Harvard and MIT to discuss Google paper on TUMIX

Most people say they want to be “ahead of the curve”. Very few are willing to read the papers that actually bend that curve. Over the last weeks I’ve started treating arXiv like a quiet, daily training ground: I picked the archives that matter to my work (for me: computer science) and subscribed by email. Every day, a stream of fresh ideas lands in my inbox: AI, systems, security, theory. Most of it I’ll never use directly. That is not the point. To keep it realistic, I pair it with a 10 minute pass through tools like PaperDigest, which highlight each paper in one sentence. It is just enough context to decide: “ignore, skim, or go deep”. No drama, no hustle culture. Just a calm habit of staying intellectually awake. This simple workflow is useful if you want to: stay aware of cutting edge research, get comfortable with reading papers, or even explore ideas for a potential PhD topic. https://lnkd.in/dUVKAg7s

#ThisDayInHistory | On October 29, 1969, the first computer-to-computer message was sent from the University of California, Los Angeles (UCLA) to the Stanford Research Institute (SRI), marking the beginning of the internet revolution. Programmers Charley Kline and Bill Duvall, under the guidance of Dr. Leonard Kleinrock, attempted to send the word “login,” but the system crashed after “lo.” This groundbreaking experiment became the first step in developing ARPANET — the network that eventually evolved into today’s global internet.

We don’t always need to throw more compute at problems. At TEDAI Vienna, we sat down with IBM’s Anna Topol to chat about the future of AI hardware. We also touched on the topic of why developers should keep their love for math alive. This is part of Off Stage, our new series of short conversations from TEDAI where we listen more than we talk and ask important questions about where the industry’s headed.

🔮 Future-proofing logic The 7B/8B range is a tipping point: small enough for edge inference and large enough to serve as a foundation for domain-finetuning (OCR, doc reasoning, or lightweight multimodal work). Open-source ecosystems (Qwen, LLaVA-Next, Mistral-Medium, Yi-VL) are optimizing around exactly that size. So, yes — it’s the smart “one-GPU-lab” target if you plan to grow your Proof-of-AI edge nodes without refitting every year.

🎉 Excited to share that we have two papers accepted to NeurIPS, including one Spotlight! 1) Parameter-Efficient Fine-Tuning for LLMs (lead: Jingjing Zheng) We propose an adaptive multi-subspace method for PEFT that outperforms SOTA on average accuracy while using far fewer trainable parameters. 2) Spotlight — Training Oblique Decision Trees at Scale (lead: Qiangqiang Mao) We introduce a highly scalable algorithm that boosts test accuracy by ~7% over strong baselines. Remarkably, a single optimal oblique tree trained by our method matches random forest accuracy while using hundreds of times fewer parameters. Huge thanks to all co-authors and collaborators for the hard work!

Many had asked me to explain why our recent work co-authored with ORCA Computing, Novo Nordisk and SiC Systems, Inc. is a breakthrough? Here it is for those of you looking for a layman answer. Enjoy. Link to full paper: https://lnkd.in/dmYMJ8RY Thanks to several colleagues of mine, I gave Google’s NotebookLM a try to create a popular summary of a very multidisciplinary and complex work we did. Pretty impressive and a good example of using AI tools for outreach and communication of scientific subjects. I had to do some additional prompting to fix a few minor mistakes but apart from that, it was created seamlessly. Rajiv Kailasanathan, Shawn Gibford, MohammadReza Boskabadi, Emmanouil (Manolis) Papadakis, William Clements, Christopher Savoie, Per Nyberg, Christopher Brown, Eric Reuthe

Good summary of the work with SiC Systems, Inc., Novo Nordisk and DTU - Technical University of Denmark on biomanufacturing anomaly detection using quantum AI.

I’m noting arXiv’s decision to stop accepting unvetted computer science preprints after a surge of low-effort, AI-generated submissions. This is a signal: uncurated information flows damage trust, derail discovery, and increase verification costs for product teams that rely on preprints for rapid innovation. For leaders, the actionable implications are clear: enforce provenance and quality gates, combine human expertise with automated triage, invest in metadata and reproducibility checks, and define KPIs that measure signal-to-noise and downstream validation costs. Treat research pipelines like product pipelines—prioritize curation, scalable governance, and feedback loops that convert raw output into reliable inputs for R&D. That approach preserves agility while protecting against wasted effort and reputational risk. I can translate this into a concrete roadmap for platform governance and research intake.

arXiv Mandates Peer Review for CS Surveys and Position Papers Amid Al Surge arXiv implemented a new policy on October 31, 2025, requiring literature surveys and position papers in its computer science category to undergo prior peer review at a journal or conference, with authors providing proof such as a DOl or acceptance letter to avoid rejection. This change addresses the surge in low-effort submissions fueled by generative Al tools, which have increased rejection rates to nearly 10% and overburdened moderators, while original research papers continue under light moderation. Researchers have mixed reactions, with some praising the quality control and others expressing concerns over delays in rapidly evolving Al fields.

🎥 Research Highlight: Heterogeneous LiFi–WiFi Network (Google’s Notebook LM (Language Model) Excited to share my research video created with Notebook LM, featuring insights from my published work - “Heterogeneous LiFi–WiFi with Multipath Transmission Protocol for Effective Access Point Selection and Load Balancing.” 💡 In this research, I developed an intelligent hybrid LiFi–WiFi model that integrates Deep LSTM for location prediction and the STIO algorithm for Access Point Selection and Load Balancing, achieving: ⚡ Ultra-low delay 🔋 High energy efficiency 📶 Seamless handover 📈 Enhanced throughput Bridging LiFi and WiFi for the next generation of smart, high-speed hybrid networks. https://lnkd.in/gtz9eA8s