wing.nus

📄 Paper: sentic.net/ai-overload.pdf ✍️Feedback and discussion from the community is welcome. Huge thanks to Philipp for the inspiring talk! @Philipp_Mayr 🧵5/5

🔍 Key insight: In AI systems, the bottleneck is no longer generation but oversight. As AI autonomy grows, validation, monitoring, and accountability burdens increase. Mitigation requires bounded autonomy, calibrated reliance, governance, and AI literacy. 🧵4/5

We were delighted to host Philipp Mayr (GESIS – Leibniz Institute for the Social Sciences) at our group meeting! 🎉 He presented the Knowledge Technologies department, the Information & Data Retrieval team, the OMINO project, and two Scholarly Document Processing projects. 🧵1/5

Bottom line: Better fact-checking isn’t about more web access. It’s about finding what’s missing. Search the gaps → better Notes. If you’re working on misinformation, moderation, or Community Notes, we’d love your thoughts. (6/6)

Human evaluation (N=100): 🏆 69% win rate vs human-written helpful notes 📈 Helpfulness: 3.87 vs 3.36 📌 Biggest gain: better context Also, 59% win vs generic web agents. Search strategy matters. (5/6)

🚨 Can smarter search beat generic web-search LLMs for Community Notes? Same model. Same web access. Different search strategy. We built GitSearch to test this. 📄 arxiv.org/abs/2602.08945 (1/6)

📌 Final thought Trust is a learnable, usable signal for LLM agents. Towards reliable multi-agent systems, we must teach agents who to believe — not just how to reason. 🔗 arXiv: arxiv.org/abs/2601.21742 🏠 Github: github.com/skyriver-2000/… 💬 Thoughts welcome! 🧵6/n

⚠️ Why this matters Without trust modeling: ❌ agents collapse under social pressure ❌ confident hallucinations dominate ❌ adversarial peers win With ECL: ✅ agents resist blind conformity ✅ trust becomes an explicit reasoning signal Analytic experiments verify this 🧵5/n

🔥 Thrilled to announce ECL, a framework for LLMs to reason with trust in multi-agent systems 📖 Key Takeaways •We introduce interaction history for LLMs to judge peer reliability and selectively refer to them •We decouple trust estimation and conditioned decision-making 🧵1/n

🔮Meet IzzyViz, a new way to see how transformer models think. 🔥Attention heatmaps | 🧠Head & layer comparison | 📊Run-to-run stability | ⏳Training-time evolution | 🎯Key-region detection 🔗Repo: github.com/WING-NUS/IzzyV…

📢 Excited to share our accepted EMNLP 2025 papers from the NUS WING group! 🎉 See you in Suzhou! #EMNLP2025

Led by @yajing08042. Huge thanks to Prof Kan Min-Yen (@knmnyn) and Tony Deng for their support! See you at #EMNLP2025! 📜 Paper: arxiv.org/abs/2509.17037 💻 Code: github.com/yajingyang/kah… #NLP #DataToText #Finance #FinNLP #LLM #AI 🧵[5/n] n=5

High-quality knowledge can be "distilled"! We used GPT-4o to generate a knowledge base for a smaller Llama3.1-8B. This "distillation" significantly boosted its performance, enabling efficient, high-quality narration. 🧵[4/n]

Thrilled to share that our paper, "KAHAN: Knowledge-Augmented Hierarchical Analysis and Narration," is accepted at #EMNLP2025 Findings! In this work, we built a framework that uses LLMs as domain experts to hierarchically extract insights from tables. 🧵[1/n]