Samy Badreddine

12 posts

Samy Badreddine

@sbadredd

Research Scientist at SonyAI - PhD student at FBK Interested in Neurosymbolic AI and Probabilistic ML https://t.co/pQK2WIBhEd

Katılım Kasım 2023

37 Takip Edilen22 Takipçiler

Samy Badreddine retweetledi

Emile van Krieken@EmilevanKrieken·28 Kas

And finally #3 🔨 Rank bottlenecks in KGEs: At Friday's "Salon des Refuses" I will present @sbadredd's new work on how rank bottlenecks limit knowledge graph embeddings arxiv.org/abs/2506.22271

English

364

Samy Badreddine retweetledi

NeSy 2026@nesyconf·10 Eyl

@luislamb We're glad to announce the NeSy 2025 Test of Time award for "Probabilistic Inference Modulo Theories"! 🏆Rodrigo de Salvo Braz was here to accept the award. This is groundwork for recent NeSy approaches like DeepSeaProbLog and the probabilistic algebraic layer.

English

852

Samy Badreddine@sbadredd·9 Eyl

🎉Check our IJCAI'25 paper on #Explainable link prediction in #KG🔗 We use a path-based #RL model to generate explanations for predictions, rewarding paths that explain best ✨ And we've got user studies 🧑‍🔬 Great work led by Susana Nunes & @CPesquita👏 👉 arxiv.org/pdf/2509.02276

English

689

Samy Badreddine@sbadredd·1 Eyl

@orionweller @EmilevanKrieken Some relevant works from the Language Modeling literature + arxiv.org/pdf/1711.03953 + arxiv.org/pdf/2203.06462 @andreasgrv

English

Samy Badreddine@sbadredd·1 Eyl

@orionweller In our work, we address this using a Mixture-of-Softmaxes output layer to score and rank concepts, breaking the bottleneck. We wonder if this could be a promising direction for the retrieval tasks you highlight as well. Great to see this parallel exploration! @EmilevanKrieken

English

Orion Weller@orionweller·29 Ağu

Instructions/reasoning are now everywhere in retrieval - we want embeddings to do it all! 🚀 But... is it even possible? 🤔 Turns out, it's not possible for single-vector models 😱 theoretically and empirically! To make it obvious we OSS a simple eval SoTA models flop on! 🧵

English

324

33.3K

Samy Badreddine@sbadredd·1 Eyl

@deedydas In our work, we address this by adapting a Mixture-of-Softmaxes output layer to score concepts and break the bottleneck. We wonder if this could be a promising direction for the retrieval tasks you highlight as well. Great to see this parallel exploration! @EmilevanKrieken

English

Samy Badreddine@sbadredd·1 Eyl

@deedydas Insightful! We tackled the same problem in Knowledge Graph Completion. Dot-product scoring on low-dim embeddings severely limits what a model can predict. We call this a “rank bottleneck” to align with the existing LM literature. Our paper for context: arxiv.org/abs/2506.22271

English

382

Deedy@deedydas·31 Ağu

This new DeepMind research shows just how broken vector search is. Turns out some docs in your index are theoretically incapable of being retrieved by vector search, given a certain dimension count of the embedding. Plain old BM25 from 1994 outperforms it on recall. 1/4

English

385

4.2K

475K

Samy Badreddine@sbadredd·19 Eyl

I had the chance to visit @CPesquita group, hear more about their AI for Science projects and visit sunny Lisbon for a few days. Thanks for the warm welcome!

Catia Pesquita@CPesquita

This week, we were very proud to host three brilliant young researchers: @sbadredd, Christoph Wehner and @olgamshk. They gave super interesting talks and engaged in inspiring discussions with our PhD students. Thank you for visiting! @lasige

English