Find us @ East Exhibit # 2502 (⏰ 4:30PM - ) if you are at #NeurIPS2024 :D

Xudong Sun is on the faculty job market. He is truly brilliant, and has been doing exciting research on System Verification and Software Testing towards the vision of provably correct cloud infra and systems. His materials can be found at, marshtompsxd.github.io Interview him;

Proud Moment 📣: DeekSeek uses P to validate correctness of their distributed file system (3FS). DeepSeek open sourced 3FS designed to address the challenges of AI training and inference workloads. The coolest part is that the team also provided formal specifications of the

Checkout Yuxiang Wei ‘s new work! Online learning plus simple patch similarity reward makes models better SWEs!

Noam Brown “Designing hard evals is easy, designing meaningful ones is hard” -unknown

🚀 I'll be launching the Formal Methods Engineering Lab (manchester-fme.github.io) – and I am hiring! If you’re interested in working with me, feel free to reach out.

multi-turn reasoning convo can drop earlier CoTs to save context, but we need to prefill assistant[i-1] + user[i], instead of just user[i] for non-reasoning models. in agentic tasks, assistant outputs can be pretty long. so ig a quick optimization can be “pre-”prefilling

I’m gonna be recruiting students thru both Language Technologies Institute | @CarnegieMellon (NLP) and CMU Engineering & Public Policy (Engineering and Public Policy) for fall 2026! If you are interested in reasoning, memorization, AI for science & discovery and of course privacy, u can catch me at ACL! Prospective students fill this form:

🪜Introducing: StepWiser🦉 📝: arxiv.org/abs/2508.19229 - Reframes stepwise reward modeling as a reasoning task: outputs CoT + judgment. - Trained by RL using relative outcomes of rollouts. Results: (1) SOTA performance on ProcessBench! (2) Improves policy at train time. (3)

Xundong is great. Come work with him if you like building cool systems!

📢 The Software Engineering group at Cornell Bowers Computing and Information Science is growing fast -- we're now 8 PhD students strong! I’m recruiting PhD students for Fall 2026! If you are interested in the intersection of SE and AI, apply to Cornell CS and reach out! Ddl: Dec 15, 2025. RT!

If you value benchmark scores above everything else you’re going to have a bad time

I have one PhD intern opening to do research as a part of a model training effort at the FAIR CodeGen team (latest: Code World Model). If interested, email me directly and apply at metacareers.com/jobs/214557081…

today i had a talk in hkust gz, one friend asked me how come we can make the bet on scaling linear attention. my answer is more about the culture that i have been trying to make. admittedly it is too hard to change the mechanism which always rewards visible contribution and

What the outside world largely does not understand about OpenAI is how decentralized and bottom-up it is, and why this is incredibly good and important for the democratization of benefits. Aidan explains it well.

Today we at OpenAI are releasing GPT-5.1-Codex-Max, which can work autonomously for more than a day over millions of tokens. Pretraining hasn't hit a wall, and neither has test-time compute. Congrats to my teammates Kevin Stone & Michael Malek for helping to make it possible!

Meet GPT-5.1-Codex-Max, our latest frontier agentic coding model, available in Codex starting today. It’s faster, more capable and token-efficient, and able to work persistently on long tasks with built-in compaction abilities.

I’ll be at #NeurIPS2025 this week to present SWE-RL (main conf poster & DL4C oral) and chat about Code World Model (with features that later trended in SOTA LLMs like preserved reasoning and budget reminders). Let’s connect and discuss the future of software agents training!

Today, OpenAI is launching a new Alignment Research blog: a space for publishing more of our work on alignment and safety more frequently, and for a technical audience. alignment.openai.com

for now i’m more interested in the easiest problem the model can’t solve, rather than the hardest one it can. we underestimate how important reliability is!