Meta AI

10 gain · 12 problem · 0 high-evidence

1. problemScienceModerate evidenceImpact 79

   Next-action prediction accuracy of the word-based model declined in later out-of-sample periods, indicating temporal drift in the learned workflow patterns.

   Researchers used more than 295 million inpatient audit log actions recorded between 2019 and 2024 to fine-tune Llama-3-8B to predict a clinician's next action in the electronic health record. They compared two encodings, symbolic field-based tokens and word-based natural-language descriptions, against a first-order Markov baseline.

   Oct 15, 2026

   1 source

   • Journal of the American Medical Informatics AssociationPeer-reviewed
2. gainHealthModerate evidenceImpact 69

   Pairing language models with the SCAI semantic knowledge pipeline raised accuracy on 313 USMLE sample questions for Gemini, Llama 3 70B, and DeepSeek-R1-Distill, with statistically significant gains on all three exam steps.

   Researchers tested four language models, Google Gemini, Microsoft Copilot, Meta Llama 3 70B, and DeepSeek-R1-Distill-Llama-70B, on official text-only USMLE sample questions from Steps 1, 2 CK, and 3. Each model was run with and without SCAI, a semantic knowledge system that supplies clinical context to the host model. Three of the four models improved significantly, while Copilot barely changed.

   Oct 17, 2026

   1 source

   • IEEE Journal of Biomedical and Health InformaticsPeer-reviewed
3. gainHealthModerate evidenceImpact 69

   Gemini 2.5 Flash and ChatGPT-4o gave comprehensive, accurate micron filter instructions and correct local harm reduction service information, in a non-judgmental manner.

   Researchers built a scoring rubric to test how accurately four widely used large language models answered harm reduction questions. The questions covered filtering oral medications before injection and finding local harm reduction services. Gemini 2.5 Flash scored 87% and ChatGPT-4o 77%, while DeepSeek R1 Distill Llama 3.3 70B scored 22% and Llama 4 scored 18%.

   Oct 16, 2026

   1 source

   • Drug and Alcohol ReviewPeer-reviewed
4. problemHealthModerate evidenceImpact 69

   DeepSeek R1 Distill and Llama 4 gave little or no filter instructions, some models generated misinformation, and some refused questions because of built-in safety parameters.

   Researchers built a scoring rubric to test how accurately four widely used large language models answered harm reduction questions. The questions covered filtering oral medications before injection and finding local harm reduction services. Gemini 2.5 Flash scored 87% and ChatGPT-4o 77%, while DeepSeek R1 Distill Llama 3.3 70B scored 22% and Llama 4 scored 18%.

   Oct 16, 2026

   1 source

   • Drug and Alcohol ReviewPeer-reviewed
5. problemHealthModerate evidenceImpact 68

   Six LLMs performed inconsistently on pre-analytical microbiology questions, and experienced specialists rated responses lower than trainees did, so the answers need critical appraisal before use.

   Researchers compared six language models on common pre-analytical questions in medical microbiology, using OpenAI's GPT-3.5 and GPT-4, Meta's Llama2-7B and Llama2-13B, and two domain-enhanced versions. Outputs were checked against expert-reviewed reference answers and scored by thirty-three medical professionals.

   Oct 14, 2026

   1 source

   • Clinical Microbiology and InfectionPeer-reviewed
6. gainHealthModerate evidenceImpact 68

   Fine-tuned LLaMA models extracted symptom presence and severity from free-text notes with high F1 scores, enabling the first structured quarterly trajectories for VNS patients.

   Researchers built a two-stage large language model pipeline to turn free-text clinical notes into structured symptom severity data for vagus nerve stimulation patients, part of NIH-funded work U54AT012307. Using 1427 annotated notes from 95 patients, LLaMA models classified symptoms and then graded severity, producing quarterly trajectories for 23 VNS patients compared to a 10-patient NSQIP surgical control.

   Oct 3, 2026

   1 source

   • JAMIA OpenPeer-reviewed
7. gainMedia & ArtsModerate evidenceImpact 68

   Quantitative stylometry using Burrows' Delta can reliably separate LLM-generated short stories from human-authored stories, providing a measurable tool for authenticity and authorship checks.

   Researchers compared human-authored short stories with stories generated by GPT-3.5, GPT-4, and Llama 70b in response to the same prompts, using Burrows' Delta and clustering methods including hierarchical clustering and multidimensional scaling to visualize stylistic relationships.

   Aug 25, 2026

   1 source

   • Humanities and Social Sciences CommunicationsPeer-reviewed
8. problemMedia & ArtsModerate evidenceImpact 68

   LLM-generated creative writing shows higher stylistic uniformity and tight clustering by model, lacking the broader heterogeneity and individual diversity seen in human-authored stories.

   Researchers compared human-authored short stories with stories generated by GPT-3.5, GPT-4, and Llama 70b in response to the same prompts, using Burrows' Delta and clustering methods including hierarchical clustering and multidimensional scaling to visualize stylistic relationships.

   Aug 25, 2026

   1 source

   • Humanities and Social Sciences CommunicationsPeer-reviewed
9. gainHealthModerate evidenceImpact 67

   LLMs such as GPT-4 and Llama can enhance analysis and interpretation of wearable sensor data to advance health monitoring, activity recognition, and personalized medicine.

   As of August 4, 2024, this peer-reviewed survey in Sensors reviewed early trends in using large language models such as GPT-4 and Llama to model vast wearable sensor data for human activity recognition, health monitoring, and behavioral modeling, integrating them with time series and deep learning methods.

   Jul 20, 2026

   1 source

   • SensorsPeer-reviewed
10. problemHealthModerate evidenceImpact 67

    Complexity and volume of wearable sensor data create substantial modeling challenges, with additional barriers of data quality, computational requirements, interpretability, and privacy concerns for LLM deployment.

    As of August 4, 2024, this peer-reviewed survey in Sensors reviewed early trends in using large language models such as GPT-4 and Llama to model vast wearable sensor data for human activity recognition, health monitoring, and behavioral modeling, integrating them with time series and deep learning methods.

    Jul 20, 2026

    1 source

    • SensorsPeer-reviewed