Mukul Ranjan

Happy to announce our recent project "Time Blindness"! We found that all current AI models are completely blind to temporal patterns that humans see effortlessly. Even GPT-4o and Gemini score 0% on tasks where humans achieve 98% accuracy. timeblindness.github.io 🧵0/n

🎉🎉🎉 Excited to introduce our recent project FigMirror - a very interesting and useful tool with a simple workflow for making any paper-style figures. - See a beautiful figure in a paper - Screenshot it - Add your own data - Get a new Matplotlib figure with the same visual style FigMirror learns the quiet details that make paper figures look polished: - typography - spacing - line weight - color restraint - layout rhythm The key mechanism is Grounded Measurement. Computer-use AI can point to coordinates inside the reference figure. Code then inspects the pixels, colors, spacing, and layout around those points. This gives the system concrete visual evidence to iterate on. Our FigMirror draws a candidate figure, compares it with the reference, keeps what works, and improves what still feels off. Outputs: - editable Matplotlib code - camera-ready PDF It works as both a local Web UI and a Codex / Claude Code skill. Open source: github.com/VILA-Lab/FigMi… Try it before your next deadline!

Excited to share our new paper: “Useful Memories Become Faulty When Continuously Updated by LLMs. Can LLM agents keep improving by turning past experience into compact, reusable memories? We find this is much more fragile than it looks. Continuously consolidated memories can perform worse than no memory at all — sometimes even on problems the agent previously solved. Episodic memories that preserve raw episodes are much more reliable. There is still limited evidence that today’s models can learn reusable abstractions from experience over the long term, which I believe is a crucial capability for agents that continuously improve. Paper: arxiv.org/pdf/2605.12978. Congrats to @dylan_works_ and team!

Paper: arxiv.org/abs/2510.14973 Code: github.com/VILA-Lab/Elast… Project page: vila-lab.github.io/elastic-cache-… @szq0214

If you're going to be at the conference, I'd love to connect and chat about diffusion LLMs, efficient inference, hw/sw co-design or anything in between, feel free to reach out! And if the project sounds interesting, give a star on our repo.

Excited to share that I'll be attending #ICLR2026 at Rio! I'll be presenting our work on Elastic-Cache, a method for faster, more efficient KV caching in diffusion LLMs. We show consistent accuracy gains alongside significant speedups (up to 45× on GSM8K).

Without Einstein's general relativity from 1915, your GPS would drift about 10 km per day, and you'd have no idea why.

#ACL2026 This year we submitted four papers to ACL 2026 and all are accepted (1 Main + 3 Findings, a 100% acceptance rate). Congratulations to my students, the best team! @SondosBsha59089 @YaxinLuo999999 @mukul_ranjan_ @ShazanGhazi 🎉🎉🎉

Look closely. Between these two moments, our species has performed miracles. We have mapped the blueprint of life within our own DNA. We have built “brains” of silicon that can outthink their creators. We have pushed back the darkness of disease. Infant mortality has plummeted, and millions of children who would have been lost to the earth in 1972 are today alive, dreaming, and contributing to the global chorus. We have sent robotic emissaries to the edge of the interstellar dark and peered back at the beginning of time itself through mirrors of gold. Technologically, we are a different species. We are more connected, more informed, and more capable than any ancestor could have imagined in their wildest fever dreams. And yet, look again. From this distance, the borders remain invisible. You cannot see the “holy” ground over which we spill the blood of our children. You cannot see the walls we build to keep our neighbors out or the ideological trenches we dig to bury our common humanity. Despite our leap from vacuum tubes to artificial intelligence, we remain haunted by the same ancient tribalisms. We use 21st century technology to prosecute Bronze Age grudges. We have changed the climate of our world, but we have yet to change the climate of our hearts. We are still a toddler civilization, playing with matches in a library of irreplaceable wonders. The contrast is our great paradox. We have the power of gods, but we still possess the temperaments of the territorial primates from which we rose. We have learned to fly between worlds, but we are still struggling to learn how to walk together on this one.

That's us! 🌍 The Artemis II crew captured beautiful, high-resolution images of our home planet during their journey to the Moon. As @Astro_Christina put it: "You guys look great."

@gowthami_s @theworldlabs Congrats 🎉

Excited to share I’ve joined @theworldlabs! Generating pixels and frames was just the prologue. Now it's time to build frontier models that actually understand physics and power living, breathing simulations. Onwards to new worlds. 🌎🚀

Congratulations to David Basin, Affiliated Professor of Computer Science at MBZUAI, on receiving the 2026 Levchin Prize for Real-World Cryptography. The Levchin Prize recognizes major innovations with lasting impact on the practical use of cryptography, and Professor David's work is a testament to exactly that. He and his collaborators were honored for Tamarin, an open-source tool for the formal analysis of security protocols. Tamarin has been used to model and verify the security of some of the most critical systems in the world, including 5G, TLS, EMV, and iMessage PQ3. Established in 2016, the Levchin Prize celebrates contributions that strengthen the foundations of secure digital infrastructure globally — and we couldn't be prouder to have David as part of the MBZUAI community. #MBZUAI #LevchinPrize #Cryptography #CyberSecurity #Tamarin #AIResearch #ComputerScience

1080p video. 30 seconds generated in 5 seconds. FastVideo from MBZUAI's IFM isn’t just faster, it changes what video is. Pair it with K2 Think → real-time intelligence + real-time generation. Try it here: dreamverse.fastvideo.org

Blog post - Transformers as Constrained Optimization Rewriting pre-norm decoder-only transformers as solutions to regularized objectives. Changing regularization to hard constraint gives a canonical temperature, generalizing to KL-divergence, ideas of cross-layer interaction.

Introducing Olmo Hybrid, a 7B fully open model combining transformer and linear RNN layers. It decisively outperforms Olmo 3 7B across evals, w/ new theory & scaling experiments explaining why. 🧵

It’s extremely good that Anthropic has not backed down, and it’s siginficant that OpenAI has taken a similar stance. In the future, there will be much more challenging situations of this nature, and it will be critical for the relevant leaders to rise up to the occasion, for fierce competitors to put their differences aside. Good to see that happen today.

Excited to share LUCID — a new attention mechanism that improves retrieval and reasoning in long-context LLMs! [1/9]🧵 Here's how it works:

This is really cool!

In January, @jonhoo, @jjgort, and I returned to @MIT_CSAIL to teach Missing Semester, a class on topics missing from most CS programs—tools and techniques that everyone should know, like Bash, Git, CI/CD, and AI tools. Today, we’re releasing the course for free online!

@o_v_shake Thanks for sharing Abhishek.

People at India AI summit: if you want to work at frontier labs and dont want to spend money to prep for interviews, here’s a new resource: workatafrontierlab.com

New art project. Train and inference GPT in 243 lines of pure, dependency-free Python. This is the *full* algorithmic content of what is needed. Everything else is just for efficiency. I cannot simplify this any further. gist.github.com/karpathy/8627f…