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?

🚀 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: embodiedbench.github.io Explore the evaluation code: github.com/EmbodiedBench/…

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:

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

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

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: chongyi-zheng.github.io/infom.

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 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: pd-perry.github.io/value-flows.