SGI-Bench is a benchmark designed to assess Scientific General Intelligence (SGI) in AI systems through the complete inquiry cycle: Deliberation, Conception, Action, and Perception. It consists of over 1,000 expert-curated samples across ten disciplines, reflecting Science’s 125 Big Questions. The framework includes an agentic evaluation approach with multiple metrics, enabling a thorough assessment of AI capabilities in scientific reasoning and research. The benchmark focuses on four key task families. Deliberation involves complex reasoning through multi-hop retrieval and synthesis. Conception emphasizes structured ideation and comparative evaluations. Action covers both dry and wet experiments, including code generation and lab protocol development. Perception incorporates various forms of experimental reasoning. The tasks are based on a robust corpus of expert materials, ensuring high fidelity and relevance. Data cleaning and difficulty filtering maintain task integrity by removing samples easily solved by strong language models. Recent updates highlight the ongoing development of SGI-Bench. The release of the SGI-Bench paper on arXiv and its adaptation for evaluation toolkits like VLMEvalKit and SciEvalKit signal significant progress. The framework also includes a detailed scoring system that enhances reproducibility and reduces bias, making the evaluation process more reliable. The results indicate strong performance from models like Gemini-3-Pro and Claude-Sonnet-4.5, showing varied strengths across the different task categories. For those interested in using SGI-Bench, the repository on GitHub provides step-by-step instructions for setting up the environment and running evaluations. Users can install the necessary dependencies and execute scripts for each task family, ensuring straightforward access to this evaluation tool. The project encourages community involvement through GitHub issues for feedback and support.