Skip to main content
Search

University of Catania

3 related articles

Subscribe to updates
  1. arXiv

    EgoTools pairs 100 hours of first-person tool-use video with a 1,000-question benchmark, exposing video models' weak tool perception and causal reasoning

    The work introduces EgoTools, a suite comprising EgoTools-Data, roughly 100 hours of tool-centric egocentric recordings across kitchen, classroom, research lab, repair workshop, craft, office, and household domains with synchronized audio, 361,332 hierarchical captions, 6,519 tool-centric narrations, and supplementary 3D information, plus EgoTools-Bench, a 1,000-question 8-way diagnostic benchmark of 900 human-crafted and 100 human-verified spatial questions across Affordance & Causality, Perception & Grounding, Procedural Dynamics, and Spatial Reasoning; evaluation shows Gemini-3.1-Pro at 66.9% overall but only 51.7% on Perception & Grounding, open-source instruct models between 38.9% and 50.

  2. Artificial Intelligence in Ophthalmology: From Diagnostic Accuracy to Clinical Application

    This paper assesses why high-performing artificial intelligence systems for ocular image processing seldom translate into improved patient outcomes, locating the central problem in the disparity between pixel-level performance metrics and their clinical significance, naming data bias, domain shift, and label noise alongside the lack of prospective randomized deployment trials as primary obstacles, and outlining a path through stringent external validation, established decision criteria, ongoing surveillance in real clinical practice, transparent reporting standards, and deliberate human-factors engineering, with the goal of converting algorithmic accuracy into meaningful diagnostic precision for glaucoma, diabetic retinopathy, and macular conditions (specifically diabetic macular edema and

  3. bioRxiv

    LLMsFold: Integrating Large Language Models and Biophysical Simulations for De Novo Drug Design

    This work presents LLMsFold, a computational framework that identifies binding pockets geometrically, has Llama-3.3-70B generate candidate small molecules as SMILES strings, evaluates each with the Boltz-2 co-folding model for bound pose and binding affinity, and iteratively refines candidates through a feedback loop, yielding molecules for ACVR1 and CD19 that pass drug-likeness, synthetic accessibility, and novelty filters, with the ACVR1 candidate reaching a predicted affinity probability of 0.953 and predicted pIC50 of about 10.72 and the CD19 Pocket 1 candidate reaching a predicted pIC50 of about 7.73.

You’re all caught up