Activity
488 followers
-
Matt Morse reposted this🚀 Big congratulations to my colleagues at Qualcomm AI Research team for their NeurIPS 2025 paper on KeyDiff! As context windows in LLMs keep expanding, making them work on edge devices is a major challenge. Phones, IoT devices, and other constrained platforms need efficient AI — but long-context inference normally requires large memory for KV caches and suffers from high latency (for Time to First Token). 🔑 KeyDiff solves this by introducing a training-free KV cache eviction method that preserves the most important tokens while trimming redundancy. ✅ This opens up doors to more capable on-device AI for assistants, copilots, and real-time reasoning. ✅ Happy to say this innovation is coming soon to #Snapdragon platforms! #TeamQualcomm #GoQualcommAIMatt Morse reposted thisExcited to share that our work at Qualcomm has been accepted at NeurIPS 2025! 𝐊𝐞𝐲𝐃𝐢𝐟𝐟: 𝐊𝐞𝐲 𝐒𝐢𝐦𝐢𝐥𝐚𝐫𝐢𝐭𝐲-𝐁𝐚𝐬𝐞𝐝 𝐊𝐕 𝐂𝐚𝐜𝐡𝐞 𝐄𝐯𝐢𝐜𝐭𝐢𝐨𝐧 𝐟𝐨𝐫 𝐋𝐨𝐧𝐠-𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐋𝐋𝐌 𝐈𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 𝐢𝐧 𝐑𝐞𝐬𝐨𝐮𝐫𝐜𝐞-𝐂𝐨𝐧𝐬𝐭𝐫𝐚𝐢𝐧𝐞𝐝 𝐄𝐧𝐯𝐢𝐫𝐨𝐧𝐦𝐞𝐧𝐭𝐬. Our approach introduces a training-free KV cache eviction technique for LLM inference. By emphasizing geometrically unique keys that align with elevated attention scores, we enhance the preservation of crucial tokens within long-context LLMs when faced with resource limitations. It's exciting to learn that NVIDIA has integrated our method into the KVPress package on GitHub! Paper: https://lnkd.in/gJP732e3 Huge thanks to my amazing co-authors Dalton Jones Matt Morse Raghavv Goel Mingu Lee Chris Lott for their collaboration and insights that made this work possible.
-
Matt Morse reposted thisMatt Morse reposted thisExcited to share that our work at Qualcomm has been accepted at NeurIPS 2025! 𝐊𝐞𝐲𝐃𝐢𝐟𝐟: 𝐊𝐞𝐲 𝐒𝐢𝐦𝐢𝐥𝐚𝐫𝐢𝐭𝐲-𝐁𝐚𝐬𝐞𝐝 𝐊𝐕 𝐂𝐚𝐜𝐡𝐞 𝐄𝐯𝐢𝐜𝐭𝐢𝐨𝐧 𝐟𝐨𝐫 𝐋𝐨𝐧𝐠-𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐋𝐋𝐌 𝐈𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 𝐢𝐧 𝐑𝐞𝐬𝐨𝐮𝐫𝐜𝐞-𝐂𝐨𝐧𝐬𝐭𝐫𝐚𝐢𝐧𝐞𝐝 𝐄𝐧𝐯𝐢𝐫𝐨𝐧𝐦𝐞𝐧𝐭𝐬. Our approach introduces a training-free KV cache eviction technique for LLM inference. By emphasizing geometrically unique keys that align with elevated attention scores, we enhance the preservation of crucial tokens within long-context LLMs when faced with resource limitations. It's exciting to learn that NVIDIA has integrated our method into the KVPress package on GitHub! Paper: https://lnkd.in/gJP732e3 Huge thanks to my amazing co-authors Dalton Jones Matt Morse Raghavv Goel Mingu Lee Chris Lott for their collaboration and insights that made this work possible.
