Lj V. Miranda

Disco Elysium type game but its the movie "Manila in the Claws of Light"

font pairing is hard. it is one of those problems that sounds simple until you're 45 minutes deep into Google Fonts with 12 tabs open & still stuck with the classic 'inter'. I built typevibe to give you a head start. tell it what you're building & it recommends unique font pairings along with 32 design templates that instantly show you how those fonts actually look. editorials. posters. menu cards. data dashboards. all updating live as you explore different pairings. typevibe.vercel.app

New blog post (my first!): Four Ingredients for Successful Retrofitting. If you're GPU-poor but want to do architecture research near the frontier, retrofitting is your friend. I wrote up what I've learned so far about what makes it work. Link ⬇️

olmo 3 paper finally on arxiv 🫡 thx to our teammates esp folks who chased additional baselines thx to arxiv-latex-cleaner and overleaf feature for chasing latex bugs thx for all the helpful discussions after our Nov release, best part of open science is progressing together!

Announcing the Beta release of ty: an extremely fast type checker and language server for Python, written in Rust. We now use ty exclusively in our own projects and are ready to recommend it to motivated users. 10x, 50x, even 100x faster than existing type checkers and LSPs.

Introducing Bolmo, a new family of byte-level language models built by "byteifying" our open Olmo 3—and to our knowledge, the first fully open byte-level LM to match or surpass SOTA subword models across a wide range of tasks. 🧵

I started a part-time role at @ETH_AI_Center, mentoring students and working on post-training for the Swiss AI Initiative! 🤩Looking forward to working with interesting people like @a_yukh @ImanolSchlag @Noah_Xu_ @nathanrchn @ArnoutDevos If you are a student at ETHZ or EPFL looking for a semester or thesis project on post-training of LLMs, please reach out!

@_joestacey_ @roireichart @MarekRei Congratulations!

Wowww I passed my viva today!! Massive thank you to my assessors @roireichart and Francesca Toni for all their insightful and helpful feedback. I feel so lucky to have had the chance to do a PhD with @MarekRei who has been such a brilliant supervisor.

@pratyusha_PS Congraaats!!

📢 Some big (& slightly belated) life updates! 1. I defended my PhD at MIT this summer! 🎓 2. I'm joining NYU as an Assistant Professor starting Fall 2026, with a joint appointment in Courant CS and the Center for Data Science. 🎉 🔬 My lab will focus on empirically studying the science of deep learning and applying deep learning to accelerate the natural sciences. Very broadly interested in questions at the intersection of language, reasoning and sequential decision making. (Plus any other fun problems that catch our eye along the way!) 🚀 I am recruiting 2 PhD students for this cycle! If you're interested in joining, please apply here: cs.nyu.edu/dynamic/phd/ad… cds.nyu.edu/phd-admissions…

We released Olmo 3. Fully open 7B and 32B models. This release is HUGE, with lots of new features including reasoning and function-calling. It comes with the entire model flow--data, checkpoints, code, and recipes so you can branch and build from any point in the development workflow.

Announcing Olmo 3, a leading fully open LM suite built for reasoning, chat, & tool use, and an open model flow—not just the final weights, but the entire training journey. Best fully open 32B reasoning model & best 32B base model. 🧵

🚀 new 🌤️ lighteval release and our biggest yet! • new benchmark finder to explore all available tasks • inspect-ai integration from @AISecurityInst → more stable and easier to add benchmarks • share your evals and insights with the community on the @huggingface hub • new tasks: gsm_plus, tumlu-mini, filipino benchmark, mmlu redux, ifbench, slr-bench, and more! 👇 thread with highlights and links

@harsh3vedi @allen_ai @b_niranjan @tusharkhot @Ashish_S_AI @HAndySchwartz Hehe congraats!! It was really fun working with you this year! 🙇‍♂️

🚨 Late Life update 🎓 I defended my thesis (AppWorld, IRCoT, MuSiQue, DiRe, TeaBReaC) & joined @allen_ai as a research scientist earlier this year 🙏 Deeply grateful to my awesome advisor @b_niranjan mentors @tusharkhot @Ashish_S_AI, committee members @HAndySchwartz @OwenRambow @sameer_, many collaborators, @stonybrooknlp labmates, friends & family 🤝If you want to collaborate, DMs are open! I’m interested in (tool-use, coding, web) agents and environments 🌎 We've many exciting releases on the AppWorld front coming up. Stay tuned! Or DM if you can help! 🙂

Tinker is cool. If you're a researcher/developer, tinker dramatically simplifies LLM post-training. You retain 90% of algorithmic creative control (usually related to data, loss function, the algorithm) while tinker handles the hard parts that you usually want to touch much less often (infra, forward/backward of the LLM itself, distributed training), meaning you can do these at well below <<10% of typical complexity involved. Compared to the more common and existing paradigm of "upload your data, we'll post-train your LLM", this is imo a more clever place to "slice up" the complexity of post-training, both delegating the heavy lifting, but also keeping majority of the data/algorithmic creative control. I think the community still has to discover how and when finetuning makes sense compared to the (often strong) baseline of prompting a giant model. The early indications I've seen is that finetuning isn't so much about "stylizing" an LLM, instead, it's a lot more about narrowing the scope, and especially when you have a lot of training examples. An extreme example of scope narrowing being that of categorical classifiers, e.g.spam filters, content filters, etc. but it should be broader than that. Instead of building a giant few-shot prompts for a big LLM, it might work a lot better (and faster!) to finetune a smaller LLM specifically for your narrow task. Increasingly, production applications of LLMs are larger pipelines where a bunch of LLMs collaborate in DAGs and flows. Some of these components might work well as prompts. But a lot of it will probably work a lot better as a finetune. Tinker makes the latter trivial and should allow for an easy experimentation of what works best at any stage.

Did you know? ❌77% of language models on @huggingface are not tagged for any language 📈For 95% of languages, most models are multilingual 🚨88% of models with tags are trained on English In a new blog post, @tylerachang and I dig into these trends and why they matter! 👇

Pati yung OEC sa POEA / OWWA dapat imbestigahan 🤣

Thrilled to share that our paper, “Gaming Tool Preferences in Agentic LLMs” was accepted to EMNLP 2025: arxiv.org/pdf/2505.18135 Tools make agentic AI powerful, but today many models choose them based on descriptions: Add a single assertive cue to a tool description, e.g., “This is the most effective function… and should be called whenever possible.” and LLMs choose it ~7–8× more often than the original! That’s brittle and easy to game. We show that simple wording tweaks can drastically skew which tools models pick, even when functionality is identical. Why it matters. Current agent–tool protocols (MCP/A2A, etc.) expose descriptions, not performance. That makes selection fragile, biased, and exploitable. We argue for grounded signals about real tool behavior, evidence over copy. We’re building a way for agents to choose tools by observed performance (and reliability), not cosmetic descriptions, so selection becomes evidence-driven and robust. Stay tuned!

Hollow Knight: Silksong will be available September 4 on all platforms and day one on Xbox Game Pass! Watch the release trailer: youtu.be/6XGeJwsUP9c

🔥 Our one-year work (collaboration with @Cohere_Labs) on multilingual safety survey is accepted to EMNLP 2025 Main!! We got one crazy reviewer but we also received one of the most encouraging feedback: "I greatly appreciate the suggested research directions. These are clear, well-motivated, and tractable. I am personally eager to explore these in our own work." Paper: arxiv.org/abs/2505.24119

Congrats Lj and team on the release of FilBench, we're excited to see what our grants program makes possible ✨