Siyin Wang @ICLR 2026

@xuningy how about the performance of recent WAMs like Dreamzero

When every generalist robot model scores 95%+ on a benchmark, the numbers become meaningless. What if we built a photorealistic benchmark that never saturates and can generate new scenes and tasks with AI Workflows in minutes? We introduce RoboLab! 🧵(1/6)

Made it to Rio after nearly 30 hours in the air ✈️ Excited to be at #ICLR2026! We'll be sharing our work on RoboOmni on Apr 25 (10:30 AM – 1:00 PM). If you're interested in VLAs, Omni-LLMs, World Models, or Multimodal Agentic AI, feel free to reach out! #multimodality #VLA

What if robots could understand what you want without being told? RoboOmni makes that possible — an omni-modal LLM that fuses speech, sound, and vision to infer human intent, confirm actions, and execute tasks. Trained on the new OmniAction dataset (140k episodes), it outperforms text- and ASR-based baselines in success rate, speed, and proactive assistance, paving the way for more intuitive human-robot collaboration. RoboOmni: Proactive Robot Manipulation in Omni-modal Context Fudan, SII, NUS Paper: arxiv.org/abs/2510.23763 Code: github.com/OpenMOSS/RoboO… Project: OpenMOSS.github.io/RoboOmni Our report: mp.weixin.qq.com/s/PXBqdEW7_Ta_… 📬 #PapersAccepted by Jiqizhixin

I will be at EMNLP from Nov 5th to Nov 8th. If you are interested in multimodal spatial reasoning, Embodied AI (like VLA), Omni-LLMs, please feel free to chat with me!👋 📍I will also present a poster (ConvSearch-R1) with Changtai at Nov 6th 16:30-18:00. #EMNLP2025 #suzhou

Unveiling the hidden fragilities of AI robotics New research introduces LIBERO-Plus, a comprehensive benchmark that systematically reveals vulnerabilities in Vision-Language-Action models under 7 real-world perturbation dimensions.

🚀Tired of Libero? Try our Libero-Plus! 🤔Libero’s at 99%, but we’ve found VLA drops points with even minor disturbances. 🤩Switch to Libero+ in just a few steps and unlock your VLA’s true generalization ability. #Embodied #VLA #Robotics

@feiliu_nlp @HuiWei15 congrats!

🏆 Thrilled that our paper #PlanGenLLMs (arxiv.org/abs/2502.11221) won the SAC Award at #ACL2025!! Couldn't have done it without the amazing team: @HuiWei15, Zihao Zhang, Shenghua He, Tian Xia, and Shijia Pan. So thankful and beyond proud! 💖 #ACL2025NLP #NLProc 🧠 Planning is a core aspect of both human and artificial intelligence. LLMs/agents have been used in various planning tasks, from navigating websites and planning trips to querying databases, but most benchmarks are narrow and task-specific. That makes it difficult to compare systems across domains, or figure out which one's best for a new planning problem. That's where our paper comes in: We offer a comprehensive overview of LLM-based planning agents, highlighting gaps, challenges, and what's next. Check it out 👉 arxiv.org/abs/2502.11221

Excited to be in Vienna! Come by if you’re around!👋#ACL2025 📌Poster: VisuoThink (Thinking with images, Multimodal reasoning) 📌Poster: D2PO (World Modeling, Embodied) 🗓️Tue, July 29 16:00–17:30 |📍Hall 4/5 (Session 3) Presenting with @ngc7293q @JinlanFu !👏

@feiliu_nlp Congrats! Looking forward to seeing you!

Happy to share our paper got selected as an Oral Presentation at #ACL2025! Out of 8,000+ submissions and 3,000+ accepted papers, only 245 were chosen for oral (<3%)! Our amazing first author Hui Wei can't travel, so I'll be presenting in Vienna. Hope to see you there! 🇦🇹 📄 Paper: arxiv.org/abs/2502.11221 💻 Resource: github.com/wll199566/Awes… #ACL2025 #NLProc #PlanGenLLMs #LLM #AgenticAI

Awesome Open-sourced work! 🚀

Exciting work on Embodied Agents! 🦾 By leveraging interactive reinforcement learning, we shatter the ceiling on ALFWorld (31.05 -> 97.78) and ScienceWorld (22.05 -> 79.92). 🔥 Huge thanks to amazing coauthors @ngc7293q li ji, junhao, @JingjingGong_ @xpqiu

Thanks for sharing our latest work! 🚀 Grateful to work with amazing collaborators: junhao @ngc7293q @JingjingGong_ qipeng @xpqiu !

1️⃣VisuoThink 🔗 arxiv.org/pdf/2504.09130 🛠️ github.com/ekonwang/Visuo… 2️⃣D²PO 🔗 arxiv.org/pdf/2503.10480 🛠️ github.com/sinwang20/D2PO Huge thanks to all my amazing collaborators for their brilliant insights! Excited to see these ideas contribute to the field of multimodal reasoning!

Thrilled to share our TWO papers accepted to #ACL2025 Main Conference! 🥳🎉 🎨VisuoThink: Empowering LVLM Reasoning with Multimodal Tree Search 🌏World Modeling Makes a Better Planner: Dual Preference Optimization for Embodied Task Planning #AI #MultimodalLearning #worldmodel

World Models are clearly booming, attracting researchers across CV, NLP, and Robotics! 🔥 We also presented our work at this workshop—thanks to everyone who stopped by! 🤩 #ICLR2025 #worldmodel

@furongh happy to chat with you today! I am also interested in VLA and world model. 🤩

I will present TraceVLA today from 3 p.m. to 5:30 p.m. in Hall 3 + Hall 2B Poster #35. Find me there, and let's chat. #VLAs #Robotics #FoundationModel4Robotics

📣 This work demonstrates that embedding world modeling into the planning leads to much better embodied intelligence. Check out the full paper for details, and great thanks to my talented collaborators @ngc7293q @cheng_qinyuan @Joey_zh_ @PanpanCai @JinlanFu @xpqiu 8/8

🔎 We also compared action-conditioned (predict state given state & action) vs. goal-directed world models (imagines future state from history & goal). While action-conditioned excel in familiar settings, goal-directed models generalize better to novel environments! 7/8

✨ Excited to share our latest research “World Modeling Makes a Better Planner: Dual Preference Optimization for Embodied Task Planning” 🤔 Current LVLMs struggle with grounding in embodied environments, how can we make AI agents understand the physical world like humans? 1/8