Zihao He

Research Scientist @Meta
LLMs, GenAI, LLM-as-a-Judge, Alignment
Previously PhD @USC

[Yuchen Jin @Yuchenj_UW]

DeepSeek is the GOAT. 🐳

They just published DSpark, a new speculative decoding method that boosts throughput by 51% to 400%.

They also open-sourced DeepSpec, the training framework behind it.

This is the real open AI.

[Zihao He @ZihaoHe95]

@ZeyuanAllenZhu Meta's loss

[Yuandong Tian @tydsh]

Several of my team members + myself are impacted by this layoff today. Welcome to connect :)

[Zihao He @ZihaoHe95]

Our team (Monetization Generative AI @Meta) is seeking research scientist interns to work on multi-modal LLM-as-A-Judge, reward modeling, and other relevant topics. PhD candidates with experience in VLMs and reinforcement learning (such as DPO, PPO, and GRPO) are preferred.

Job post: https://t.co/Rlwlc1TovB. Please do NOT apply directly.

The start date is in Spring (no earlier than March) or Summer 2026. If interested, please fill out this form: https://t.co/qWuk02oCVc. Thank you! We will reach out if there is a match.

[Zihao He @ZihaoHe95]

@xinformatics @openreviewnet same

[Kai Chen @kaichen23]

Our benchmark paper "STEER-Bench" was accepted by #EMNLP2025 Main Conference🥳. Huge thanks to my collaborators!

[Zihao He @ZihaoHe95]

Finally!!!

[OpenAI @OpenAI]

Our open models are here.
Both of them.

https://t.co/9tFxefOXcg

[Avi Chawla @_avichawla]

KV caching in LLMs, clearly explained (with visuals):

[Peng Qi @qi2peng2]

Seven years ago, I co-led a paper called 𝗛𝗼𝘁𝗽𝗼𝘁𝗤𝗔 that has motivated and facilitated many #AI #Agents research works since. Today, I'm asking that you stop using HotpotQA blindly for agents research in 2025 and beyond.

In my new blog post, I revisit the brief history of 𝗛𝗼𝘁𝗽𝗼𝘁𝗤𝗔, why it defined a good problem in its historical context, some of my own research attempts at solving this problem that I believe are on the critical path to autonomous agents, and what's changed today to make the dataset less relevant in its original form. I also reflect on the possible paths forward for AI agents research, and share a skeleton of a research proposal if I ever find the time / need to write one from my perspective today: https://t.co/K2dFHE4P9j

[Zihao He @ZihaoHe95]

Glad to be part of this very promising project!

[CLS @ChengleiSi]

Are AI scientists already better than human researchers?

We recruited 43 PhD students to spend 3 months executing research ideas proposed by an LLM agent vs human experts.

Main finding: LLM ideas result in worse projects than human ideas.

[Zihao He @ZihaoHe95]

Models can follow instructions—but can they speak like a community?

STEER-BENCH shows that even top LLMs fall short on norm alignment across diverse online communities.

Cultural steerability is still an open challenge. 👇

[Kai Chen @kaichen23]

🤔How well do LLMs adapt to different norms?
🧵We introduce STEER-BENCH, a benchmark for assessing steerability in LLMs.
📉 Human: 81% | Top LLM: ~65%
🚨 Norm alignment ≠ solved.
📄 Paper: https://t.co/YG4uDp37Zk
@ZihaoHe95 @taiwei_shi @KristinaLerman

[Zihao He @ZihaoHe95]

@YuYang_i @OpenAI Congrats!

[Zihao He @ZihaoHe95]

Check this work out! 🤗
We found that smoothed labels provided by a teacher model can mitigate the hallucinations from a student model! 🧐

[Hieu Nguyen @hieuhocnlp]

I'm excited to announce that my first-author research, "Smoothing Out Hallucinations: Mitigating LLM Hallucination with Smoothed Knowledge Distillation", is now available as a preprint!

https://t.co/TxSnfZbzhZ

[Zihao He @ZihaoHe95]

@muhao_chen Thank you! I'm joining Meta

[Zihao He @ZihaoHe95]

🚀 Officially Dr. Zihao He! 🎉

I’m beyond excited to share that I have successfully defended my PhD thesis, “Aligning Large Language Models with Human Perspectives” at USC! This marks the culmination of an incredible journey—years of research, collaboration, and growth.

Throughout my PhD, I’ve explored how we can align LLMs to better reflect human perspectives, evaluate alignment beyond simple stance replication, and investigate the ethical risks of ideological manipulation. It’s been a challenging yet deeply rewarding experience, and I’m grateful to have worked on such an impactful topic at the intersection of AI, NLP, and society.

A heartfelt thank you to my advisor Kristina Lerman, my committee members Emilio Ferrara and Marlon Twyman, and everyone who has supported me—mentors, collaborators, friends, and family. Your guidance, encouragement, and thought-provoking discussions have made this journey so much more meaningful.

As I step into the next chapter, I’m excited about the opportunities ahead in AI research and its real-world implications. If you’re working on LLMs, AI alignment, or NLP applications, I’d love to connect and exchange ideas!

Onward to new challenges! 🚀 #PhDDefense #AIAlignment #LLMs #NLP

[TikTok Policy @TikTokPolicy]

STATEMENT FROM TIKTOK:

In agreement with our service providers, TikTok is in the process of restoring service. We thank President Trump for providing the necessary clarity and assurance to our service providers that they will face no penalties providing TikTok to over 170 million Americans and allowing over 7 million small businesses to thrive.

It’s a strong stand for the First Amendment and against arbitrary censorship. We will work with President Trump on a long-term solution that keeps TikTok in the United States.

[Jiao Sun @sunjiao123sun_]

Mitigating racial bias from LLMs is a lot easier than removing it from humans!

Can’t believe this happened at the best AI conference @NeurIPSConf

We have ethical reviews for authors, but missed it for invited speakers? 😡

[USC ISI @USC_ISI]

🙌 WTG to ALL🏆 the @USCViterbi students at #EMNLP2024 including ISIers @rebeccadorn_ @BrihiJ @AlexanderSpangh @TenghaoHuang45 who had Outstanding Papers 🏆

[Weiyan Shi @shi_weiyan]

It feels emotional to hear that #EMNLP is going back to China after 10 years🥹🥹🥹 thanks @emnlpmeeting ❤️❤️❤️

[Taiwei Shi @taiwei_shi]

We show that LLMs’ political ideology can be easily manipulated by just a few synthetic samples! 😈

Join @kaichen23 on Nov 12 (Tue) from 16:00 to 17:30 in poster session 4 for this exciting paper. #EMNLP2024