SGI-Bench: Scientist-Aligned Benchmark for Scientific General Intelligence

More than 1,000 expert-curated samples across 10 scientific disciplines probe the Scientific General Intelligence of LLMs. The SGI-Bench evaluates models across the full inquiry cycle—Deliberation, Conception, Action, and Perception—inspired by Science's 125 Big Questions. It was created by InternScience and last updated in June 2026.

Use Cases

• Benchmarking LLM performance on scientific reasoning tasks based on the described inquiry cycle stages
• Evaluating agentic AI workflows in scientific contexts based on the agentic evaluation framework
• Analyzing model strengths across different scientific disciplines based on the 10 covered domains
• Training or fine-tuning models for scientific problem-solving based on expert-curated samples

Strengths

• More than 1,000 expert-curated samples provide a substantial testbed
• Covers 10 scientific disciplines for broad evaluation
• Structured around the four-stage inquiry cycle (Deliberation, Conception, Action, Perception)

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Data may reflect bias inherent to the expert curation process and source questions

Provenance

Source

InternScience via Hugging Face

Collection Method

Expert-curated samples inspired by Science's 125 Big Questions

Freshness

Last updated 2026-06-02 12:03:49; freshness should be verified