AI safety isn’t universal.
#UbuntuGuard, led by researchers at @brownpublichealth.bsky.social and collaborators, shows that models that pass English benchmarks can fail when tested against locally grounded policy rules across African languages.
Read more:
rebootdemocracy.ai/blog/ubuntu-...
6/n: We release:
- The dataset for KG-based inductive reasoning: huggingface.co/Tassy24
- Paper: arxiv.org/abs/2502.13344
- Repo: github.com/rsinghlab/K-Pa….
Big thanks to all my incredible collaborators- this wouldn’t be possible without your brilliance!
5/n: This open doors in biomedical AI:
• Structured KG reasoning + LLMs' generative power = transparent, scientific discovery.
• Imagine helping scientists understand why a treatment works, not just what!
4/n: Why this matters:
• Interpretability baked in - paths double as explanations.
•Efficiency - huge KG reductions help scale to real-world biomedical workloads.
•Generalization - works in unseen (inductive) settings.
3/n: 🔥 Results across model families (LLMs & GNNs):
• Tx-Gemma 27B: +19.8 F1 on interaction severity
• Llama 70B: +8.5 F1 on similar tasks.
• EmerGNN: trains on 90% smaller KG with no loss in accuracy
• That cross-architecture gain is rarely seen and hard to ignore.
2/n: K-Paths is not just a retrieval method.
It’s a pipeline that combines:
• Heuristics path finding (Yen’s algorithm)
• Diversity-aware selection
• Graph pruning
• Natural language transformation to aid reasoning for LLMs.
•GNN/LLM integration
1/n: Most KG methods for drug discovery tasks rely on GNNs or focus on narrow QA tasks.
K-Paths is model-agnostic and reframes KG use for reasoning & discovery, especially in inductive settings (e.g. emerging drugs/diseases).
Thrilled to share our KDD 2025 paper: Can large language models (LLMs) reason like biomedical scientists?
We introduce K-Paths, a retrieval framework for extracting reasoning paths from knowledge graphs (KGs) to aid drug discovery tasks.
👇 Thread:
Amazing! Congratulations!!!
Speech tech is everywhere, even healthcare. But how well does it work for African accented English? 🤔
We benchmarked SOTA models for speaker diarization, ASR & LLM summarization on medical & general conversations.
Find me at the 11 am poster session in Hall 3 to learn more!
#NLP4Healthcare
I am pleased to share the report of the Text Retrieval Conference (TREC) Biomedical Generative Retrieval (BioGen) Track, a challenge evaluation to assess reference attribution in LLMs for clinical questions.
arxiv.org/abs/2411.18069
Thank you for responding. I have not done UX work but my research focuses on human-centered AI, exploring how AI can be adapted to better serve human needs across various applications. I’m happy to share more details of my work in a less constrained setting. Here is my resume: tassabdul.github.io
Hi Dr @mihaelav.bsky.social, I am super interested. My research lies at the intersection of NLP and IR aiming to derive factual insights from extensive literature or knowledge graphs to support accurate, reliable, and evidence-based decision-making in various domains. Thanks your consideration.
I’d like to join. Thanks!
Hi Dr @boghuma.bsky.social , I am computer science researcher and I would love to be included in the starter pack. Thanks.
Thank you Dr. @kevinkaichuang.bsky.social for the follow back! As my first follower you deserve a shoutout. Have a great week ahead!
