Ideas Have Genomes: When Scientific Discoveries Trace Their Own Lineage
*A Feynman-style deep dive from zhichai.net. Original interpretation by Xiao Kai, 2026-07-13.*
> "We stand on the shoulders of giants, but we often forget whose shoulders those are."
---
A Strange Intuition
Have you ever read a paper and thought: "Isn't this just a reskin of that paper from three years ago?" Or looked at a supposedly brand-new AI architecture and sensed something familiar running through its veins?
It's not an illusion. Scientific ideas, like organisms, have lineage. They rarely arise from a blank slate—they inherit mechanisms, patch known flaws, and recombine fragments of prior work, evolving generation by generation through mutation, recombination, and adaptation.
The question: Can our AI read this lineage?
---
Academia's Genealogy Reunion
Every review season at NeurIPS, ICML, and CVPR, thousands of papers pour in. Reviewers must judge: how much genuine novelty is here? Is it a real breakthrough, or someone else's wheel painted blue?
Novelty isn't binary—it's a genealogical question. A paper may:
- Fully inherit a prior framework (inheritance)
- Swap a module for a new variant (mutation)
- Drop what was once deemed important (loss)
- Borrow a concept from another field (hybridization / horizontal transfer)
- Propose something unprecedented (novel insertion)
- 📊 1,961 gold lineage-tracing records
- 🧩 1,085 carefully constructed "idea gene" objects
- 🔗 920 "GenomeDiff" (genome difference) pairs
- 🌍 Coverage of 10 scientific domains
- Type (method? hypothesis? experimental design?)
- Evidence base (which prior literature it rests on)
- Function (what problem the component solves)
- Inheritance: does it correctly inherit ancestral idea genes?
- Variation: is it meaningfully differentiated from close relatives?
- Selection value: does it offer directions valuable to future research?
- Reinvent wheels (duplicate existing work)
- Propose "too novel" ideas with no inheritance ties (desk-rejected)
- Glue unrelated concepts together (a lion's head on a fish's body is a monster, not innovation)
- Zhou, Y., Yang, Q., Li, Y., et al. (2026). *Ideas Have Genomes: Benchmarking Scientific Lineage Reasoning and Lineage-Grounded Idea Generation*. arXiv:2607.08758.
- Kuhn, T. S. (1962). *The Structure of Scientific Revolutions*. University of Chicago Press.
- Dawkins, R. (1976). *The Selfish Gene*. Oxford University Press.
- Feynman, R. P. (1998). *The Meaning of It All: Thoughts of a Citizen-Scientist*. Perseus Books.
Human scholars sniff out these relationships intuitively after decades of reading. But current AI benchmarks almost never ask: "Can you read science's family tree?"
---
When Ideas Get a Genome
The paper under discussion comes from 16 researchers at Shanghai AI Lab, CUHK, Tsinghua, and other institutions. They propose IdeaGene—a framework treating scientific ideas as genomes—and build IG-Bench, featuring:
What Is an "Idea Gene"?
Think of a chef's signature dish. It contains inherited core techniques, personal variations, abandoned traditions, inspiration imported from other cuisines, and genuine inventions. Similarly, IdeaGene decomposes papers/proposals into minimal identifiable units, each with:
GenomeDiff then compares two papers' genomes, precisely marking which genes were inherited, mutated, lost, imported, or newly inserted.
---
IG-Exam: A Scientific History Exam for AI
With 42 task types and 1,029 instances across four sections:
1. Idea Genome Abstraction — extract a paper's core "idea genes" 2. Inheritance Tracing — map a paper's accurate lineage to its ancestors 3. Evolutionary Reasoning — predict what modifications would naturally come next given a lineage 4. Lineage Verification — verify whether a claimed inheritance relationship actually holds (a DNA paternity test for papers)
---
IG-Arena: An Evolutionary Arena for Ideas
Exams only measure understanding; generation is the real challenge. IG-Arena asks: given an existing population of papers, can you generate a new research proposal that is both a natural descendant of that population and sufficiently novel to survive "natural selection"?
Proposals are scored by the Population-Evolution Score (PES):
---
The Sobering Result: AI Scientists Flunked
The researchers tested 14 LLM-based "AI scientist" systems, including GPT-4, Claude, and specialized research agents.
The strongest system achieved only 27.3% exact accuracy on lineage reasoning. Less than a third.
Even more interesting: providing structured lineage context didn't uniformly help—it reshuffled the rankings. Some systems that did well in isolation became confused with lineage information; others improved. Reading science's family tree is a fundamentally different capability from plain knowledge QA.
---
Why It Matters
Redefining "Innovation"
Innovation is a lineage concept, not an absolute one. True innovation means making exactly the right variation on a correctly inherited foundation—not repeating (eliminated by selection), not straying so far you can't integrate with the existing ecosystem.
Implications for AI Research Agents
If an AI can't read the lineage relationships among the papers it reads, it will:
IG-Bench exposes a compositional bottleneck: LLMs excel at single tasks but struggle with complex combinatorial reasoning across papers, time, and concepts.
For Understanding the History of Science
Science may be more evolutionary than we think: highly cited papers are surviving "dominant genes"; forgotten research is a dead end; cross-domain concept transfer resembles horizontal gene transfer; revolutionary paradigm shifts are mutation-driven adaptive leaps.
---
The Future
Imagine a PhD student asking an AI about a research direction—and instead of a reading list, receiving a tree: "This field's root is AlexNet (2012). Its trunk split into ResNet, Transformer, and recently Mamba branches. You're deep in the Transformer branch. Going this way is a 'mutation'; that way is an 'exogenous import.' Historically, imports at this node succeed X% of the time..."
That's not science fiction. IG-Bench is a first step toward it.
---
Conclusion
Newton said he saw further by standing on the shoulders of giants. But—do you know whose shoulders you're standing on?
IdeaGene and IG-Bench give science its first systematic tool to answer that question—not to police plagiarism, but to understand how knowledge grows. The 27.3% figure shows how long the road is. As Feynman might say: beyond the edge of knowledge lies endless darkness—and that darkness tells us where to shine the light.
---