We recently hosted an insightful talk by Gorjan Radevski, Researcher at NEC Laboratories Europe, on compositional steering tokens – a new method for guiding large language models (LLMs) to follow multiple behaviors simultaneously by embedding behavioral instructions directly into input tokens. Gorjan explained how these tokens generalize to unseen behavior combinations and outperform existing steering approaches across different LLM architectures. To learn more, watch: https://lnkd.in/dQT4eJwv. #NECLabs #AI #largelanguagemodels
NEC Laboratories Europe’s Post
More Relevant Posts
-
The Ideal Text Community in the Era of AI A theoretical framework extending Jürgen Habermas's "ideal communication community" into the age of machine-mediated, planetary-scale textual discourse—and the design principles required to realize, rather than merely simulate, uncoerced rational deliberation. Transitioning this framework to an AI-mediated paradigm necessitates a critical re-evaluation of communicative action. As algorithmic intermediaries begin to shape the linguistic constraints and access points of public discourse, the primary challenge lies in preventing the instrumentalization of reason by opaque optimization models. We must interrogate whether current large-scale language systems facilitate the mutual understanding constitutive of a truly democratic public sphere or whether they inevitably incentivize performative rhetoric over substantive engagement. The stakes of designing these planetary-scale deliberative systems involve more than mere technical interface optimization; they demand an institutional architecture capable of fostering uncoerced consensus. By embedding normative democratic requirements—such as transparency, discursive symmetry, and the suspension of status-based power dynamics—directly into the infrastructure of machine-mediated interaction, we may construct a digital agora that serves as a robust foundation for contemporary collective intelligence. https://lnkd.in/eYUyJ6QQ #Habermas #IdealTextSituation #DeliberativeInfrastructures
-
New research from the College of IST and Massachusetts Institute of Technology finds that AI chatbots become more agreeable and begin mirroring users' views over extended conversations — even when it means sacrificing accuracy. In a real-world study of five large language models, 4 of 5 grew more sycophantic when storing user memory profiles, raising concerns about echo chambers and misinformation 💻📝 Read more: https://loom.ly/E2Gf8UY #PennState #PennStateIST #informationsciencestechnology #WeAre #PSU
-
-
The complexity of modern systems, particularly with large language models (LLMs), extends far beyond just the number of tools implemented. Protocols define a much broader surface area, including resources, prompts, and sampling flows that allow servers to delegate work to the LLM. Furthermore, concepts like Roots and Elicitation primitives introduce even more layers of potential complexity. It's crucial to understand that the current tool count is merely a baseline, not a limit. This expansive nature means the actual system footprint can be significantly larger than initially perceived. Ignoring these deeper protocol elements can lead to underestimating system requirements and potential challenges. #SystemDesign #LLM #SoftwareEngineering #AI #TechInnovation
-
Do AI models actually reason or do they just sound like they do? In this MI Colloquium, SC&I Alumnus Debanjan Ghosh (PhD 2018) explores how small language models (SLMs) handle commonsense reasoning. Do SLMs truly understand everyday physical and social logic or simply produce convincing text? Join us to see what Debanjan's research reveals on March 12. Online via Zoom - REGISTER-https://ow.ly/NSjv50YpKJY
-
-
AI is no longer just answering questions. It’s starting to think with us. 🧠⚡ NotebookLM + long-context LLMs may change how humans read, learn, and create knowledge. Read here: https://smpl.is/ai6it #AI #NotebookLM #FutureTech
-
AI is no longer just answering questions. It’s starting to think with us. 🧠⚡ NotebookLM + long-context LLMs may change how humans read, learn, and create knowledge. Read here: https://lnkd.in/djgpAAEC #AI #NotebookLM #FutureTech
-
AI is no longer just answering questions. It’s starting to think with us. 🧠⚡ NotebookLM + long-context LLMs may change how humans read, learn, and create knowledge. Read here: https://lnkd.in/djgpAAEC #AI #NotebookLM #FutureTech
-
⚡ RAG vs Fine-Tuning — Which One Should You Use for LLM Applications? One of the most common questions in Generative AI today is: 👉 Should we use Retrieval-Augmented Generation (RAG) or Fine-Tuning? Both approaches improve Large Language Models, but they solve different problems. #GenerativeAI #LLM #RAG #FineTuning #ArtificialIntelligence #MachineLearning #AIEngineering #AIArchitecture #DataScience #VectorDatabase #AIInnovation #TechTrends
-
-
Most people still use AI to ask questions. The real power is using AI to think. 🧠🔥 NotebookLM shows what knowledge engines will look like. Full article: https://lnkd.in/djgpAAEC #ArtificialIntelligence #SecondBrain #LLM
