Google Research just published a framework that quietly undercuts one of the more expensive assumptions in enterprise AI procurement: that better performance requires a bigger, newer model. WikiSkill gives an LLM agent a persistent "wiki" — a structured, compounding knowledge base that sits between raw execution logs and the reusable "skills" (packaged instructions and workflows) the agent actually runs. A four-agent loop does the work: an Inference Agent executes tasks, a Wiki Maintainer performs root-cause analysis on what succeeded and failed and writes the lessons into the wiki, a Skill Proposer drafts skill updates informed by that accumulated knowledge, and a Gating mechanism validates each update before accepting or rolling it back. The critical design choice is that the wiki itself is never rolled back, even when a skill update is rejected — failed attempts and their root causes are preserved permanently, so the same mistake is never proposed twice.
The headline numbers hold up under scrutiny. Tested across five benchmarks (math reasoning, web search, spreadsheet manipulation, long-document QA, and embodied tasks) and five models spanning 4B to 27B+ parameters, WikiSkill beat the strongest existing skill-evolution method by 3.3 to 12.0 points on average, and beat the no-skill baseline by as much as 39.6 points — Gemini-3.5-Flash went from 33.0% to 72.6% on math reasoning and from 50.5% to 76.6% on spreadsheet manipulation using the exact same underlying model, purely from accumulated procedural knowledge.
The number that should get a board's attention isn't the top-line lift — it's the substitution effect. Qwen-3.5-9B equipped with a WikiSkill wiki (47.4% average) beat Qwen-3.6-27B with no skills at all (39.4%) — a smaller, cheaper model with accumulated institutional knowledge outperforming a substantially larger one running cold. And the knowledge is portable: skills evolved by one model transferred effectively to different models and even different model families, in several cases outperforming the skills a model developed for itself. A skill library built on a 27B model lifted a 9B model to 70.2% on an embodied-task benchmark, versus 63.4% using its own self-evolved skills.
What this actually means for the "which model should we standardize on" debate:
Health systems evaluating clinical AI agents — documentation, prior auth, triage support — routinely frame the decision as a model choice: which frontier LLM, which vendor, which version. WikiSkill's result suggests that framing undersells the actual asset. The accumulated wiki of what worked, what failed, and why — built through real deployment experience — degrades much less gracefully than the underlying model choice, and in several tested configurations it was worth more than a full parameter-scale upgrade. That reframes a live procurement question: is the vendor selling a model subscription, or an accumulating knowledge asset that happens to run on a model? Those are different products with different lock-in profiles, and only one of them should get more valuable the longer a health system uses it.
The caveats matter here as much as the numbers, and the authors are direct about them. Transfer is not universal — skills built by a smaller model sometimes encode narrow workarounds that actively degrade performance on a larger one; one configuration dropped Gemini-3.5-Flash's spreadsheet accuracy from 50.5% to 18.1% when it inherited a smaller model's skill set. This is a Google preprint, not yet peer-reviewed, and every benchmark is a controlled task — math problems, spreadsheets, web search, a simulated household — none of them a clinical workflow, so the read-across to documentation or triage agents is inferential, not demonstrated. Skills were also injected directly into the prompt rather than retrieved from a growing library, so the retrieval-at-scale question — what happens once the wiki holds hundreds of accumulated skills instead of a handful — remains untested.
For an advisory practice, the actionable question isn't whether to believe the headline lift. It's whether a health system's AI vendor contract gives them ownership or export rights over the accumulated knowledge layer their own deployment generates — the skill library, the failure logs, the institutional "wiki" — or whether that knowledge is locked inside a specific model version the vendor can deprecate on its own timeline. That is a contract and architecture diligence question boards should be asking before the next model-upgrade cycle, not after.
Pre-Print Intelligence (arXiv)
LUCAID: Agentic Multimodal AI for Lung Cancer Precision Pathology
Brief: LUCAID is an agentic multimodal AI system that integrates nine pathology modules—including quality control, tumor detection, subtyping, microenvironment profiling, and biomarker scoring—to support precision lung cancer diagnostics. It achieved F1 scores of 0.82–0.95 against expert ground truth and demonstrated 93.0% concordance with an expert-panel reference standard in prospective clinical validation, surpassing the performance of experienced thoracic pathologists.
Methodological Integrity: The study relies on prospective validation but does not disclose the geographic diversity or sample size of the validation cohort, raising potential spectrum bias. Additionally, the agentic architecture introduces complexity that may obscure failure modes and limit external generalizability without further multi‑center testing.
