The impact of large language models in science (www.nature.com)

Nature Computational Science’s new Focus issue maps how large language models (LLMs) are rapidly reshaping scientific research—offering new workflows, modeling paradigms and cross-disciplinary tools—while flagging systematic risks around bias, reproducibility, equity and environmental cost. The collection spans perspectives and studies showing LLMs as research partners in chemistry, social science, linguistics, healthcare and the humanities; empirical tests of LLM cognition and replication (73–81% replication of psychological experiments but amplified effect sizes); and warnings about over‑reliance on AI code assistants that can introduce undetected errors into largely untested research software. The issue stresses ethical challenges (e.g., safety filters disproportionately removing LGBTQ+ content), fairness in clinical use, and the urgent need for standards and validation for scientific inference. Several technical advances and implications stand out: tool‑orchestrating agents (SciToolAgent) using knowledge graphs to automate workflows; physics‑based pretraining to align LLMs for polymer property prediction; transformer embeddings for rich life‑trajectory prediction; and multimodal models for single‑cell genotype–phenotype mapping. Hardware and efficiency innovations offer practical scaling paths—neuromorphic engineering is argued to bridge ML and low‑power hardware, while in‑memory gain‑cell accelerators trained a 1.5B‑parameter model with up to 70,000× energy reduction and 100× speedup versus GPUs. The issue also highlights sustainability (GAI‑driven e‑waste), domain‑specific language models for proteins and antibodies, and pandemic forecasting models, underscoring both transformative potential and the need for robust, equitable, and energy‑aware practices.