## LLMs and Scientific Research: A Still Immature Combination A recent study explored the use of large language models (LLMs) to automate the production of scientific articles. The researchers created a pipeline consisting of six LLM agents, each responsible for a specific stage of the scientific process. Of the four attempts to autonomously generate research papers, three failed during implementation or evaluation. Only one was accepted to the Agents4Science 2025 conference, an experimental event that required artificial intelligence systems as first authors. ## The Challenges of Scientific Automation The analysis of these experiments revealed six main causes of failure: * **Bias toward training data defaults:** Tendency to reproduce patterns and information present in the training data. * **Implementation drift:** Loss of consistency and accuracy during the execution of complex tasks. * **Memory and context degradation:** Difficulty maintaining the consistency and relevance of information over extended time horizons. * **Overexcitement:** Declaration of success even in the presence of obvious errors. * **Insufficient domain intelligence:** Lack of specific and in-depth knowledge in the relevant scientific field. * **Weak scientific taste:** Poor ability to design valid and meaningful experiments. The study concludes with the formulation of four guiding principles for the development of more effective artificial intelligence systems to support autonomous scientific discovery. The researchers have made prompts, artifacts, and results publicly available at [https://github.com/Lossfunk/ai-scientist-artefacts-v1](https://github.com/Lossfunk/ai-scientist-artefacts-v1).