research & publications
Memory, reasoning, and learning to act in the world.
Has GPT-6-Astra Achieved Embodied Intelligence?
Probing embodied competence and practical limitations, then asking what a next-generation benchmark should measure in a world that keeps moving.
CO-FIRST AUTHOR · BENCHMARK PROTOTYPE & BEHAVIOR-1K
EgoTools: Towards Tool-Centric Reasoning in Real-World Egocentric Videos
Understanding why and how people use tools: 100 hours of egocentric video and a 1,000-question diagnostic benchmark.
DATA COLLECTION · ANNOTATION PLATFORM & CURATION
InnerFeel: Benchmarking Multi-Perspective Emotion Report Prediction in Egocentric Social Interactions
Studying emotion understanding as per-user belief tracking, with memory isolation and evaluation across different perspectives.
RESEARCH FRAMING · EVALUATION DESIGN · VISUALIZATION
Manuscript under review · Public link to follow.
FileGram: Grounding Agent Personalization in File-System Behavioral Traces
Grounding personalized memory in what people do with their files, through a simulation engine, diagnostic benchmark, and memory architecture.
DATA ADAPTATION · BEHAVIOR SIMULATION ENGINE
Understanding Secret Leakage Risks in Code LLMs: A Tokenization Perspective
Investigating how tokenization turns high-character-entropy secrets into predictable token sequences, and why this increases leakage risks.
Data Compressibility Quantifies LLM Memorization
A shift from instance-level to set-level entropy uncovers a near-linear relationship between data compressibility and LLM memorization.
UNDERGRADUATE THESIS · PIPELINE LEAD
RAUIE: A Relation-Augmented Document-Level Event Extraction Model Based on UIE
Using relation-aware prompts to improve document-level event extraction.
* Equal contribution.
WORK IN PROGRESS
On my deskright now
A tiny practice runScroll to move ↓
Drag to turn · slide to replay
A playful illustration, not a research demo.
01 MEMORY
Memory for world
action models.
Connecting what a robot remembers to what it does next.
I assist in designing dual memory, co-develop RMBench++, and build the real-robot pipeline for data acquisition, model deployment, and robot control.
A closer look
- Endogenous memory compresses older history into gist tokens.
- Exogenous memory records where predictions and new observations disagree.
- Evaluation combines training, simulation, and real-robot experiments.
With Dr. Zhonghua Wu
02 ACTION
Learning actions
from video.
Taking learned actions to real robots.
I integrated register-token latent actions into π0.5 and ran closed-loop manipulation evaluation in RoboTwin.
A closer look
- Policy integration connects native latent actions with downstream robot learning.
- Real-robot experiments cover teleoperation, data formatting, model inference and deployment, and success-rate evaluation.
With Dr. Qingyi Tao
EXPERIENCE & EDUCATION
Along the way
SenseTime
Algorithm Intern · Embodied AI
Singapore
SenseTime
Algorithm Intern · Embodied AI
SingaporeCore contributor to the GPT-6-Astra technical report, including the benchmark prototype and BEHAVIOR-1K evaluation. Current projects study memory for world action models and native latent action models.
MMLab · NTU
Research · Multimodal agents & video understanding
Advised by Chen Change Loy & Ziwei Liu
MMLab · NTU
Research · Multimodal agents & video understanding
Advised by Chen Change Loy & Ziwei LiuLed HippoCamp as first author and research lead. Contributed to FileGram, EgoTools, and InnerFeel, spanning personalized agents, tool-centric video reasoning, and multi-perspective emotion understanding.
Nanyang Technological University
M.Sc. in Artificial Intelligence · GPA 5.0 / 5.0
MMLab · Chen Change Loy & Ziwei Liu
Nanyang Technological University
M.Sc. in Artificial Intelligence · GPA 5.0 / 5.0
MMLab · Chen Change Loy & Ziwei LiuMMLab@NTU, advised by Prof. Chen Change Loy and Prof. Ziwei Liu. Research on versatile robots and multimodal agents for everyday tasks.
CUHK · ARISE Lab
Undergraduate thesis · LLM Memorization & Privacy
Advised by Michael R. Lyu
CUHK · ARISE Lab
Undergraduate thesis · LLM Memorization & Privacy
Advised by Michael R. LyuUndergraduate thesis research on the relationship between entropy and LLM memorization, including a 400K+ sample inference pipeline and investigation of code-secret leakage.
The Chinese University of Hong Kong
B.Sc. in Computer Science · First Class Honours
Major GPA 3.80 / 4.0
The Chinese University of Hong Kong
B.Sc. in Computer Science · First Class Honours
Major GPA 3.80 / 4.0B.Sc. in Computer Science, First Class Honours. Major GPA: 3.80/4.0. Dean’s List on four occasions.
NOTES TO SELF01 / 02
What I want
wants me more.
Zhe
