Zishuo Zheng

2nd-year PhD @OhioStateCSE, working with Prof. @shocheen on LLMs

[Zishuo Zheng @ZishuoZheng]

That why you need our work to ensure the robot will obey the three laws of Asimov 🤣
https://t.co/3AbA28rhqD

[Jon Hernandez @JonhernandezIA]

So this guy plugged an Ai to a robot and ask the Ai to shoot him, but the Ai refused... Until it actually shoot him 😂

We are so fucked

[Zishuo Zheng @ZishuoZheng]

@xiangyuli 而且丹阳的学术水平怎么样我不知道，她实验室出来的学生确实不怎么样。她上位的手段也很了得。

[Zishuo Zheng @ZishuoZheng]

@xiangyuli 我接触过poo后对他的印象不好，一个纯粹的政治家而不是科学家。不关心学生，把学生当耗材压榨，去培养他所谓的年轻科学家。
神经所丑闻频出，不是诬陷，都是实打实的丑闻。神经所招来的人中很少有真才实学，很多都是骗子。

[Zishuo Zheng @ZishuoZheng]

OpenAI further validated our method
on large-scale frontier model GPT-5-mini and proposal IH-Challenge in their work "IH-Challenge: A Training Dataset to Improve Instruction Hierarchy on Frontier LLMs"
https://t.co/MsCZa9k42X

[Zishuo Zheng @ZishuoZheng]

Our paper "Reasoning Up the Instruction Ladder for Controllable Language Models" has been accepted as ACL Findings 🎉🎉🎉

I’ll be at #ACL2026 (July 2-8) and present this work at Poster Session C on Sunday, July 5th in GRAND HALL & CORONADO
Come and chat if your are interested!

[Zishuo Zheng @ZishuoZheng]

✨We are excited to introduce Reasoning for Instruction Hierarchy!

We train LLMs to explicitly reason over instruction hierarchies (IH) with RLVR, producing flexible and robust behavior!

Thanks my coauthers @vidhisha_b , @chan_young_park , @faeze_brh , and my advisor @shocheen

[Zishuo Zheng @ZishuoZheng]

We have extended reasoning for instruction hierarchy into larger-scale MoE model Qwen3.6-35B-A3B.
Our method achieves further gains even when model shows an strong baseline on instruction hierarchy tasks (90%->92% on IHEval-conflict setup).
Check our paper for details!

[Hanane Nour Moussa @HananeNMoussa]

Gym environments have played a key role in advancing LMs and agents for general coding tasks. But how do we build them for scientific coding?

Introducing D3-Gym, the first automatically constructed dataset of verifiable environments for data-driven scientific discovery. 🧵

[Zishuo Zheng @ZishuoZheng]

#NeurIPS2025
On December 7 (Sunday), I’ll present our recent work on Reasoning for Instruction Hierarchy. We are mainly working on post-training with RLVR on reasoning models to build controllable and flexible LLMs. Check my last poster for more details!

The Workshop is at Upper Level Room 33ABC, and the poster session is at 10:15-11:30 and 15:00-16:15. Welcome to meet and chat with me if you are interested!

I’m also currently seeking Summer 2026 internship opportunities. If you know relevant openings, please contact me via LinkedIn or speak with me in person at NeurIPS!

[Dylan Foster 🐢 @canondetortugas]

Don’t miss our workshop on Foundations of Reasoning in Language Models at NeurIPS tomorrow!

Upper Ballroom 33ABC, starting 8:50 AM

Featuring talks by @Azaliamirh, @aviral_kumar2, @natolambert, and @mhahn29 on self-improvement, exploration, Olmo 3, chain-of thought, and more!

[Zishuo Zheng @ZishuoZheng]

I’ll be at #NeurIPS2025 (Dec 2-8) and present this work at the FoRLM Workshop on Sunday, Dec 7, 15:00–16:15 in Room 33ABC.
I'm also actively seeking a Summer 2026 internship in RL, Post-Training, Reasoning, or Multimodal Models. Please let me know if any opportunities!

[Zishuo Zheng @ZishuoZheng]

✨We are excited to introduce Reasoning for Instruction Hierarchy!

We train LLMs to explicitly reason over instruction hierarchies (IH) with RLVR, producing flexible and robust behavior!

Thanks my coauthers @vidhisha_b , @chan_young_park , @faeze_brh , and my advisor @shocheen

[Zishuo Zheng @ZishuoZheng]

🔗𝗠𝗼𝗿𝗲 𝗗𝗲𝘁𝗮𝗶𝗹𝘀?
Check our paper and GitHub repo!
📑https://t.co/eznCsKiZII
🔧https://t.co/rATc9C32EH

[Zishuo Zheng @ZishuoZheng]

✨We are excited to introduce Reasoning for Instruction Hierarchy!

We train LLMs to explicitly reason over instruction hierarchies (IH) with RLVR, producing flexible and robust behavior!

Thanks my coauthers @vidhisha_b , @chan_young_park , @faeze_brh , and my advisor @shocheen

[Zishuo Zheng @ZishuoZheng]

By adding simple safety rules in the system prompt after RLVR training with VerIH, our method provides up to 20% gains on jailbreak and prompt injection robustness!