Have arrived in Suzhou! I will present DISCO paper in EMNLP 2025 Thursday’s noon poster session. Feel free to reach out and discuss! If you’re interested in Meta’s current position for both FTE or internships, also let me know! #EMNLP2025

🧠 What if large models could read each other’s minds? Our new paper (#neurips2025 spotlight), “Thought Communication in Multiagent Collaboration”, explores how large model agents can share latent thoughts, not just messages. 📷arxiv.org/abs/2510.20733 (CMU × Meta AI ×

Unified multimodal models can generate text and images, but can they truly reason across modalities? 🎨 Introducing ROVER, the first benchmark that evaluates reciprocal cross-modal reasoning in unified models, the next frontier of omnimodal intelligence. 🌐 Project:

Scaling Agent Learning via Experience Synthesis 📝: arxiv.org/abs/2511.03773 Scaling training environments for RL by simulating them with reasoning LLMs! Environment models + Replay-buffer + New tasks = cheap RL for any environments! - Strong improvements over non-RL-ready

🚀Introducing Lumine, a generalist AI agent trained within Genshin Impact that can perceive, reason, and act in real time, completing hours-long missions and following diverse instructions within complex 3D open-world environments.🎮 Website: lumine-ai.org 1/6

🚀 Launching DeepSeek-V3.2 & DeepSeek-V3.2-Speciale — Reasoning-first models built for agents! 🔹 DeepSeek-V3.2: Official successor to V3.2-Exp. Now live on App, Web & API. 🔹 DeepSeek-V3.2-Speciale: Pushing the boundaries of reasoning capabilities. API-only for now. 📄 Tech

When I asked people what they wanted from a good tutorial, someone said, “Beautiful figures so I can take photos of the slides as souvenirs.” 📸 (Joking… kind of) If you’re at NeurIPS, come to our (w/ Michael Saxon ✈️ NeurIPS SD Xiang Yue) tutorial “The Science of Benchmarking: What’s

There are competing views on whether RL can genuinely improve base model's performance (e.g., pass@128). The answer is both yes and no, largely depending on the interplay between pre-training, mid-training, and RL. We trained a few hundreds of GPT-2 scale LMs on synthetic

One key takeaway we want to highlight💡: pure token-level expert selection is fundamentally limited.😮‍💨 FusionRoute addresses this by letting the router also generate, making token-level collaboration strictly more expressive, without joint training or model merging.

Token-Level LLM Collaboration via FusionRoute huggingface.co/papers/2601.05…