I'm an undergraduate student in Software Engineering at Xidian University. My research interests lie in multimodal understanding, memory, and reliable reasoning for intelligent agents.
I am seeking PhD opportunities starting in Fall 2027, with a particular interest in video understanding, multimodal memory, and embodied AI.
- Streaming video understanding: understanding continuous visual observations and retaining useful information from long video histories.
- Multimodal memory: selecting, compressing, and retrieving task-relevant historical information for reasoning and decision-making.
- Embodied AI: studying and evaluating vision-language-action (VLA) models for robotic tasks.
- Reliable reasoning: evaluating the faithfulness of model reasoning and exploring reinforcement learning for more reliable behavior.
My work explores how multimodal models can manage historical information during streaming video understanding. I am interested in visual information compression, KV-cache management, and event-level memory, with the goal of preserving useful context for later queries and reasoning.
My experience includes evaluation development and quantization experiments for VLA models, with work involving OpenVLA / OpenVLA-OFT, LIBERO, and SimplerEnv.
- Working on OpenVLA-OFT quantization and simulation-based policy evaluation.
- Studying how model compression affects task success and inference efficiency.
- Studying the source code of π0 and π0.5, focusing on the information flow from multimodal inputs to action generation.
Related working repository: Quant, based on the OpenVLA-OFT codebase.
I am a co-first author of a research manuscript on faithful reasoning using spatio-temporal latent dynamics.
My contributions focus on experiments, including reinforcement-learning runs, reward ablation studies, and analysis of model behavior and results. This experience has motivated my interest in how to evaluate whether a model's reasoning faithfully reflects its decision process.
- FMD Python Reimplementation — a project on implementing Feature Mode Decomposition in Python.
Python · PyTorch · CUDA · Linux · Git
Research frameworks and benchmarks I have worked with or studied include OpenVLA, OpenVLA-OFT, LIBERO, and SimplerEnv.
- Email: 23049200290@stu.xidian.edu.cn
- GitHub: WentongZhao
I welcome conversations about research collaborations and PhD opportunities in multimodal learning, video understanding, and embodied AI.