What it is
A benchmark for testing whether LLMs can trace how scientific ideas inherit, mutate, and recombine, and generate lineage-coherent proposals. Instead of paper-level relevance or free-form ideation, it evaluates science at idea-object level: papers are decomposed into typed Idea Genomes and aligned with GenomeDiffs to test inheritance and score descendant proposals.
Key results
- Best IG-Exam exact accuracy is 27.3% across 1,029 lineage tasks
- IG-Exam covers 42 task types and 1,029 instances across 4 capability axes
- IG-Bench includes 1,961 lineage traces, 1,085 idea objects, 920 diffs
Shows that fluent literature-aware generation does not imply the ability to track and extend mechanism-level scientific lineage.