-
Matt Morse liked thisMatt Morse liked thisExcited to share that our paper "MeshFEM: A Block-accelerated Solver for Nonlinear Finite Elements" has been published in ACM Transactions on Graphics, and I'll be presenting it at SIGGRAPH 2026! We built a new solver that dramatically speeds up simulations used in computer graphics and geometry processing, like elastic body simulation with contact and surface parametrization. Our approach exploits hidden structure in these problems to accelerate every step of the underlying math, delivering up to 14x faster results than existing tools. Thank you to my coauthors, Xinzhuo Hu, Prof. Roi Poranne, and my PhD. supervisor Prof.Julian Panetta for their invaluable contributions to this work. https://lnkd.in/gFqxqx62 #SIGGRAPH2026 #ComputerGraphics #Simulation
-
Matt Morse liked thisMatt Morse liked this🚀 We’re hiring at Qualcomm AI Research! We’re looking for talented software engineers at multiple levels to help build and optimize cutting-edge Generative AI models and features for Qualcomm’s NPU and GPU accelerators. You’ll work alongside experienced engineers and researchers across AI, systems, software, and hardware. If you’re excited about working at the intersection of AI models, high-performance software, and edge computing, take a look at the open roles below. - https://lnkd.in/gPbTTXqF - https://lnkd.in/gi-XZwUz - https://lnkd.in/gG4VDsk5 - https://lnkd.in/gNMVeTxH #Hiring #SoftwareEngineering #GenerativeAI #EdgeAI #AIResearch #NPU #GPU #Qualcomm
-
Matt Morse liked thisLLM quantization and compression remain critical to personal AI. Check out the summary below from Markus for the latest research on LLM model efficiency at #ICML2026 from #Qualcomm AI Research team. #HiringMatt Morse liked thisI am excited to be at #ICML2026 in Seoul this week and present several of our recent works on model efficiency, quantization and LLM compression. Let me know if you are in town and want to catch up or stop by at one of our posters or invited talk. FPTQuant: Function-Preserving Transforms for LLM Quantization 👥 Authors: Boris van Breugel, Yelysei Bondarenko, Paul Whatmough, Markus Nagel 📅 Thu, Jul 9, 10:30 AM – 12:15 PM KST (poster session #6) 📍 Hall A, Poster #2905 🔗 ICML: https://lnkd.in/eYUffSzV 📄 arXiv: https://lnkd.in/eH_nk3Jg Dissecting Quantization Error: A Concentration-Alignment Perspective 👥 Authors: Marco Federici, Boris van Breugel, Paul Whatmough, Markus Nagel 📅 Thu, Jul 9, 5:00 PM – 6:45 PM KST (poster session #8) 📍 Hall A, Poster #2602 🔗 ICML: https://lnkd.in/euZgsJs5 📄 arXiv: https://lnkd.in/eDQrAysS Accurate Low Bit Quantization Through Function Preserving Transformations 🎤 Invited talk at #AdaptFM workshop 📅 Sat, Jul 11, 10:15 AM - 11:00 AM KST 📍 ASEM Ballroom 202 🔗 ICML: https://lnkd.in/e6PP5HkV Leech Lattice Vector Quantization for Efficient LLM Compression 👥 Authors: Tycho van der Ouderaa, Mart van Baalen, Paul Whatmough, Markus Nagel 📅 Sat, Jul 11, 01:00 PM or 02:30 PM KST 📍 ASEM Ballroom 202 📄 arXiv: https://lnkd.in/epHuR8NU Looking forward to catching up with many of you this week! P.S. If you are excited about these research directions, we are also hiring full time research positions in my team. https://lnkd.in/eHw5MsE6 #ICML2026 #LLMs #Quantization #ModelEfficiency #Compression #EfficientAI
-
Matt Morse liked thisMatt Morse liked thisExcited to be presenting 3 papers at ICML 2026 in Seoul! I'll be presenting work spanning speculative decoding, multi-token prediction, and accelerating diffusion LLMs. Main Conference 🔹 ConFu: Contemplate the Future for Better Speculative Sampling (with Zongyue Qin) 📍 Poster #2303 🗓 Wed, Jul 8 • 2:30–4:15 PM KST • Hall A Key idea: Explicitly generate and verify future-direction signals to improve token acceptance and robustness in speculative decoding. 🔹 Efficient Training-Free Multi-Token Prediction via Embedding-Space Probing 📍 Poster #191 🗓 Thu, Jul 9 • 2:30–4:15 PM KST • Hall A Key idea: Unlock latent multi-token prediction ability in frozen LLMs without retraining, draft models. Workshops (Jul 10) 🔹 Structuring The Future: Diffusion LLM Speculative Decoding via Calibrated Draft Graphs Presented at: Foundation Models for Generative Modeling Poster sessions: 9:10–10:00 AM, 2:00–2:50 PM Probabilistic Inference & Generative Modeling Poster sessions: 11:40 AM–1:30 PM, 4:10–5:00 PM A big thanks to all co-authors: Mukul Gagrani, Risheek Garrepalli, Sudhanshu Agrawal, Zongyue Qin, Mingu Lee If you'll be attending ICML, feel free to stop by the posters — happy to chat about efficient and scalable foundation models, learning dynamics, reasoning, generative modeling, and future directions in AI systems. #ICML2026 #MachineLearning #GenerativeAI #FoundationModels #LLMs #ReasoningModels
