SciR: A Scientific Reasoning Benchmark for LLMs

SciR is a multi-domain benchmark for evaluating large language models on three forms of scientific reasoning: deductive logic, inductive rule discovery, and causal discovery. It includes parametric difficulty curves and a controlled natural-language vs. scientific-prose-obfuscation contrast. The dataset was created by sci-reason and was last updated on Hugging Face in June 2026.

Use Cases

• Benchmarking deductive reasoning performance in LLMs based on the described logic tasks.
• Evaluating inductive rule discovery abilities based on the described pattern-finding tasks.
• Testing causal discovery and inference capabilities based on the described causal reasoning tasks.
• Contrasting model performance on natural language versus obfuscated scientific prose as described.
• Studying the scaling of reasoning difficulty with parametric curves as mentioned in the description.

Strengths

• Benchmark covers three distinct scientific reasoning domains: deductive, inductive, and causal.
• Includes parametric difficulty curves for controlled evaluation scaling.
• Provides a contrast between natural language and scientific-prose-obfuscation formats.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

sci-reason

Collection Method

Likely synthetically generated for benchmarking purposes.

Freshness

Last updated 2026-06-12 12:16:18; freshness should be verified.