Interview (in polish) here:

Our @_FedeSiciliano_ yesterday was interviewed by @hnonline, a Slovak newspaper, because of his expertise in AI. They discussed about the current status of AI, with a focus on the European Union vision of the topic. Bringing Italian style to Slovak 🇸🇰🇮🇹.#AI #Research #Recommender

Long day in the office for @andreag_difra00, presenting at #IJCNN2025:.A poster for the competition on Learning with Noisy Graph Labels .Link Prediction with Physics-Inspired GNNs (GNN session).And served as session chair for both events!.Link:

On July 4th from 11 to 12.30, room Belli @giuditeodoro will present the work “A theoretical analysis of recommendation loss functions under negative sampling”. On July 2nd will present the Competition on Learning with Noisy Graph Labels from 10.30 to 12.30, room Manzoni. #IJCNN25.

Next Monday will start the #IJCNN2025 conference! And our group will present lots of papers! On July 5th at 2.30 PM CET in Room Loyola @frances55037016 will present “Renormalized Graph Representations for Node Classification”. Preprint: #Classication #ML.

RT @FlorinCuconasu: 1/n.While positional bias affects LLMs, it's actually not a real problem for practical RAG systems. In our new paper,….

Paper:

4/4 These wavelet-like characteristics become more pronounced in larger models and align with optimal solutions to the fundamental uncertainty principle between positional precision and frequency resolution.

3/4 We discovered that attention heads naturally evolve to implement multi-resolution processing analogous to wavelet transforms, with heads organizing into complementary frequency bands.

2/4 Our research shows how transformer models develop robust wavelet-like properties that effectively compensate for the theoretical limitations of Rotary Position Embeddings (RoPE).

1/4 Excited to share that our paper "Beyond Position: the emergence of wavelet-like properties in Transformers" has been accepted to #ACL2025NLP! 🎉 @fabreetseo @ValeriaRuscio.

Arxiv:

"The Distracting Effect: Understanding Irrelevant Passages in RAG" co-authored by @FlorinCuconasu was accepted at #ACL2025NLP main conference!. This work analyzes how different types of distracting passages affect LLMs and achieves 7.5% accuracy gains via targeted fine-tuning.

Work dome by: @filippobetello @AntonioPuri00 @VinVit0920 @gtolomei and @fabreetseo.

With just 3 lines of code, our lightweight saver lets you update EcoTaskSet with new training dynamics. Simple, fast, and ready to use:.

GREEN: an inference-time method using our new dataset EcoTaskSet to recommend Pareto-optimal models balancing accuracy and emissions. Our UI lets users choose dataset & accuracy-emissions trade-off and recommends the best model.📚💻

Deadline extended!.The IR-RAG Workshop @ SIGIR 2025 has postponed its submission deadline to May 5. Don’t miss the chance to contribute to the future of Retrieval-Augmented Generation!.Submit: Info: #RAG #IR #SIGIR #AI #NLP #LLM.

📢 Our paper "A Theoretical Analysis of Recommendation Loss Functions under Negative Sampling" has been accepted at #IJCNN2025!. ✅ Done by @_FedeSiciliano_ , with @giuditeodoro , @ntonellotto , and @fabreetseo . 📍 See you in Rome!. #RecommenderSystems

🚀 Join us at IR-RAG @ SIGIR 2025 (July 17)!.Explore how IR powers Retrieval-Augmented Generation systems. 🧠💡. 📢 Submit your work by April 23: More info: #RAG #IR #AI #SIGIR2025`.

Link: