Jiachen Liu

On AI-native research infrastructure & AI co-scientists, I founded Orchestra Research — a unified substrate where hypothesis generation, experiment execution, analysis, and writeup share one memory and one state, with a kernel that orchestrates AI co-scientists across the full research lifecycle. I co-lead Curie, the first AI-agent framework for rigorous automated scientific experimentation, and co-created EXP-Bench (ICLR 2026), the benchmark measuring whether agents can conduct AI research experiments end-to-end.

On the broader AI-native research ecosystem, I argue that the academic paper itself must evolve for an agent-driven world. In The Last Human-Written Paper: Agent-Native Research Artifacts, I propose ARA — a machine-executable research package that replaces narrative papers with structured layers (scientific logic, executable code, exploration graph, raw evidence), eliminating the storytelling and engineering taxes that block AI from reproducing and extending science.

Underneath, I build the systems layer that lets this ecosystem run at scale. Through intern, part-time, and full-time roles at Meta, I contributed to the pre-training and post-training systems powering one of the largest LLM training stacks in the world. My broader work spans efficient LLM serving (Andes), frontier evaluation benchmarks (Humanity's Last Exam, Nature 2026), and energy-aware LLM serving (ML.ENERGY Benchmark, NeurIPS 2025 Spotlight).