Akshika (Aki) Wijesundara, PhD

New paper from #NVIDIAResearch 🎉 🧩 Puzzle: Distillation-Based NAS for Inference-Optimized LLMs+ Puzzle is a distillation-based NAS approach that transforms trained LLMs into far more efficient, inference-aware (heterogeneous) architectures. It works by replacing each Transformer block in the original model into a lightweight alternative, suitable for this specific block’s importance and role in the model. Some notable models created with Puzzle are Nemotron-Super-49B and Nemotron-Ultra-253B: https://lnkd.in/g6Szsc6y 📗 The Puzzle paper will be presented at ICML 2025: https://lnkd.in/gnzq6E3C

221

9 Comments

No pretraining. No datasets. Just pure inference-time gradient descent on the target ARC-AGI puzzle itself, solving 20% of the evaluation set. Introducing *ARC‑AGI Without Pretraining* Check out MLD PhD student Isaac Liao's technical blog post here: https://lnkd.in/eY_kvbQj

17

🚀 Scaling reinforcement learning (RL) for post-training LLMs is challenging, especially at large scale. The PyTorch team at Meta has open-sourced TorchForge, a PyTorch-native RL library designed to simplify large-scale post-training. In collaboration with Stanford University and CoreWeave, the team evaluated TorchForge on a 512-GPU cluster and shares what they learned about scaling RL infrastructure, running experiments with Weaver as the verifier, and enabling faster iteration on post-training workflows. 🔗 Read the blog: https://lnkd.in/g88ZMAb9 #PyTorch #torchforge #ReinforcementLearning #LLM #OpenSource #AIResearch

126

1 Comment

In today’s Google Research Blog post, we demonstrate how an LLM-powered coding agent can help discover new mathematical structures that push the boundaries of our understanding of complexity theory (a sub-field of theoretical computer science). We use AlphaEvolve, a system developed by Google DeepMind for iterative code evolution. Starting with populations of code snippets, AlphaEvolve evaluated the structures produced by them, and used an LLM to morph the most successful snippets toward better solutions. This approach led to two new results in complexity theory: 1) improving the state-of-the-art for the limit on our ability to approximate the outcome of the Maximum Cut problem for 4 slices, and 2) tightening the bounds on the average-case hardness of certifying properties of random graphs. Importantly, these improved proof elements can be automatically verified by a computer program. Read all about it at: https://goo.gle/3WhbCRI

456

3 Comments

Neural Machine Translation takes a whole sentence and processes (ie translates) it using one integrated model. And in this course, you'll learn all about the history and evolution of NMT via hands-on PyTorch replications of key papers. You'll learn about phrase representation, attention, transformers, and lots more. https://lnkd.in/gJDQvYs9

75

2 Comments

Our Cardiff NLP team and collaborators have been releasing amazing NLP models and datasets over the last few years. We've now organised the content as Hugging Face collections 🤗 Here is a short summary of these resources, in particular those related to social media applications: - TweetNLP. This is definitely our most popular collection 🔥 With millions of downloads every month and over 500 million downloads overall, this collection contains specialised NLP models for sentiment analysis, emotion detection, offensive language identification and more. - SuperTweetEval. This collection contains a benchmark with challenging datasets for NLP tasks in social media such as question answering, topic-based sentiment analysis, emoji prediction or tweet similarity. In addition to the datasets, we release custom models for each task. - Sensitive content. In social media we can find sensitive content such as hate speech, conflictual language, drug and sexual-related content, profanity, self-harm or spam. We release a dataset with these categories and models so you can customise your sensitive filters. - Sentiment analysis. Since sentiment analysis is one of the most popular tasks when it comes to social media, we’ve created a separate collection with some of our popular sentiment-related resources, including multilingual models and a unified benchmark for 8 languages 🌏 - Topic classification. With a topic taxonomy tailored for social media, we release datasets and multilingual models for topic classification. With these models, you can filter or analyse content related to specific topics such as sports, science, news, music and 15 others. - Hate speech detection. Unfortunately, hateful content is widespread across the web, and it’s hard to detect it without automatic tools. We have developed general-purpose models for hate speech detection, including to classify according to the target community. 🌐 Link to Cardiff NLP Hugging Face page: https://lnkd.in/dnxvVSix

