Fig. 2

Illustration of the scientific discovery process: Scientific research can be formulated as a search for rewards in an abstract knowledge space. By synthesizing existing knowledge—represented by blue disks (human-discovered) and stars (human-machine discovered))—in novel ways, new knowledge (indicated by red stars) can be explored. For specific research, scientists or LLMs need to traverse the hypotheses-experiment-observation loop, where hypotheses are proposed based on existing knowledge (including LLM knowledge, and additional literature provided via RAG methods), observation, and the creativity of LLMs. Then, with aid of external tools such as programming languages, formal validations, and other methodologies, experiments are conducted to test the hypotheses or gather data for further analysis. The experimental results can be observed and described through the observation process, facilitated by domain-specific models and the multi-modality capabilities of language models. All these parts–observation, proposing hypotheses, conducting experiments, and automation–can be assisted by LLMs and LLM-agents, considering the non-trivial implementations of scientific environments in silico.