Yu He (@yuhe441) header
Yu He (@yuhe441) profile picture

CS PhDing @stanford | BA+MEng @Cambridge_CL

Joined October 2014
Photo by Yu He (@yuhe441)Photo by Yu He (@yuhe441)Photo by Yu He (@yuhe441)Photo by Yu He (@yuhe441)Photo by Yu He (@yuhe441)Photo by Yu He (@yuhe441)Photo by Yu He (@yuhe441)Photo by Yu He (@yuhe441)Photo by Yu He (@yuhe441)Photo by Yu He (@yuhe441)
Happy to share our @NeurIPSConf '26 paper with @yuhe441, Yingxi Li, and @yifeiwang77! We study whether LLMs can combine skills learned through RL and find that training on individual skills transfers poorly to composed tasks, while transfer in the other direction is stronger.
🎉 Excited to share that our paper "Compositional Reasoning in Language Models under Reinforcement Learning Post-Training" has been accepted to NeurIPS 2026! 🎉 In this paper, we ask: Can LLMs learn individual reasoning skills and compose them to solve complex problems? (1/8)
1
5
37
4,645
🎉 Excited to share that our paper "Compositional Reasoning in Language Models under Reinforcement Learning Post-Training" has been accepted to NeurIPS 2026! 🎉 In this paper, we ask: Can LLMs learn individual reasoning skills and compose them to solve complex problems? (1/8)
28
57
4
690
46,498
(7/8) 💡 The practical implication: RL training data should include reasoning tasks that expose models to compositional structures, enabling them to learn both individual skills and how those skills interact. Training on isolated skills alone is insufficient.
1
1
17
1,415
My team @GoogleDeepMind is hiring! We are seeking a Research Scientist to rigorously study the internals and generalisation of frontier AI systems, w/ a particular view to AI security. This is an in-person role at our Platform 37 office in London 🎡 Link to apply in thread 👇
8
46
1
428
41,609
Check out Henry's talk on our upcoming @NeurIPSConf 2026 paper, "FLARE: Verifying MILP Reformulations with LLM-Based Theorem Proving." Paper: arxiv.org/abs/2608.25220 Code + Benchmark: flare.henryrobbins.com Led by Henry, with @LawlessOpt and @madeleineudell, using @leanprover.
Henry Robbins presents "Verified Mixed-Integer Programming in Lean" at the SF Lean Meetup on September 14 0:00 Introduction 1:15 Overview and Motivation 6:01 Mixed-Integer Programming 18:36 MILP Formulations and the TSP 26:54 MILP Reformulations 36:12 Verifying Reformulations Automatically 38:29 Quasi-Karp Equivalence and EquivaMap 1:01:33 Pitfalls of Instance-Level Verification 1:17:00 FLARE: Constructive Reformulations in Lean 1:21:47 Results on FormulationBench 1:24:16 Future Directions 1:25:10 Q&A youtu.be/c--uaHWwiC0
1
1
16
947
Scientific 🔬 progress often starts with an unexpected observation that sparks a new hypothesis. But AI agents often start with a hypothesis, then look for observations to support it. We believe agents should actively explore both observations and hypotheses. When observations are only a byproduct of hypothesis testing, we risk missing discoveries we weren’t looking for. ⚡ Meet SHELDON: a scientific reasoning harness that models separate hypothesis and observation spaces and generalizes across medical diagnosis, system-log investigations, ARC-AGI 2, and even autoresearch. Bazinga!
4
23
7
101
7,494
Honored to be named an HAI Data Science Scholar!
Meet our 15 new HAI Data Science Scholars. This year’s cohort spans computer science, law, neuroscience, sociology, education, and materials science, with research ranging from AI governance to drug discovery to how the brain builds models of the world: hai.stanford.edu/news/stanfo…
11
954
Sign up for SASP if you need help with your Stanford CS PhD application this year!
Applying to Stanford's CS PhD program this year? We are organizing the 2026 Student-Applicant Support Program (SASP) and would love to help with your application materials. SASP is a volunteer initiative run by Stanford CS graduate students. We provide feedback on statements of purpose and resumes for applicants from diverse backgrounds, broadly defined, before they submit their applications. We know how much work goes into PhD applications, and we hope an extra set of eyes on your materials can make the process a little easier. Please submit your materials by October 25 (Pacific Time). We plan to return feedback by November 20. More details and how to participate: cs.stanford.edu/sites/g/file… Please share this with anyone who might find it helpful! Questions are welcome at [email protected]. -- Bailey Trang (@baileytrangn), Bo Ai (@BoAi0110), Hermann Kumbong (@KumbongHermann), Kwon Ko (@hyungkwonko), Michael Paper, Swapnil Gandhi (@sw2pnil), and Yu He (@yuhe441)
12
988
Introducing Mosaic 🧩, a constrained decoding framework for diffusion language models. - significantly improves DLMs' performance on general function calling and JevBench. - compatible with vLLM, SGLang and HF, supporting DiffusionGemma, LLaDA2, and many others. 1/n 🧵
9
35
2
190
32,462
How can theoretical computer science help us think about the interaction between algorithms and society? I wrote some personal reflections for @KnightHennessy about my path into this area and the research community working on these questions knight-hennessy.stanford.edu…
2
10
109
7,291
Check out our poster on neural algorithmic reasoning in the AI4Math workshop at ICML.
1
2
16
968
🍡Introducing a new constrained decoding approach for agentic LLMs: 1⃣guarantees successful tool calls with max_tokens=k (no truncation); 2⃣improves downstream performance with test time compute. Presented now at #ICML2026 poster session 1 #1912 by @jieyuzhao11.
4
7
1
40
8,075
We are at poster session 8 today 5PM at Hall A #2108! I won’t be at ICML this year sadly, but Yingxi is there to chat!
[0/n] Can LLMs 𝘢𝘤𝘵𝘶𝘢𝘭𝘭𝘺 reason about structure—order, hierarchy, connectivity, and how parts fit together?🧩 We introduce 𝗗𝗦𝗥-𝗕𝗲𝗻𝗰𝗵: a data-structure benchmark designed to test this 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝘁𝗼𝗼𝗹𝘀. Even SOTA LLMs still struggle in the hardest settings.
1
10
1,048