Yinghui He, Simran Kaur, Adithya Bhaskar, Yongjin Yang, Jiarui Liu, Narutatsu Ri, Liam Fowl, Abhishek Panigrahi, Danqi Chen, Sanjeev Arora
Download Paper
Arxiv Preprint, 2026
SD-Zero improves reasoning by having one model generate answers and then revise its own mistakes to turn sparse outcome feedback into dense token-level training signals, boosting 10%+ performance without an external teacher.
Yinghui He*, Abhishek Panigrahi*, Yong Lin, Sanjeev Arora
Download Paper
ICLR 2026, 2025
We introduce a new training paradigm, Skill-Targeted Adaptive Training (STAT), which offers a principled path to overcoming SFT saturation and advancing generalization in LLMs.
Yinghui He, Abhishek Panigrahi, Yong Lin, Sanjeev Arora
Download Paper
COLM 2025, 2025
Kids improve when a good teacher offers adaptive, targeted feedback. Can a small LLM benefit if a large LLM provide helpful feedback, in-context?? Naive ideas fail here. We propose AdaptMI: adaptive, skill-based in-context supervision that boosts 1B models by 6% on challenging math tasks.
Jiahao Qiu*, Yinghui He*, Xinzhe Juan*, Yiming Wang, Yuhan Liu, Zixin Yao, Yue Wu, Xun Jiang, Ling Yang, Mengdi Wang
Download Paper
EMNLP 2025 Main Conference, 2025
Can AI Be Blamed for a Teen’s Suicide? Do AI Chatbots encourage suicide? 🧒📱What if your teen’s favorite AI character crossed the line? 💔 A 14-year-old boy in Florida took his own life after forming a deep bond with an AI character on http://Character.AI. The AI chatbot — modeled after a Game of Thrones persona — reportedly discussed his suicidal thoughts and encouraged these dangerous ideas. ⚠️AI can help, but unfortunately, it can also harm.
Xi Ye, Fangcong Yin*, Yinghui He*, Joie Zhang*, Howard Yen*, Tianyu Gao, Greg Durrett, Danqi Chen
Download Paper
COLM 2025, 2025
“🤔Now most LLMs have >= 128K context sizes, but are they good at generating long outputs, such as writing 8K token chain-of-thought for a planning problem？ 🔔Introducing LongProc (Long Procedural Generation), a new benchmark with 6 diverse tasks that challenge LLMs to synthesize highly dispersed information and generate long, structured outputs.
Yinghui He, Yufan Wu, Yilin Jia, Rada Mihalcea, Yulong Chen, and Naihao Deng
Download Paper
Findings of EMNLP 2023; ICML 2023 Workshop on Theory of Mind in Communicating Agents, 2023
“They don’t know that we know they know we know” 🤯 — Does GPT-4 have Higher-Order Theory of Mind? Introducing 👋 Hi-ToM: a benchmark pushing LLMs to their limits in higher-order ToM (3rd order & beyond). LLMs’ performance declines drastically to near 0 📉 on 3rd and 4th!
Jianhao Ma, Rui Ray Chen, Yinghui He, Salar Fattahi, and Wei Hu
Download Paper
ICLR 2024 Workshop on Bridging the Gap Between Practice and Theory in Deep Learning, 2023