Research Scientist · ByteDance
Research on intelligent systems that model, reason about, and act in the world — spanning weather forecasting, generative modeling, and real-world agents.
WeatherGenerative modelingAgents & LLMs
Scholar snapshot June 2026
01 / Selected work
Representative work across physical forecasting, agent evaluation, and LLM performance scaling.
Weather intelligence · Nature 2023
A generative nowcasting system that combines physical-evolution priors with data-driven learning for large-area, short-term extreme precipitation prediction.
Agent evaluation · arXiv 2025
A profession-aligned evaluation suite for measuring real-world productivity scaling and market readiness of AI agents.
LLM scaling · ICLR 2026
A Clustering-On-Difficulty framework that groups tasks by their scaling behavior and uses predictable subsets to forecast downstream LLM performance before large-scale pretraining.
02 / Publications
Ten papers across weather intelligence, generative modeling, multimodal reasoning, scaling, and agents.
Open the filterable archive01
arXiv · Agents · LLMs
02
arXiv · Visual reasoning
03
arXiv · Agent benchmark
04
05
arXiv · Reasoning · RL
06
arXiv · Agents · Scaling
07
arXiv · LLMs · Scaling
08
arXiv · LLMs · Scaling
09
ICML 2024 · Diffusion
10
Nature 2023 · AI for weather
Citation counts are a dated Google Scholar snapshot from June 2026.
03 / Profile
Kaiyuan Chen is a Research Scientist at ByteDance. His work spans forecasting, generative modeling, multimodal reasoning, evaluation, and real-world agents.
Education
2021 — 2024
Machine Learning Lab, School of Software
Tsinghua University
2017 — 2021
Department of Mathematical Sciences
Tsinghua University