Research Interests

I am committed to long-term, in-depth research on pretraining data reformulation, including:

• Long-context Data Reformulation — Synthesizing long-context training data to enhance long-range modeling of LLMs.
• Knowledge Data Reformulation — Creating diverse entry points (narratives, contexts, co-occurring entities) for the same knowledge, so that long-tail facts are no longer long-tail in pretraining corpora.
• Long Agentic Data Reformulation — My latest focus. Agentic post-training struggles to generalize; the key is agentic pretraining. Since environments (e.g., GPU clusters, distributed systems, commercial platforms, etc.) are the hardest to scale, pretraining should lower the activation cost of post-training — making correct paths easier to sample.

I also explore scalable oversight — what do we do when humans can no longer effectively supervise AI? Welcome to discuss or join me (wuxing@iie.ac.cn).

Work Experience

News

• 🎉🎉🎉 NextLong and EntropyLong validated on Tencent HY 3.0, Zhipu GLM-5, and XiaoHongShu Dots, demonstrating consistent improvements across production LLMs.
• 2026.01 EntropyLong accepted as ICLR 2026 conference paper!
• 2025.11 LiteLong accepted as AAAI 2026 conference paper!
• 2025.09 LongMagpie accepted as NeurIPS 2025 conference paper!
• 2025.05 NextLong accepted as ICML 2025 conference paper!
• 2025.01 Quest accepted as ICLR 2025 conference paper!

Selected Publications

For a full list of publications, please visit my Google Scholar profile.

Selected Blogs