What a great day to open 🦋: @adamgol.*.* it's been way too long #MyXAnniversary

Our research team is hiring PhD interns 🍏 Spend your next summer in Paris and explore the next frontiers of LLMs for uncertainty quantification, calibration, RL and post-training, and Bayesian experimental design. Details & Application ➡️

📢 We’re looking for a researcher in in cogsci, neuroscience, linguistics, or related disciplines to work with us at Apple Machine Learning Research! We're hiring for a one-year interdisciplinary AIML Resident to work on understanding reasoning and decision making in LLMs. 🧵

LLMs are notorious for "hallucinating": producing confident-sounding answers that are entirely wrong. But with the right definitions, we can extract a semantic notion of "confidence" from LLMs, and this confidence turns out to be calibrated out-of-the-box in many settings (!)

Come do a research internship with us at FAIR Coding in Paris or Tel Aviv 🤗 We've just released CWM ( https://t.co/nLXj35VSZV) and are now looking for strong students interested in working on the next generation of reasoning and coding models. Apply as soon as you can!

(1/n) Introducing KaVa ( https://t.co/xPyMoCtCSE) – the first latent reasoning framework leveraging compressed KV-Cache to guide the latent generation. We beat previous approaches, especially on a realistic, Natural Language GSM8K dataset.

We found a new way to get language models to reason. 🤯 No RL, no training, no verifiers, no prompting. ❌ With better sampling, base models can achieve single-shot reasoning on par with (or better than!) GRPO while avoiding its characteristic loss in generation diversity.

I love this line of research from my colleagues at Apple: Augmenting a language model with a hierarchical memory makes perfect sense for several reasons: - Intuitively the memory parameters should be accessed much less frequently than the weights responsible for reasoning. You

Introducing Pretraining with Hierarchical Memories: Separating Knowledge & Reasoning for On-Device LLM Deployment 💡We propose dividing LLM parameters into 1) anchor (always used, capturing commonsense) and 2) memory bank (selected per query, capturing world knowledge). [1/X]🧵

How do diffusion models generate images for prompts like "A cat eating sushi with chopsticks in the style of van Gogh" that were (probably) not seen during training? Models seems to compose known concepts (cat+sushi+style), but how?

Nice paper

Yeah we did exactly that

Uncertainty quantification (UQ) is key for safe, reliable LLMs... but are we evaluating it correctly? 🚨 Our ACL2025 paper finds a hidden flaw: if both UQ methods and correctness metrics are biased by the same factor (e.g., response length), evaluations get systematically skewed

I'll present my view on the future of uncertainties in LLMs and vision models at @icmlconf, in penal discussions, posters, and workshops. Reach out if you wanna chat :) Here's everything from me and other folks at Apple: https://t.co/MnSE4anJRS

Why does RL struggle with tasks requiring long reasoning chains? Because “bumping into” a correct solution becomes exponentially less likely as the number of reasoning steps grows. We propose an adaptive backtracking algorithm: AdaBack. 1/n

Can LLMs access and describe their own internal distributions? With my colleagues at Apple, I invite you to take a leap forward and make LLM uncertainty quantification what it can be. 📄 https://t.co/uhoCJfPdZK 💻 https://t.co/pQY1DfaKtS 🧵1/9

There’s a lot of confusion around uncertainty in machine learning. We argue the "aleatoric vs epistemic" view has contributed to this and present a rigorous alternative. #ICML2025 with @janundnik @eleanortrollope @markvanderwilk @adamefoster @tom_rainforth 1/5

Is the mystery behind the performance of Mamba🐍  keeping you awake at night? We got you covered! Our ICML2025 paper demystifies input selectivity in Mamba from the lens of approximation power, long-term memory, and associative recall capacity. https://t.co/dWDYyIWLzt

I recently gave a short talk at the International Workshop on Reimagining Democracy. The first half focused on feeling the AGI. The second half briefly outlined a new research direction I'm very excited about: leveraging AI to build unprecedentedly trustworthy institutions.

We are happy to welcome our next speaker to MLSS 2025! 🎤 @BarzilayRegina is a School of Engineering Distinguished Professor of AI & Health in the Department of Computer Science and the AI Faculty Lead at MIT Jameel Clinic. She develops machine learning methods for drug

Here is an RL perspective on understanding LLMs for decision making. Are LLMs best used as: policies / rewards / transition functions ? How do you fine-tune them ? Can LLMs explore / exploit ? 🧵 Join us down this rabbit hole... (ICLR 2025 paper, done at  ML Research)