Vision-Language Models (VLMs) can describe the environment, but can they refer within it? Our findings reveal a critical gap: VLMs fall short of pragmatic optimality. We identify 3 key failures of pragmatic competence in referring expression generation with VLMs: (1) cannot

P.S., We are building GrowAIlikeAChild, an open-source community uniting researchers from computer science, cognitive science, psychology, linguistics, philosophy, and beyond. Instead of putting growing up and scaling up into opposite camps, let's build and evaluate human-like AI

🔨🔧⚙️ Have Vision Language Models solved mechanical reasoning? If you are at ICLR 2026, please come and check out our poster "Probing Mechanical Reasoning in Large Vision Language Models" Bidirectional Human-AI Alignment #ICLR2025 ! 📷Room Garnet 216-214 🗓️April 28 My work with GrowAIlikeAChild

ICLR 2026 is this week! Check out new research by Rada Mihalcea Martin Ziqiao Ma Haizhong Zheng Atul Prakash Zhijing Jin Honglak Lee and more! cse.engin.umich.edu/stories/eleven…

#ICLR please checkout our poster‼️We evaluated 209 models and all of them are stochastic parrots 🦜 🙀Models either believe "the bigger the ball, the quicker it falls" (illusions), or, "not matter how big, physics textbook, Pisa's tower, fall at the same time" (shortcuts).

#ICLR2025 #ICLR Small poster, big insights ⁉️ Vision Language Models Know Law of Conservation without Understanding More-or-Less 🙀🙀 Come to our poster at ICLR 2026 Bidirectional Human-AI Alignment 📷Room Garnet 216-214 📷 ❗️April 28 ‼️Work by GrowAIlikeAChild

#ICLR Bidirectional Human-AI Alignment‼️🤔Intention understanding and perspective-taking are core theory-of-mind abilities that humans typically develop starting around age 3. However, in VLMs, these two abilities dissociate.🧐 📅 April 28, Garnet 216-214 openreview.net/forum?id=rmHnN… 👏GrowAIlikeAChild

#ICLR Spurious Correlation & Shortcut Learning Workshop‼️introducing "Concept-Hacking" 🙀 We evaluated 209 models and found all of them are stochastic parrots🦜 🙀 Models either believe "the bigger the ball, the quicker it falls" (illusions), or, "not matter how big, fall at the same time" (shortcuts). 😲

#ICLR 💐 GrowAIlikeAChild We have 4 posters tomorrow‼️ Evaluating Multi-modal Language Models Through Concept Hacking 🦜Spurious Correlation & Shortcut Learning Workshop openreview.net/forum?id=B2QXX… Vision Language Models See What You Want but not What You See 😍 🙈Bidirectional Human-AI Alignment openreview.net/forum?id=rmHnN… Probing Mechanical

Finally finish making our website GrowAIlikeAChild growing-ai-like-a-child.github.io 🙌 with William Yijiang Li Dezhi Luo Martin Ziqiao Ma Zory Zhang Pinyuan Feng (Tony) Pooyan Rahmanzadehgervi and many others 🚶🚶‍♂️🚶‍♀️

👁️ 𝐂𝐚𝐧 𝐕𝐢𝐬𝐢𝐨𝐧 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 (𝐕𝐋𝐌𝐬) 𝐈𝐧𝐟𝐞𝐫 𝐇𝐮𝐦𝐚𝐧 𝐆𝐚𝐳𝐞 𝐃𝐢𝐫𝐞𝐜𝐭𝐢𝐨𝐧? Knowing where someone looks is key to a Theory of Mind. We test 111 VLMs and 65 humans to compare their inferences. Project page: grow-ai-like-a-child.github.io/gaze/ 🧵1/11

Gaze has been on my mind for a long time. In real life, we don’t use mouse cursors—we use gaze, head turns, and gestures to refer nonverbally. We (GrowAIlikeAChild) ask: can VLMs interpret gaze like humans do? Spoiler: they mostly chase head direction, not eye gaze. Having

#ICML #cognition #GrowAI We spent 2 years carefully curated every single experiment (i.e. object permanence, A-not-B task, visual cliff task) in this dataset (total: 1503 classic experiments spanning 12 core cognitive concepts). We spent another year to get 230 MLLMs evaluated

🚀 Dive in 👇 🌐 Project: williamium3000.github.io/core-knowledge/ 📄 Paper: arxiv.org/abs/2410.10855 📝 OpenReview: openreview.net/forum?id=EIK6x… 📊 Dataset: huggingface.co/datasets/willi… 💻 Code: github.com/williamium3000… 😎 Team: growing-ai-like-a-child.github.io 💥 Huge thanks to my amazing collaborators

🔥 Huge thanks to Yann LeCun and everyone for reposting our #ICML2025 work! 🚀 ✨12 core abilities, 📚1503 tasks, 🤖230 MLLMs, 🗨️11 prompts, 📊2503 data points. 🧠 We try to answer the question: 🔍 Do Multi-modal Large Language Models have grounded perception and reasoning?

Thanks to SenseTime to comprehensively investigate our #CoreCognition framework and eval on GPT-5 OpenAI Extremely interestingly, GPT-5s have achieved significantly improvements on concrete operational stage (Alexander Wei Noam Brown), namely object permanence, intuitive

Hokin Deng SenseTime OpenAI Alexander Wei Noam Brown YuanLiuuuuuu Yubo Wang Brian Bo Li Ziwei Liu Thanks, Hokin! CoreCognition has been a major source of inspiration for us. What we find particularly fascinating is how perspective-taking seems largely uncorrelated with other multimodal capabilities. Congrats to your team too, and we’re excited to see what’s next. 😄

Thanks Hokin Deng for the attention to our work! We’re inspired by many insights in #CoreCognition. Particularly, Fig.6 is a key motivation—showing perspective-taking as a unique ability with low correlation to others, even though CC isn’t specifically about spatial intelligence.

#GrowAI #cognition #philoLLM I am very excited that many frontier LM scientists start to look into insights from human cognition and philosophy of mind. I talked to many philosophers recently and conveyed to them that frontier AI practitioners are not just engineers who chase