Strategic Implication: If cleared for clinical use, LUCAID could reduce inter‑observer variability and accelerate reporting in lung cancer pathology, addressing a high‑volume diagnostic bottleneck. However, its impact will depend on seamless integration with existing laboratory information systems and demonstration of cost‑effectiveness across varied healthcare settings.
Executive Summary: LUCAID combines agentic reasoning with nine specialized modules to automate lung cancer pathology workflow. In prospective validation it reached 93% concordance with expert consensus, outperforming individual pathologists.
Innovation: 8/10 | Applicability: 8/10 | Commercial Viability: 7/10
STRIVE: Multi-Agent Structured Temporal Reasoning with Integrated Verification for Longitudinal Radiology Report Generation
Brief: STRIVE proposes a multi-agent framework that separates diagnosis, attribute estimation, and temporal change reasoning into specialized modules, each producing explicit evidence. The Temporal Change Agent is refined with Progression-Aware GRPO, a shaped reward that penalizes direction reversals more heavily than missed changes. Verification is enforced via a deterministic Consistency Gate and a Validation Agent that checks report fidelity to aggregated evidence.
Methodological Integrity: Evaluation relies on the Longitudinal-MIMIC dataset, which may suffer from limited diversity and potential label noise; external validation on heterogeneous institutional data is absent. The study does not report ablation of agent interactions or potential data leakage between training and test splits.
Strategic Implication: If validated clinically, STRIVE could reduce radiologist workload and improve report consistency, but adoption hinges on seamless integration with PACS/RIS workflows and demonstration of real‑world error reduction. Its current form remains a research prototype without clear path to ambient, proactive deployment.
Executive Summary: STRIVE introduces a modular, verifiable approach to longitudinal radiology report generation, achieving superior performance on Longitudinal-MIMIC. The method emphasizes explicit intermediate evidence and structured temporal reasoning to improve traceability.
Innovation: 9/10 | Applicability: 7/10 | Commercial Viability: 6/10
Speech-to-SOAP: End-to-End Summarization of Medical Dialogues: KIT@BeTraC 2026
Brief: The paper proposes an end-to-end model that converts clinician-patient dialogue directly into SOAP notes, bypassing intermediate transcription. It uses a scalable data augmentation pipeline that generates synthetic speech and automatically creates SOAP labels to adapt a speech foundation model for this task. The approach aims to preserve paralinguistic cues and reduce documentation burden.
Methodological Integrity: Evaluation relies on synthetic augmentation and limited real-world dialogue datasets, raising concerns about generalization to diverse clinical accents and settings. No external validation or prospective clinical trial is reported, and potential bias from automatically generated SOAP supervision is not addressed.
Strategic Implication: If proven robust, the technology could support ambient clinical documentation, aligning with trends toward proactive, screen‑less AI in orthopedics and MSK care. However, widespread adoption would require integration with EHRs, rigorous clinical validation, and compliance with privacy regulations.
Executive Summary: The study presents a novel end‑to‑end speech‑to‑SOAP summarization method using synthetic data augmentation. It demonstrates feasibility on benchmark datasets but lacks real‑world validation.
Innovation: 6/10 | Applicability: 8/10 | Commercial Viability: 8/10
MyoMechanix: Biomechanically-Grounded Compositional Skilled Activity Understanding and Coaching
Brief: MyoMechanix introduces a multimodal dataset linking video, 3D pose, and surface electromyography for weight-loaded actions, accompanied by a Fitness Knowledge Graph and the CUBIST reasoning engine for compositional action quality assessment. The system enables fine-grained error attribution and feedback by aligning motion with muscle activity patterns. It also proposes a Video2EMG task to infer muscle signals from video alone.
Methodological Integrity: The dataset comprises 7,500+ samples from only 38 subjects, limiting demographic diversity and raising concerns about overfitting to a narrow population. Validation relies primarily on internal benchmarks; external clinical or real-world testing is absent, and sEMG signal quality can vary with electrode placement and skin conditions.
Strategic Implication: If successfully miniaturized, the approach could support wearable‑based coaching in sports training and postoperative rehabilitation, offering objective biomechanical feedback. However, dependence on multimodal sensing may hinder adoption in low‑resource settings unless the Video2EMG alternative proves robust enough to replace hardware.
Executive Summary: MyoMechanix provides a large‑scale multimodal benchmark and a compositional reasoning framework for action quality assessment. It demonstrates improved interpretability and performance on AQA and VideoQA tasks, with preliminary success in predicting EMG from video.