-
Matt Morse liked thisMatt Morse liked thisSo apparently the open source MILP solver HiGHS is used in Qualcomm NPU compiler to optimize VTCM (Vector Tightly Coupled Memory) which is usually done by heuristics. I wonder what kind of efficiency gain they got🤔
-
Matt Morse liked thisMatt Morse liked thisI just wrapped up my Fulbright PhD scholarship in the Arizona and made my way back to France, and I've been wanting to put into words what this experience meant. This trip gave me far more than I expected. Scientifically, I spent my time working on combining theoretical argumentation and large language models, and I'm leaving with more ongoing projects and ideas than I arrived with. Personally, I grew in ways I didn't anticipate, mostly through the people I had the chance to meet along the way. A few thank-yous are in order. To Mihai Surdeanu, who invited me and made time to meet with me so often, I always looked forward to our discussions, and I learned a great deal from you. And to Eduardo Blanco, thank you for the thoughtful exchanges we had. To everyone at the CLU Lab, thank you for the warm welcome. The moments we shared made this experience what it was. A special thank-you to Enfa Fane and Fahmida Alam for your help, but above all for your friendship. And thank you to the Computer Department of Computer Science at The University of Arizona for the poster session and picnic. To my partner Gaël Van der Lee and my friend Marianne THIERY who shared parts of this journey with me: thank you for being there. To Srdjan Vesic and Maxime BROUAT thank you both for being part of this journey from near and far. And finally, to the Fulbright France commission, thank you for your support throughout this experience. None of it would have been possible without you. Grateful doesn't quite cover it.
-
Matt Morse liked thisThe startup that I'm spending my sabbatical with, EliseAI, just hit $200M annual recurring revenue! There's lots of interesting applied ML / OR / AI problems here, and the company is hiring for both software engineering and research scientist roles. If you're in my network and interested in either of these things, shoot me a message :)Matt Morse liked thisThis quarter, EliseAI crossed $200M in annual recurring revenue, our fifth straight year of doubling. Our first $100M took years, the next $100M took twelve months. When we started, a lot of people told us housing and healthcare were the wrong industries to build in. Too regulated, too many workflows, too many edge cases, too much still happening over the phone and on paper. They were right. These industries are genuinely hard. That's exactly why we chose to serve them. Most AI companies are building for other tech companies. We're building for the industries every American depends on. Housing and healthcare account for more than 40% of the average household budget. If AI is going to create meaningful impact, it should start where people's lives are most affected. Before we wrote any code, I took a front desk job at a real estate firm. I answered phones, replied to emails, and worked directly with residents, leasing teams, and maintenance staff. It was the fastest way to understand what was actually broken. A missed renewal email becomes a notice to vacate. A family qualifies for an affordable apartment but loses it because no one can answer questions about a complex application process. A resident gets charged late fees because their payment method expired and nobody reached them in time. From a distance, these look like operational inefficiencies. Up close, they're moments that affect people's lives. Building for these industries isn't glamorous. Progress comes from years of listening to customers, solving real operational challenges, and earning trust one workflow at a time. It requires a team that treats the mission like it's their own. Today, 1 in 6 apartments in the United States runs on EliseAI, and this is just the beginning of what AI can do for these industries. We're hiring.
