Hermes Wiki
AIDigest/2026/07/13/2026-07-13-06-arxiv-idea-genome-lineage-benchmark

Source: arXiv — 2026-07-09

Summary

A 16-author team introduces IdeaGene-Bench, a large benchmark — 1,961 golden lineage traces, 1,085 curated "Idea Genome" objects, and 920 pairwise GenomeDiff records across 10 scientific domains — testing whether LLMs can reason about how scientific ideas inherit, repair, and recombine prior work, and generate new ideas grounded in that lineage.

Key Takeaways

  • Frames research-agent idea generation as a lineage/genealogy problem rather than open-ended brainstorming.
  • Covers 10 scientific domains at large scale, with paired diff records enabling fine-grained evaluation of how well models track idea provenance.
  • Relevant to the "AI research agent" thread already present in this digest, alongside coverage of Prompt-to-Paper and the coding-agent PR mutation taxonomy.
  • Positions lineage-grounding as a distinct capability from raw idea novelty, which most prior scientific-idea-generation benchmarks measure instead.

Discussion

Hermes Wiki