💥Excited to share our paper “AlphaOne: Reasoning Models Thinking Slow and Fast at Test Time” at #EMNLP2025 🚀 this Friday, Nov. 7, during Gather Session 4. Come say hi virtually!👋 📄Paper: https://t.co/CksN8hEuoF 🪩Website & Code: https://t.co/AwMLAQFvtz #AI #LLMs #Reasoning

1/ How can we model the future rewards (returns) for RL agents? While prior methods round the returns into discrete bins or predict a finite number of quantiles, we use flexible models to predict the fine-grained structure of the full return distribution: https://t.co/YyGEEQnnEH.

Also huge thanks to all my incredible collaborators and friends @RunpeiDong @HanWang178023 @krystal_ning @HaoranGeng2 @peihao_goblue @Xialin_He @YutongBAI1002 @JitendraMalikCV @_saurabhg @huan_zhang12 @drdhxi. Without them, nothing could be done.

🎯Key Takeaways 1⃣Slow thinking first, then fast thinking, leads to better LRM reasoning. 2⃣Slow thinking can bring efficient test-time scaling. 3⃣Dense reasoning modulation brings improved reasoning.

This led us to systematically study test-time reasoning modulation and develop AlphaOne, a universal framework that scale LRM reasoning at test time. 🔬

Through our research, we found that LRMs struggle to find the optimal schedule between these two thinking modes the way we do effortlessly every day!!!

Back in April, Runpei and I were fascinated by how AI reasoning differs from human thinking. Humans naturally switch between fast intuition⚡(System 1) and slow deliberation🐢(System 2). But how about large reasoning models?

For years I have dreamt of a tool that could neutralize pathogens the moment they emerge. Today we unveil ApexOracle—an AI that, from a pathogen’s genome and phenotypic knowledge alone, predicts which antibiotics will work and invents new molecules for threats it has never seen.

1/ How should RL agents prepare to solve new tasks? While prior methods often learn a model that predicts the immediate next observation, we build a model that predicts many steps into the future, conditioning on different user intentions: https://t.co/QYPy1Tlv4x.

EmbodiedBench got an ICML 2025 Oral! Time to challenge your MLLMs on embodied tasks!🤖

Huge thanks for sharing our work @_akhaliq! AlphaOne deep dive & code release coming soon 🚀

Reasoning Models Thinking Slow and Fast at Test Time Another super cool work on improving reasoning efficiency in LLMs. They show that slow-then-fast reasoning outperforms other strategies. Here are my notes:

🚀 New model results on EmbodiedBench! 🚀 🔹 Qwen2.5 VL surpasses Qwen2 VL as embodied agents! 🔹 InternVL2_5 MPO leads as the best-performing open-source model! Check out the latest results: https://t.co/91OLQaa9ml Explore the evaluation code: https://t.co/Q0FP6RI2dC

When we successfully built a framework that enables MLLM-based agents to plan for low-level manipulation tasks (a key component of EmbodiedBench), I was super excited! Could this be a step toward MLLM-based agents becoming so versatile that we no longer need dedicated VLA models?

🔍 Can Vision-Language Models (VLMs) truly reason about math, or are they just reflecting patterns they’ve seen before? 🤔 We tested GPT-4o and Claude-3.5 with our new benchmark, DynaMATH, and the results were eye-opening! 🧩 GPT-4o struggles to recognize when a shifted

Here are the links to our paper and project website. Paper: https://t.co/wTno35WCow Website:

Thrilled to share our work HERD in #ICLR2024. Thanks much to my co-author Heng @drdhxi for a year of collaboration in exploring Robot Design. Hope to see further research and attention on algorithmic aspects in this field!

Our work in #ICML2023: Symmetry-Aware Robot Design with Structured Subgroups. https://t.co/JtUsDvALP3