Haonan He 何浩楠

I am a student at the University of Science and Technology of China (USTC), working at the intersection of large language models, parameter-efficient fine-tuning, and AI for science. My research focuses on making LLMs more efficient to train and adapt, while pushing their scientific reasoning capabilities to Olympiad-level frontiers.

I have first-authored works on gradient-driven adaptive low-rank adaptation (GoRA, NeurIPS 2025), a unified study of LoRA variants (LoRAFactory), and a multi-omics instruction dataset for LLMs (Biology-Instructions, EMNLP 2025 Findings). I also contributed to physics-reasoning models that achieved gold-medal-level performance at the International Physics Olympiad (IPhO).

On the RL and distillation side, I contributed to P1 and P1-VL (gold-medal performance at IPhO 2025), first-authored SimpleOPD (+21.2 on ProofBench, surpassing Gemini-2.5-Pro), and contributed to Agents-A1 and ExoMind. Our open-sourced models have been downloaded 450K+ times on Hugging Face.