Yun Chen

Text world models can transform LLM-based agents from reactive responders to proactive planners, fundamentally changing how they interact with complex environments.

Overweighting easy-to-reconstruct features in generative CTR models is leaving performance on the table, especially for cold-start and long-tail users.

Get 5.3% more clicks by intelligently scaling your CTR model's inference depth only when it's uncertain, without retraining or increasing worst-case latency.

Forget painstakingly optimizing each urban scene individually: GenRe distills diffusion-based generative priors into 3D Gaussian representations, fixing reconstruction deficiencies in minutes and generalizing to challenging viewpoints.

Forget about re-balancing losses – gradient geometry is the key to unlearning in LLMs without sacrificing retention.

PPO can be made sample-efficient and stable for long-horizon reasoning in LLMs by treating the problem as a sequence-level contextual bandit, sidestepping the need for computationally expensive multi-sampling.

Finally, a diffusion model lets you puppeteer multiple objects in a video with just text prompts, opening the door to complex scene editing.