LLM post-training · Reinforcement learning · Multimodal systems
Senior Machine Learning Engineer at Meta · London
I build and evaluate machine learning systems. My research projects span prompt selection for LLM post-training, credit-assignment diagnostics, and offline reinforcement learning. I also implement and profile multimodal training systems.
I’m interested in how training choices and evaluation design affect what we can conclude about a model’s improvements.
Exact-oracle diagnostics that separate identifying helpful actions from inducing the intended policy gradient.
My contribution. Built a CPU benchmark for turn-level credit assignment and compared advantage computations across RL frameworks.
Investigating whether lightweight probes can select informative prompts before spending compute on full rollouts.
My contribution. Contributed post-training experiments and coding evaluation on OctoThinker-1B, comparing probe-guided selection with GRPO, DAPO, and LILO baselines.
Implementing a VLM and investigating suspicious scaling measurements across 2, 4, and 8 V100 GPUs.
My contribution. Built the image-token insertion and training pipeline using SigLIP-2, Qwen2.5, and a projector; profiled FSDP communication and tested forward prefetching.
At Meta, I work on production ML in Integrity, including LLM post-training, model distillation, and automated labeling. Previously, I built ML systems at Farfetch and Hypotenuse AI.
I hold an MSc in Computer Science from NYU and first-class honours degrees in Statistics and Computer Science from UNSW.
I welcome conversations about post-training, evaluation, and multimodal learning.
ng.yiyan314@gmail.com ↗