Innovation: 8/10 | Applicability: 6/10 | Commercial Viability: 5/10
Boot-and-Feedback Framework for Generalist-Expert Model Collaboration in Breast Ultrasound Diagnosis
Brief: The Boot-and-Feedback (BooF) framework pairs a multimodal large language model with a vision‑expert network for breast ultrasound diagnosis. In the Boot stage, the language model is constrained by BI‑RADS terminology and preliminary expert predictions to reduce hallucinations; in the Feedback stage, its textual output is fused with image features via an attention‑gated cross‑modality module to refine the expert’s decision. Experiments on several public BUS datasets show improved accuracy and interpretability over prior methods.
Methodological Integrity: The study relies on publicly available breast ultrasound datasets that may overlap with those used for baseline comparisons, raising concerns about potential data leakage. Validation is limited to internal cross‑validation without external multi‑center testing, and the expert model’s performance is assumed rather than independently measured, which could bias the reported gains.
Strategic Implication: If successfully integrated, BooF could reduce operator dependence and improve diagnostic consistency in breast ultrasound, offering a decision‑support tool for radiology clinics. However, its current form requires a separate expert model and language model, complicating deployment and limiting immediate impact to well‑resourced settings.
Executive Summary: BooF demonstrates that guiding a multimodal LLM with domain lexicons and expert priors can mitigate hallucinations and enhance diagnostic performance in breast ultrasound. The approach yields measurable gains in accuracy and interpretability on benchmark datasets.
Innovation: 8/10 | Applicability: 6/10 | Commercial Viability: 5/10
PubMed Gems
Electrocardiogram-Based Deep Learning to Prioritize Testing for Transthyretin Amyloid Cardiomyopathy.
Brief: The study developed a deep learning model that analyzes routine 12-lead ECG images to flag patients for further transthyretin amyloid cardiomyopathy (ATTR-CM) evaluation. It was trained on over 28k ECGs from Yale New Haven Health and validated temporally and across five external multinational cohorts, showing AUROCs between 0.78 and 0.89. Sequential use with an AI‑enabled echocardiography step improved positive predictive value from 0.24 to 0.66 while modestly reducing sensitivity.
Methodological Integrity: Temporal splitting reduces leakage risk, but the development set had a low ATTR‑CM prevalence (~1%), raising concerns about class imbalance and threshold generalizability. External validation across diverse cohorts mitigates some bias, yet the absence of prospective real‑world testing leaves calibration and clinical impact unproven.
Strategic Implication: If prospectively validated, the tool could serve as a low‑cost, first‑line triage to reduce missed ATTR‑CM diagnoses in settings with limited imaging access, aligning with preventive cardiology goals. However, its reliance on a discrete ECG acquisition and clinician‑reviewed output limits ambient, proactive deployment envisioned in the 2026 AI Alpha Manifesto.
Executive Summary: The AI‑ECG model demonstrates solid discriminative performance for ATTR‑CM across internal and external retrospective cohorts. Prospective studies are needed to confirm utility in workflow integration and patient outcomes.
Innovation: 5/10 | Applicability: 8/10 | Commercial Viability: 7/10
AI Clinical Trials (ClinicalTrials.gov)
Artificial Intelligence Support for Stroke
Brief: The study tests an AI decision‑support tool that assists physicians in determining eligibility for intravenous thrombolysis in acute ischemic stroke patients presenting within 24 hours. Using cluster randomization by attending physician, the intervention adds the AI recommendation to standard assessment while the control follows existing guidelines. The primary outcome is the proportion of patients receiving thrombolysis, with blinded endpoint adjudication.
Methodological Integrity: Being a single‑center, open‑label trial introduces potential performance bias, though endpoint blinding mitigates detection bias. Cluster randomization at the physician level may lead to contamination and requires appropriate statistical adjustment for intra‑cluster correlation, which is not detailed.
Strategic Implication: If effective, the tool could improve thrombolysis rates in underserved primary hospitals, reducing stroke‑related disability and costs. However, its reliance on explicit physician interaction limits alignment with ambient, proactive AI paradigms that drive higher workflow integration and ROI.
Executive Summary: The trial evaluates whether AI‑assisted decision support increases thrombolysis use in acute stroke care at a Chinese primary hospital. It employs a cluster‑randomized, open‑label design with blinded outcome assessment.
Innovation: 5/10 | Applicability: 7/10 | Commercial Viability: 6/10