Clinical & healthpaperarXiv.org4 minResidencyRL: Reinforcement Learning in Simulated Clinical EnvironmentsResidencyRL is a reinforcement learning method for training clinical AI agents through simulated, multi-turn patient encounters with adversarial LLM simulators and rewards aligned to diagnostic accuracy, management, communication, documentation, and safety. In held-out evaluations, the agent improved diagnostic accuracy under adversarial conditions, reduced missed red flags, and outperformed the base model across reported clinical benchmarks, though real-world validation remains necessary.Why we saved thisThis paper is relevant to researchers and builders developing clinical AI agents because it evaluates sequential decision-making, safety, and generalization beyond static medical benchmarks using simulated encounters.Read ResidencyRL: Reinforcement Learning in Simulated Clinical Environments (opens in a new tab)Discover
12 posts
Ideas worth sharing.
Research, reporting, and ideas at the edges of health, biology, and intelligent systems.
Showing 12 of 12 posts
- Clinical & healthpaperarXiv.org4 minResidencyRL: Reinforcement Learning in Simulated Clinical EnvironmentsResidencyRL is a reinforcement learning method for training clinical AI agents through simulated, multi-turn patient encounters with adversarial LLM simulators and rewards aligned to diagnostic accuracy, management, communication, documentation, and safety. In held-out evaluations, the agent improved diagnostic accuracy under adversarial conditions, reduced missed red flags, and outperformed the base model across reported clinical benchmarks, though real-world validation remains necessary.Why we saved thisThis paper is relevant to researchers and builders developing clinical AI agents because it evaluates sequential decision-making, safety, and generalization beyond static medical benchmarks using simulated encounters.Read ResidencyRL: Reinforcement Learning in Simulated Clinical Environments (opens in a new tab)
Clinical & healtharticleresearch.google10 minSensorFM: Towards a general intelligence and interface for wearable health dataGoogle Research presents SensorFM, a foundation model for wearable health data pre-trained on more than one trillion minutes of multimodal sensor data from five million people. The model transfers across 35 health prediction tasks, supports label-efficient adaptation and data infilling, and can help ground a Personal Health Agent in individual physiology.Why we saved thisThis work is relevant to researchers and builders developing general-purpose health models, wearable-data applications, and AI health agents, particularly where labeled clinical data is limited.Read SensorFM: Towards a general intelligence and interface for wearable health data (opens in a new tab)Clinical & healthblogThe Guardian5 minScientists develop AI tool to spot heart failure risk five years before it strikesAn Oxford-developed AI tool predicted heart-failure risk five years before onset with 86% accuracy in a study of 72,000 patients in England.Read Scientists develop AI tool to spot heart failure risk five years before it strikes (opens in a new tab) Biology & genomicsarticlearXiv15 minGeneral Multimodal Protein Design Enables DNA-Encoding of ChemistryDISCO co-designs protein sequence and 3D structure around arbitrary biomolecules, producing enzymes that catalyse new-to-nature reactions without pre-specified catalytic residues.Why we saved thisA striking step beyond templates, theozymes, and inverse folding: joint diffusion over protein sequence and structure produced an evolvable enzyme in a single plate.Read General Multimodal Protein Design Enables DNA-Encoding of Chemistry (opens in a new tab) Clinical & healthblogMIT News5 minNew sensor sniffs out pneumonia on a patient’s breathMIT engineers developed a breath test that detects disease-related compounds and could diagnose pneumonia and other lung conditions within minutes.Read New sensor sniffs out pneumonia on a patient’s breath (opens in a new tab) Tools & infrastructurearticleLabClaw — Stanford / Princeton4 minLabClaw: A Skill Library for Autonomous Biomedical ResearchAn open-source library of 206 agentic skills spanning biology, clinical medicine, data science, literature search, and autonomous lab workflows.Why we saved thisA useful reference architecture for agentic lab infrastructure: modular skills inside a persistent runtime, shifting AI from a research tool toward a research operator.Read LabClaw: A Skill Library for Autonomous Biomedical Research (opens in a new tab) AI & intelligent systemsarticlearXiv15 minSycophantic Chatbots Cause Delusional Spiraling, Even in Ideal BayesiansA modelling and simulation study probing the causal link between chatbot sycophancy and users becoming dangerously confident in false beliefs.Why we saved thisFor clinical agents, sycophancy is a patient-safety concern: a model can reinforce mistaken beliefs about symptoms and shape how someone later presents to a clinic.Read Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians (opens in a new tab) Biology & genomicsarticleNature15 minAdvancing regulatory variant effect prediction with AlphaGenomeAlphaGenome interprets the non-coding regions of human and mouse genomes that regulate gene expression, predicting where genes begin and end, RNA output, and splicing behaviour.Why we saved thisAlphaGenome makes it practical to compare functional differences between normal and mutated genes, opening up connections between non-coding DNA and biological processes that previously required painstaking experimentation. Biology & genomicsvideoNature Video22:05CRISPR and the Future of Gene Editing in MedicineA look at how CRISPR-Cas9 and next-generation gene-editing tools are being used in clinical trials for genetic diseases, cancer, and beyond.Read CRISPR and the Future of Gene Editing in Medicine (opens in a new tab) 
AI & intelligent systemspaperGoogle Research35 minTowards Expert-Level Medical Question Answering with Large Language ModelsMed-PaLM 2 advances medical question answering, reaching expert physician performance on benchmarks including MedQA and PubMedQA.Read Towards Expert-Level Medical Question Answering with Large Language Models (opens in a new tab)
AI & intelligent systemspaperNature30 minLarge language models encode clinical knowledgeMed-PaLM demonstrated that large language models could approach physician-level performance on medical question-answering benchmarks.Why we saved thisThe paper that started the clinical LLM wave, and still a useful reference point for what these models can and cannot do in medicine.Read Large language models encode clinical knowledge (opens in a new tab)
AI & intelligent systemsarticleMIT Technology Review12 minAI is dreaming up drugs that no one has ever seen. Now we've got to see if they work.AI automation across drug development could produce faster, cheaper pharmaceuticals—but the real test is whether generated candidates work in humans.Why we saved thisA grounded counterweight to the AI drug-discovery hype. The bottleneck is not generating candidates; it is validating them in the clinic.Read AI is dreaming up drugs that no one has ever seen. Now we've got to see if they work. (opens in a new tab)