108

3 Comments

Quantum needs data—not just qubits. aqora’s new public datasets hub marks a practical step forward for the field. By creating a central space to upload, access, and collaborate on quantum-specific datasets, they’re addressing a quiet but critical challenge in the ecosystem: reproducibility and shared benchmarks. Why does this matter? Quantum machine learning and benchmarking rely on consistent, high-quality data—but until now, much of it has been scattered, inconsistent, or locked in silos. This hub offers: ➡️ Versioned, well-structured datasets ➡️ Integration with tools like pandas and polars ➡️ Options for both public and private data sharing ➡️ A base for teaching, testing, and real-world use cases As the field matures, tools like this can help shift the conversation from isolated experiments to collaborative progress. 🔗 Full article in the comments. #QuantumComputing #QML #OpenData #QuantumResearch #TQI #TheQuantumInsider #Aqora #EcosystemBuilding

79

5 Comments

SFI welcomes Complexity Postdoctoral Fellow Max Jerdee Jerdee focuses on building interpretable, unbiased methods for analyzing network data. Using tools such as Bayesian inference and related probabilistic approaches, he develops models that identify network structures while rigorously quantifying uncertainty.  “I think a lot about how our methods shape what we believe we see in data,” he says. “If the instrument is biased, everything downstream will be too.” https://lnkd.in/emw_aqJ4

114

2 Comments

Impact Scholars Spotlight: Biological Connectivity Patterns as a Blueprint for Efficient Neural Architectures in Reinforcement Learning 🎉 Congratulations to this Impact Scholars team: Zahra Moradimanesh, Ruman Ahmed Shaikh, Nima Zargarnezhad, Austin Meek, Paolo Muratore 👏 And to the team's mentor Ilya Kuzovkin This team participated in the #NeuroAI steam of Neuromatch 's Impact Scholars Program and developed an interesting project: 📘 Biological Connectivity Patterns as a Blueprint for Efficient Neural Architectures in Reinforcement Learning Check out their final project: 🎬 Presentation: https://lnkd.in/g4T_pSGZ 🤩 General public summary: https://lnkd.in/gzr4Kk6p 📃 Publication: https://lnkd.in/gQFeEDvn 🧠 About Neuromatch's Impact Scholars Program: The Impact Scholars Program supports scientists in expanding the research they began during Neuromatch Academies—including Computational Neuroscience, Deep Learning, NeuroAI, and Climate Science. Over six months, scholars gain hands-on experience, mentorship, and the chance to translate data into a real-world research project. Stay tuned as we share more projects from the 2025 cohort!

45

2 Comments

💡“You can’t avoid reinforcement learning because the problem you care about solving is just everywhere in AI.” — James MacGlashan, Senior Research Scientist, Sony AI In our latest blog, the Sony AI Reinforcement Learning team opens up about why RL remains the hardest and most important problem in AI today. From GT Sophy to robotics, these researchers are solving real-world challenges that go far beyond passive pattern recognition. This August, we were also a proud sponsor at the recent Reinforcement Learning Conference (RLC 2025). 📚 Read the full interview: https://bit.ly/4myLn4K #SonyAI #ReinforcementLearning #MachineLearning #AIResearch #GTsophy #Robotics #AI

51

The next wave of AI breakthroughs won't come from scaling alone. They'll come from understanding intelligence itself.   We're partnering with UCL's Gatsby Computational Neuroscience Unit to fund 4 PhD scholarships starting in 2026.   The Gatsby Unit has spent decades asking foundational questions about learning, representation, and intelligent systems. The kind of questions that don't always have clean answers. The questions that tend to matter.                We're interested in how complex systems learn under uncertainty. Markets are one example. Brains are another. Both process incomplete information. Both adapt to changing environments. Both reward models that can separate signal from noise.   Our research and infrastructure are built around that reality.   This partnership builds on the Jump Trading Fellowships we established in 2025 and our broader support for research across mathematics, physics, AI, computer science, and the sciences.   👉 Discover how we're building the future of AI/ML at Jump, and who's building it with us: https://lnkd.in/g9GPGiCE   👉 Read about our partnership here: https://lnkd.in/ekcHZaNh

314

5 Comments

Want to build a better understanding of how LLMs work? Shuyang Xiang’s hands-on guide provides a practical experiment to find universal neurons, helping you to interpret the function of neurons within a transformer’s MLP block.

10