How can we be sure a generative model (LLMs, Diffusion) has truly unlearned something? What if existing evaluation metrics are misleading us? In our new paper, we introduce FADE, a new metric that assesses genuine unlearning by measuring distributional alignment, moving beyond

I'm excited to share recent research I've been working on with my amazing co-authors since joining @thomsonreuters Foundational Research, tackling two critical challenges in LLM evals: 𝗿𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗮𝘂𝘁𝗼-𝗲𝘃𝗮𝗹s & 𝗰𝗼𝗺𝗽𝘂𝘁𝗲-𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗲𝘃𝗮𝗹s (Thread 👇)

Happy to introduce my internship work at @Google and @GoogleDeepMind, collab w/ @googlecloud. We introduce TIR-Judge, an end-to-end agentic RL framework that trains LLM judges with tool-integrated reasoning 🧠🛠️ 🔗 https://t.co/rtfqlvuzJ0 #Agents #LLMs #Judges #RL #reasoning

Launching our Research Lab : Advancing experience powered, decentralized superintelligence - built for continual learning, generalization & model-based planning. Press Release : https://t.co/iPYXb1nzYr We’re solving the hardest challenges in real-world industries, robotics,

New changes for ICML 2026: - attendance not required for acceptance - original submission published along side camera ready version - new reciprocal reviewing requirements

Let’s go to Seoul!

First principle of Context Engineering: Human-Machine Intelligence Gap — Humans naturally "fill in the blanks," machines don't. Context Engineering is fundamentally about entropy reduction, translating high-entropy human intent into machine-understandable signals. Every

𝗦𝘁𝗼𝗽 𝘄𝗿𝗶𝘁𝗶𝗻𝗴 𝗯𝗲𝘁𝘁𝗲𝗿 𝗽𝗿𝗼𝗺𝗽𝘁𝘀. Start engineering the system that feeds your LLM the right context at the right time. We've just released our new e-book on 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 going into details on exactly this 🔽 Download it for free

Cohere’s models and the fabulous @JayAlammar are hard at work, to help us explore all that NeurIPS 2025 has to offer!

This seems like a major breakthrough for AI advancement Tencent and Tsinghua introduced CALM (Continuous Autoregressive Language Models), a new approach that replaces next token prediction with continuous vector prediction, allowing the model to think in ideas instead of words.

Prompt Engineering is dead. The GAIR team just dropped Context Engineering 2.0 and it completely reframes how we think about human–AI interaction. Forget prompts. Forget few-shot. Context is the real interface. Their core idea: “A person is the sum of their contexts.”

Google DeepMind release: Towards Robust Mathematical Reasoning Introduces IMO-Bench, a suite of advanced reasoning benchmarks that played a crucial role in GDM's IMO-gold journey. Vetted by a panel of IMO medalists and mathematicians. IMO-AnswerBench - a large-scale test on

𝗔𝗜 𝗔𝗴𝗲𝗻𝘁’𝘀 𝗠𝗲𝗺𝗼𝗿𝘆 is the most important piece of 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴, this is how we define it 👇 In general, the memory for an agent is something that we provide via context in the prompt passed to LLM that helps the agent to better plan and

This was almost 4 years before the term "Foundational model" was even coined. Early Continual Learning research was genuinely ahead of its time. It was also nice to see an early synergy of nascent ideas in this paper (Bayesian CL + replay through Coresets) 🧠

🚨 RIP “Prompt Engineering.” The GAIR team just dropped Context Engineering 2.0 — and it completely reframes how we think about human–AI interaction. Forget prompts. Forget “few-shot.” Context is the real interface. Here’s the core idea: “A person is the sum of their

This broke my brain. A team at Sea AI Lab just discovered that most of the chaos in reinforcement learning training collapse, unstable gradients, inference drift wasn’t caused by the algorithms at all. It was caused by numerical precision. The default BF16 format, used across

How can we be sure a generative model (LLMs, Diffusion) has truly unlearned something? What if existing evaluation metrics are misleading us? In our new paper, we introduce FADE, a new metric that assesses genuine unlearning by measuring distributional alignment, moving beyond

This thought is converging from many sides. Transformer based LLMs are not going take us to human level AI. That famous Yann LeCun interview. "We are not going to get to human level AI by just scaling up MLMs. This is just not going to happen. There's no way. Okay, absolutely

@prfsanjeevarora As a researcher deeply interested in unlearning, this is a fascinating paper! Your "path-dependence" theory explains why true RE is theoretically impossible & why we observed the "universal failure" in our new work. We proposed FADE, a metric measuring distributional similarity

Going to San Diego for Neurips? We at @evaluatingevals, along with the UK @AISecurityInst are hosting a closed door state of evals workshop at @UCSanDiego on Dec 8th. Request to join below! :) https://t.co/45Wmjc0Lwo

Very excited with this new finding from our lab: machine unlearning as currently defined (i.e. model should behave as if it had never seen the unlearned data) may be impossible. Main reason: we show that outcome of unlearning --mathematically speaking, gradient ascent --is