Chronological feed of everything captured from Karina.
VitaminP is a cross-modal learning framework that leverages paired H&E and multiplex immunofluorescence (mIF) data to transfer cytoplasmic boundary information into H&E-only inference, circumventing the inherent lack of cytoplasmic contrast in routine staining. Trained on 14 public datasets spanning 34 cancer types and over 7 million annotated instances, it outperforms four state-of-the-art segmentation methods and generalizes to unseen datasets including 24 rare cancer types. The work establishes cross-modal supervision as a reusable strategy for recovering biological structure that is absent or ambiguous in a target imaging modality, with practical deployment enabled via the open-source VitaminPScope platform.
This study presents a hybrid computational pipeline for wildfire spread modeling in Patagonia, Argentina, combining a Reaction-Diffusion-Convection (RDC) physical model with a Genetic Algorithm (GA) for parameter recovery and XGBoost for fine-tuning. The framework ingests high-resolution geospatial inputs — slope, wind velocity, and vegetation — and optimizes model parameters by maximizing spatial overlap between simulated and observed burned areas across three scenarios of increasing complexity. The GA reliably recovers reference parameters in all scenarios, while the XGBoost refinement step yields measurable accuracy gains particularly in lower-complexity cases. The approach addresses a core challenge in wildfire science: estimating difficult-to-measure physical parameters in a systematic, data-driven manner.
CyberScholar, a GenAI formative feedback tool using Retrieval-Augmented Generation (RAG) to integrate teacher rubrics and exemplars, was evaluated across 143 K–12 students and 5 teachers in a multiple-case study. Students reported tangible writing improvements through iterative, rubric-aligned feedback on organization, elaboration, and style, while teachers gained time savings and could shift toward higher-order instruction. However, the automated rating system exhibited inconsistencies and occasional misalignment with assignment expectations, surfacing a key limitation of criterion-specific scoring at scale. The findings suggest rubric-grounded GenAI feedback is viable for writing development but requires ongoing human calibration and contextual oversight to be reliably deployed.
AttuneBench introduces a multi-turn LLM emotional intelligence benchmark grounded in 200 real human-model conversations with turn-by-turn participant annotations, directly addressing the shortcomings of synthetic, single-turn, or third-party-annotated EI benchmarks. Evaluating 11 models across four dimensions — emotion recognition, behavioral classification, preference prediction, and response quality — reveals that these capabilities are largely independent, meaning a model strong in one area is not reliably strong in another. Crucially, preference alignment and response quality are far more model-discriminating than emotion-label accuracy, suggesting that true EI in LLMs is primarily about predicting the contextually appropriate response for a specific user, not just correctly labeling emotional states.