Researchers introduce MedicalRec, a transformer-based recommender system that recommends optimal models for medical image classification without retraining, reducing energy consumption and trial-and-error waste. The system uses a benchmark dataset MedicalRec-Bench with over 5,000 records from 3,000 articles, achieving a maximum HitRate@100 of 75.5%.

Offline reinforcement learning offers a promising route for developing plasma controllers from historical tokamak data, but progress is hindered by the lack of a standardized benchmark. This paper introduces RL4F, an offline RL benchmark for plasma control based on real DIII-D tokamak discharge data, covering four full-profile tracking tasks: rotation, density, temperature, and pressure. Evaluation shows that offline model-based RL methods achieve the best average performance on most objectives, though no single method dominates all tasks, highlighting the importance of dynamics modeling. The codebase, datasets, and evaluation framework are open-sourced.

The paper proposes a novel method called Worst Dimension Optimization for multimodal reasoning. Unlike current Process Reward Models that weigh all dimensions equally and may mask individual failures, this method focuses on optimizing the weakest dimension in the reasoning path, ensuring overall validity.

This study introduces GNOVA, a GRU-Neural ODE Variational Autoencoder, for bidirectional prediction of cognitive scores (CDR-SB and MMSE) in Alzheimer's disease using only routine clinical data, without MRI, PET, or CSF. The model achieves MAEs of 1.35 and 2.28 on the ADNI dataset, and feature ablation highlights age, BMI, and APOE4 as strong predictors. The framework enables reconstruction of incomplete patient histories and anticipation of future cognitive states.

A study using the Prithvi-EO-2.0 model across 19 global flood events (2017-2025) found that detection accuracy depends jointly on land cover and flood type: cropland showed highest agreement (IoU=52%) and riverine events strongest detection (F1=0.69), while tree cover and built-up areas had near-zero detection (IoU=4%). Dual-reference validation revealed some apparent model error is due to reference inconsistencies. 23 failure modes were identified, with pipeline engineering dominating initial error.

An interdisciplinary study explores the nature of chatbots as conversation partners for problem-solving. The authors argue that basic chatbots (consisting of an LLM and simple interface) cannot be genuine thinking partners because their training data only partially mimics human 'metaphorical problem propagations'. The conclusion aligns with Yann LeCun's view that current AI lacks human-like learning abilities, contrasting with Big Tech optimism. Despite limitations, the widespread use of chatbots makes understanding them socially and politically important.

A study evaluated LLaMA 3.1 for extracting structured data from Dutch brain MRI reports. The model showed high performance on categorical variables like visual rating scores but lower performance on numerical variables. Few-shot prompting improved numerical extraction accuracy significantly.

Researchers identify a 'concept bottleneck' in CoCoNuT latent reasoning where intermediate hidden states are overwritten, causing performance loss. They propose AGCLR with a persistent gated memory (write, read, forget gates) that consistently improves performance on GSM8K, HotpotQA, and ProsQA using GPT-2, with the gap widening as curriculum depth increases.

This empirical study assesses general-purpose coding agents on a fly optogenetics data-to-discovery pipeline. While agents can solve individual stages, they struggle when lacking predefined criteria, requiring scientific judgment. End-to-end automation remains beyond current capabilities.

PathoSage is a three-stage framework for patch-level multimodal reasoning in pathology, featuring Structured Evidence Deliberation to independently evaluate heterogeneous evidence and reduce anchoring bias, along with a training-free Beta-Bernoulli experience system for modeling tool reliability. Experiments show it effectively mitigates VQA hallucinations and classifier disagreement, outperforming strong baselines.

This article explains how to migrate GitHub Actions CI to Hugging Face Jobs to overcome limitations of GitHub-hosted runners, such as slow speed and lack of GPU access. By setting up a dispatcher Space, a GitHub App, and modifying the runs-on label, CI jobs can run on Hugging Face infrastructure with CPU or GPU hardware, streaming logs in real-time. Trackio's experience shows a ~30% reduction in CPU job time.

At WWDC 2026, Apple unveiled next-generation Siri AI with a custom Gemini-derived model and Vision-LLM technology to extract information from the user's screen, eliminating the need for app-specific integrations. The new Core AI library based on Meta's PyTorch enables developers to leverage Apple hardware for running models. iOS 27 Developer Beta is available, but access to Siri AI requires a waitlist.