💡Today, we speak with HLF alumna Letiția Pârcălăbescu @aicoffeebreak.bsky.social who will be joining us at the 13th HLF as a panelist! 🖥️🎓 During her PhD in computer science, Letitia developed tools to help multimodal LLMs complete tasks more "honestly". 👉 scilogs.spektrum.de/hlf/?p=14524

Keeping AI Honest - Heidelberg Laureate Forum - SciLogs - Wissenschaftsblogs

We speak to computer scientist Letiția Pârcălăbescu on her research to help train more transparent and "honest" AI models.

scilogs.spektrum.de

🧠 Do Vision & Language Decoders Use Images and Text Equally? In our latest episode, we speak with Letitia Parcalabescu about her ICLR 2025 paper examining how vision–language *decoder* models use images and text — and how self-consistent their explanations really are. (1/8🧵)

If you love @aicoffeebreak.bsky.social, this one's for you — Letitia Parcalabescu is our next guest on the #WiAIR_podcast! Stay tuned for our conversation: 🎬 YouTube: www.youtube.com/@WomeninAIRe...

LLMs can memorize even a phone number seen once in training.🔒 Google’s VaultGemma fixes that, being the first open-weight LLM trained from scratch with differential privacy, so rare secrets leave no trace. ☕ new video explaining Differential Privacy through VaultGemma 👇 🎥 youtu.be/UwX5zzjwb_g

What's up with Google's new VaultGemma model? – Differential Privacy explained

YouTube video by AI Coffee Break with Letitia

youtu.be

We explain diffusion models and flow-matching models side by side. Flow-Matching models are the new generation of AI image generators that are quickly replacing diffusion models. They take everything diffusion did well, but make it faster, smoother, and deterministic. 🎥 youtu.be/firXjwZ_6KI

Diffusion Models and Flow-Matching explained side by side

YouTube video by AI Coffee Break with Letitia

youtu.be

Ever wondered how Energy-Based Models (EBMs) work and how they differ from normal neural networks? ☕️ We go over EBMs and then dive into the Energy-Based Transformers paper to make LLMs that refine guesses, self-verify, and could adapt compute to problem difficulty.

The world’s largest NLP conference with almost 2,000 papers presented, ACL 2025 just took place in Vienna! 🎓✨ Here is a quick snapshot of the event via a short interview with one of the authors whose work caught my attention. 🎥 Watch: youtu.be/GBISWggsQOA

My friend Vivi Nastase is working on a short science communication film called "Puppets of a Digital Brain". It aims to explain the tech behind AI chatbots (the good, the bad, the environmental) in an accessible, visual way. 💡 GoFundMe: gofund.me/453ed662

How do LLMs pick the next word? They don’t choose words directly: they only output word probabilities. 📊 Greedy decoding, top-k, top-p, min-p are methods that turn these probabilities into actual text.

Excited to be at ACL 2025 in Vienna this week 🇦🇹 #ACL2025 I’m always up for a chat about reasoning models, NLE faithfulness, synthetic data generation, or the joys and challenges of explaining AI on YouTube. If you're around, let’s connect!

🤖 Can we trust AI in science? I'm excited to be speaking at the final event of the Young Marsilius Fellows 2025, themed "Dancing with Right & Wrong?" – a title that feels increasingly relevant these days. I'll be joining a panel on "(How) can we trust AI in science?" to discuss questions like:

We train AI on human-selected or -generated data (yes, even taking a photo is concept selection – we capture what we find interesting; text even more so, expressing our conceptualisation of the world). Then we’re surprised when the AI's concepts and representations are similar to ours. 🤷‍♀️

I'm very excited to finally share the main work of my PhD! We explored the evolutionary dynamics of gene regulation and expression during gonad development in primates. We cover among others: X chromosome dynamics (incl. in a developing XXY testis), gene regulatory networks and cell type evolution.

Kaessmann Lab@kaessmannlab.bsky.social · last yr.

We are delighted to share our new preprint “The evolution of gene regulatory programs controlling gonadal development in primates” www.biorxiv.org/content/10.1...

💡 AlphaEvolve is a new AI system that doesn’t just write code, it evolves it. It uses LLMs and evolutionary search to make scientific discoveries. We explain how AlphaEvolve works and the evolutionary strategies behind it (like MAP-Elites and island-based population methods). 📺 youtu.be/Z4uF6cVly8o

AlphaEvolve: Using LLMs to solve Scientific and Engineering Challenges | AlphaEvolve explained

YouTube video by AI Coffee Break with Letitia

youtu.be

Excited to share that I’ll be joining the Summer School “AI and Human Values” this September at the Marsilius-Kolleg of Heidelberg University as a speaker. I'll be giving an introduction to how large language models actually work—before the summer school dives deeper into broader implications.

Long videos are a nightmare for language models—too many tokens, slow inference. ☠️ We explain STORM ⛈️, a new architecture that improves long video LLMs using Mamba layers and token compression. Reaches better accuracy than GPT-4o on benchmarks and up to 8× more efficiency. 📺 youtu.be/uMk3VN4S8TQ

Token-Efficient Long Video Understanding for Multimodal LLMs | Paper explained

YouTube video by AI Coffee Break with Letitia

youtu.be

Follow @aicoffeebreak.bsky.social!! Letitia is very effective in communicating research papers in just a few mins! Perfect for your coffee break. 😉

We all know quantization works at inference time, but researchers successfully trained a 13B LLaMA 2 model using FP4 precision (only 16 values per weight!). 🤯 We break down how it works. If quantization and mixed-precision training sounds mysterious, this’ll clear it up. 📺 youtu.be/Ue3AK4mCYYg

4-Bit Training for Billion-Parameter LLMs? Yes, Really.