-
Matt Morse liked thisOur paper just won the Best Industry Paper Award at MLSys 2026! Building a framework that runs across a wide range of devices took years of unglamorous engineering from a lot of people. Happy to see the work recognized and grateful to the team and the hardware partner engineers who built this with us. Paper: https://lnkd.in/e22stMPwMatt Morse liked thisBig congrats to the ExecuTorch team. Their paper just won the Best Industry Paper Award at MLSys 2026. 🏆 ExecuTorch is PyTorch's framework for running models on-device, in production today across phones, wearables, desktops, and embedded systems. Backends include Apple, Arm, Qualcomm, MediaTek, Samsung, NXP, Cadence, Intel, NVIDIA, and others. Catch the team Tuesday morning in the Best Paper Session (8:45 AM PT) and at the poster session Thursday evening. Paper: https://lnkd.in/g5UmCu4D Discord: https://lnkd.in/g2e4xEcw X: https://x.com/Executorch Project: https://lnkd.in/grE63qJ8 Schedule: https://lnkd.in/gAkzQfPY ✍ Mergen Nachin, Digant Desai, Stephen Jia, Chen Lai, Mengwei Liu, Jacob Szwejbka, Raziel Alvarez, RJ Ascani, Dave Bort, Manuel Candales, Andrew Caples, Yanan Cao, Zhengxu Chen, Soumith Chintala, Gregory Comer, Tanvir Islam, Songhao Jia, Tarun Karuturi, Jack Khuu, Abhinay Kukkadapu, Tugsbayasgalan Manlaibaatar, Andrew Or, Kimish Patel, Siddartha Pothapragada, Lucy Qiu, Supriya Rao, Orion Reblitz-Richardson, Max Ren, Scott Roy, Anthony Shoumikhin, Scott Wolchok, Guang Y., Angela Y., Mengtao Yuan, Hansong Zhang, Jack Zhang, Zhenrui(Jerry) Zhang, Shunting Zhang, Cagatay Bilgin #MLSys2026
-
Matt Morse liked thisMatt Morse liked thisMost entrepreneurs fail for one reason. And that reason isn't what you think it is. It isn't: → Lack of funding → Wrong market → Bad timing → Bad ideas Most entrepreneurs fail because they never develop the discipline to do the unglamorous work consistently. → They're incredible at starting and terrible at sustaining. → They launch with energy and momentum, then reality sets in. The results don't come as fast as expected, the excitement fades, and without a system to fall back on, they stop showing up. It's never one catastrophic decision that kills a business. It's a thousand small abandonments. → The hard conversation that never got had. → The follow-up email that never got sent. → The content that never got posted. Failure is just inconsistency compounded over time. The entrepreneurs who make it aren't necessarily smarter or more talented, they just developed the discipline to show up when the excitement wore off. Success isn't built in the moments of inspiration, it's built in the moments when inspiration is completely gone and you show up anyway. Want daily frameworks to build the discipline that keeps you showing up long after the excitement fades? Check out BeBetter - https://lnkd.in/e_96B6ut
Experience
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Education
-
-
-
-
-
-
-
-
-
Other similar profiles
Explore more posts
-
James Brand
Wrote a second blog post! This time it's about using LLMs for imputing missing data. I've seen a few papers about imputation with LLMs, but most seem to either train custom imputation transformers or horserace flagship models against standard methods. Mert Demirer and I recently played with an alternative idea which instead uses LLMs as part of an ensemble imputation approach, relying on their "knowledge" of the world to provide additional prediction signal. I haven't seen a paper about it yet, so we thought it'd be fun to make a short post: https://lnkd.in/gMzC6iaP
8 Comments -
RISE Robotics
🤖 The continuum robotics challenge just got solved. Researchers have cracked a unified mathematical model for tendon-actuated concentric tube robots—a breakthrough that combines the best of two worlds. For years, engineers faced a trade-off: tendon-driven robots lacked flexibility in their degrees of freedom, while concentric tube designs suffered from instability. This new Cosserat rod-based framework handles multiple tubes with multiple tendons actuating each one, validated through experiments achieving <4% tip prediction error. The model is so robust it accurately predicts behavior across existing robots in the field. • Why it matters: Accurate modeling means better shape estimation and control—unlocking precision surgical applications, delicate manipulation tasks, and next-gen minimally invasive procedures. Follow for more breakthroughs in robotics and AI innovation! 🚀 #Robotics #ContinuumRobotics #Biomimetics #ControlSystems #RoboticSurgery #Innovation Source: https://lnkd.in/gQtDFW7Z
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top contentOthers named Matt Morse
283 others named Matt Morse are on LinkedIn
See others named Matt Morse