Xiaoquan Kong’s Post

Very interesting. I like particularly the open approach, experimenting and exploring the different types of chatbot services that could be presented to users. We can clearly see the diversity of usages in and out of academia. For examples, in the teaching context, we could imagine teachers working with their students with notebooks to explore topics together. Several articles report that use of AI is poisoning teachers-students relations at the moment. Using GoTriple chatbot in classes could be a good way to overcome the challenge and create more positive and collaborative relations around AI in this context. Many other examples come in mind in other contexts. This is extremely exciting.

New article published - How can we leverage embedding techniques to measure how a particular domain is covered by a domain ontology? https://lnkd.in/d4wqf6Eq

Confused by the jump from CNNs to transformers for computer vision? You're not alone. Destin Gong shares a clear, interactive guide to key CV tasks using models like ViT, DETR, and BLIP, demystifying the new state-of-the-art.

Confused by the jump from CNNs to transformers for computer vision? You're not alone. Destin Gong shares a clear, interactive guide to key CV tasks using models like ViT, DETR, and BLIP, demystifying the new state-of-the-art.

“People don’t set out to have emotional relationships with these chatbots,” she says. “The emotional intelligence of these systems is good enough to trick people who are actually just out to get information into building these emotional bonds. And that means it could happen to all of us who interact with the system normally.” The paper, which is currently being peer-reviewed, has been published on arXiv.

🧩 Came across an interesting research paper today! 📄 “Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning” — by researchers from Cognée Inc., Innorenew CoE, and University of Primorska. The paper explores how systematic hyperparameter tuning can improve performance when integrating Knowledge Graphs (KGs) with Large Language Models (LLMs) — especially for multi-hop reasoning tasks (HotPotQA, TwoWikiMultiHop, MuSiQue). What I found intriguing 👇 They emphasize end-to-end optimization across chunking, graph construction, retrieval, and prompting. They use Cognée, a modular framework for KG–LLM pipelines. Results highlight the importance of tuning and the limits of standard evaluation metrics like EM, F1, and DeepEval correctness. 💬 Curious to hear what others think: How practical is this tuning-based approach for real-world RAG or reasoning systems? Do you see similar variability when combining graphs and LLMs in your projects? Could modular frameworks like this shape the next generation of explainable reasoning pipelines? #AIResearch #KnowledgeGraphs #LLMs #RAG #DeepEval #ComplexReasoning #GraphAI #Cognée #AIOptimization #MachineLearning cognee

The team’s implementation, an environment named Delethink, structures the reasoning chain into fixed-size chunks, breaking the scaling problem that plagues very long LLM responses. https://lnkd.in/ey_fwpFi

Does DQN Learn?” (https://lnkd.in/gF6AbbvY) by Aditya Gopalan (@today_itself) and Gugan Thoppe (@GThoppe) has been accepted to the IEEE Transactions on Automatic Control. In practice, DQN often fails to identify good policies—this work provides a rigorous explanation for why.

Tl;dr: The future of building with AI is less about being a prompt engineer and more about becoming an 'eval designer'. This was my main takeaway after watching a summary of Denny Zhou's Stanford lecture on the essence of AI reasoning. The video (https://lnkd.in/g_WcHRTc) offers a deep dive into the evolution of LLM thought processes. It highlights a paradigm shift from forcing LLMs to reason like humans via hyper-specific prompts, to instead crafting robust evaluations and letting the model find its own optimal path. This reframes my own work. I've found that over-engineering prompts has diminishing returns, especially for novel cases. The real leverage is simplifying the instructions and trusting the LLM to navigate complexity, guided by clear success criteria. #LLM #eval