YouTube video by AI Coffee Break with Letitia

youtu.be

Just say “Wait…” – and your LLM gets smarter?! We explain how just 1,000 training examples + a tiny trick at inference time = o1-preview level reasoning. No RL, no massive data needed. 🎥 Watch now → youtu.be/XuH2QTAC5yI

s1: Simple test-time scaling: Just “wait…” + 1,000 training examples? | PAPER EXPLAINED

YouTube video by AI Coffee Break with Letitia

youtu.be

🎙️ Yesterday, I gave a keynote on large language models outfitted with visual understanding, and the faithfulness of their chain-of-thought reasoning at the National Conference on Governing the Digital Society and Human-Centered AI.

Ece Takmaz@ecekt.bsky.social · 2y ago

It was very nice seeing @aicoffeebreak.bsky.social again after many years😊 She gave a wonderful keynote at the Railway Museum in Utrecht, as part of the National Conference on AI Transformations!💫

The National Conference on AI Transformations: Language, Technology, and Society organised by Utrecht University @utrechtuniversity.bsky.social was a success, and indeed Letiția‘s @aicoffeebreak.bsky.social talk was very inspiring.

Ece Takmaz@ecekt.bsky.social · 2y ago

It was very nice seeing @aicoffeebreak.bsky.social again after many years😊 She gave a wonderful keynote at the Railway Museum in Utrecht, as part of the National Conference on AI Transformations!💫

🎉 Exciting news from our team! The final paper of @aicoffeebreak.bsky.social's PhD journey is accepted at #ICLR2025! 🙌 🖼️📄 Check out her original post below for more details on Vision & Language Models (VLMs), their modality use and their self-consistency 🔥

aicoffeebreak.bsky.social@aicoffeebreak.bsky.social · 2y ago

The last paper of my PhD is accepted at ICLR 2025! 🙌 🎊 We investigate the reliance of modern Vision & Language Models (VLMs) on image🖼️ vs. text📄 inputs when generating answers vs. explanations, revealing fascinating insights into their modality use and self-consistency. Takeaways: 👇

We explain 🥥COCONUT (Chain of Continuous Thought), a new paper using vectors for CoT instead of words. We break down: - Why CoT with words might not be optimal. - How to implement vectors for CoT instead words and make CoT faster. - What this means for interpretability. 📺 youtu.be/mhKC3Avqy2E

COCONUT: Training large language models to reason in a continuous latent space – Paper explained

YouTube video by AI Coffee Break with Letitia

youtu.be

The last paper of my PhD is accepted at ICLR 2025! 🙌 🎊 We investigate the reliance of modern Vision & Language Models (VLMs) on image🖼️ vs. text📄 inputs when generating answers vs. explanations, revealing fascinating insights into their modality use and self-consistency. Takeaways: 👇

Do Vision & Language Decoders use Images and Text equally? How Self-consistent are their Explanations?

Vision and language model (VLM) decoders are currently the best-performing architectures on multimodal tasks. Next to answers, they are able to produce natural language explanations, either in post-ho...

arxiv.org

An educational and a bit historical deep dive into LLM research. 💡Learn what breakthroughs since 2017 paved the way for AI like ChatGPT (it wasn't overnight). We go through: * Transformers * Prompting * Human Feedback, etc. and break it all down for you! 👇 📺 youtu.be/BprirYymXrg

LLMs Explained: A Deep Dive into Transformers, Prompts, and Human Feedback

YouTube video by AI Coffee Break with Letitia

youtu.be

Transformer language models like Chat GPT, when using chain-of-thought reasoning, are Turing complete. Specifically, they can execute probabilistic algorithms and generate any computable weighted language youtu.be/MMIJKKNxvec?...

Transformer LLMs are Turing Complete after all !?

YouTube video by AI Coffee Break with Letitia

youtu.be

Don't forget to register! 🤭👇

EAVI - Media Literacy for Citizenship @eavi-medialiteracy.bsky.social · 2y ago

📅 Artificial Intelligence: Which skills do I need? 🎤 Dr. Letiția Pârcălăbescu and Paolo Celot 🕒 04 February 2024, 15:00-16:30 (CET, Brussels) 🖥️ Don’t miss this chance to boost your AI literacy and media skills! 🔗 Registration is required and can be done here: forms.gle/vhbSt2gzLe1m...

New video about: REPA (Representation Alignment), a clever trick to align diffusion transformers’ representations with pretrained transformers like DINOv2. It accelerates training and improves the diff. model’s ability to do things other than image generation (like image classification).

REPA Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You ...

YouTube video by AI Coffee Break with Letitia

youtu.be

Why does math feel so intimidating? Prof. Yael Tauman Kalai shares her insights, and we look at what this says about how we view intelligence🧠, both human🧑and AI🤖. 📺 youtu.be/Su1puD4xQwI

Why do people fear math? – Prof. Yael Tauman Kalai 🔴at #HLF24

YouTube video by AI Coffee Break with Letitia

youtu.be

How to make powerful LLMs understand graphs and their structure?🕸️ With Graph Language Models! They take a pre-trained LLM and fit it with the ability to process graphs. Watch if you're curious!👇 📺 youtu.be/JcHeaONGbmQ (Hint: it's about position embeddings, as the author explained at #ACL2024 🔴)

Graph Language Models EXPLAINED in 5 Minutes! [Author explanation 🔴 at ACL 2024]

YouTube video by AI Coffee Break with Letitia

youtu.be