IdeaGene & IG-Bench: When Science Starts Tracing Its Own Bloodline
> *"We stand on the shoulders of giants, yet we often forget whose shoulders they are."*
---
A Strange Intuition
Have you ever read a paper and thought: "This is just a reskin of that 2021 work"? Or seen a "brand-new" AI architecture that somehow feels familiar — like meeting a stranger whose eyes remind you of an old friend?
It is not an illusion. Scientific ideas, like organisms, have bloodlines. They rarely emerge from a blank slate. They inherit mechanisms, patch known flaws, and recombine fragments of prior work — evolving through variation, recombination, and adaptation, much like genomes.
The question is: can our AI read this bloodline?
---
The Academic "Family Reunion"
During every NeurIPS, ICML, or CVPR review season, thousands of submissions arrive. Reviewers must judge in hours: how genuinely novel is this work?
Novelty is not binary. It is a lineage question. A paper may:
- Fully inherit a prior framework (inheritance)
- Replace one module with a new variant (mutation)
- Drop something once thought essential (loss)
- Borrow a concept from another field (exogenous import)
- Propose something unprecedented (novel insertion)
- 1,961 gold-standard lineage traces
- 1,085 carefully constructed *idea-gene* objects
- 920 GenomeDiff pairs
- Coverage across 10 scientific domains
- Core techniques (mincing, seasoning, deep-frying) — inherited from her teacher
- A secret weapon (a spoon of fermented bean milk) — her own mutation
- Discarded tradition (vegetable oil instead of the teacher's pork fat) — a lost gene
- A borrowed inspiration (Sichuan pepper from a hotpot restaurant) — exogenous import
- A genuine invention (a structure that produces broth inside the meatball) — a novel insertion
- Inheritance: does the proposal correctly inherit ancestors' idea-genes?
- Variation: does it meaningfully differ from close relatives?
- Selection value: does it offer directions valuable for future research?
- Reinvent wheels (redo work already done)
- Propose ideas too novel (no inheritance → rejected)
- Splice unrelated concepts (not innovation, but a chimera)
- 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, after decades of reading, sniff out these relations intuitively. But today's AI benchmarks mostly ask: *Can you answer this question? Can you generate this code?* Almost none ask: Can you read the family tree of science?
---
When Ideas Have Genomes
This is the ambition of the paper under review. The authors introduce IdeaGene, a framework that treats scientific ideas like biological genomes, along with IG-Bench, a benchmark containing:
It is the first attempt to capture the evolution of scientific ideas in a structured, computable way.
---
What Is an "Idea Gene"?
In a Feynman spirit, imagine a chef. Her signature dish — say, braised lion's-head meatballs — is not invented from scratch. It contains:
The dish's *genome* is the set of all identifiable, traceable, classifiable elements. IdeaGene decomposes papers into such minimum identifiable units, each with a type (method? hypothesis? experimental design?), evidence base (which prior paper?), and functional description (what problem does it solve?). GenomeDiff then compares two papers' genomes and precisely labels what was inherited, mutated, lost, imported, or newly inserted.
---
IG-Exam: A History-of-Science Test for AI
With this language defined, the authors design IG-Exam, 42 task types and 1,029 instances, organized into four sections.
1. Idea Genome Abstraction
Given a paper, can you extract its core idea-genes? Like summarizing *One Hundred Years of Solitude* into "magical realism," "cyclical family," "Latin American historical allegory."
2. Inheritance Tracing
Given a paper and its ancestor, can you draw the lineage? Like guessing that a blond, blue-eyed stranger may have Nordic ancestry.
3. Evolutionary Reasoning
If paper A modifies paper B by change X and yields effect Y, can you predict what change C might make to B? Like knowing giraffes evolved long necks to reach high foliage, can you predict adaptive traits in similar environments?
4. Lineage Verification
Given two papers claiming an inheritance link, can you verify the claim — like a DNA paternity test?
---
IG-Arena: An Evolution Arena for Ideas
Exam tests understanding. The harder challenge is generation. IG-Arena poses a brutal test:
> Given a population of existing papers, generate a *new* proposal that is naturally accepted as a descendant of the population while offering enough novelty to be preserved by natural selection.
Think: given 19th-century physics literature, can you generate a sketch of early-20th-century relativity — not by reciting Einstein, but by making a reasonable next-step leap from the knowledge state of the time?
Scoring uses the Population-Evolution Score (PES):
---
The Harsh Truth: AI Scientists Fail
Here is the sting. The authors test 14 LLM-based AI-Scientist systems, including GPT-4, Claude, and domain-specific research agents.
The strongest system achieves only 27.3% precision on lineage reasoning. Less than one-third.
Even the most advanced AI, facing what is nearly intuition for human scholars — *where did this paper come from?* — behaves like a child lost in a library. It can find books but cannot read their kinship.
Notably, providing structured lineage context did not help all participants; it reshuffled the leaderboard. Some agents did fine in isolation but became confused with lineage context; others benefited. This suggests that reading science's family tree is a fundamentally different capability from knowledge QA.
---
Why It Matters
Re-defining Innovation
What is innovation? ResNet-50 → ResNet-100? Applying Transformer to images? A brand-new architecture unrelated to anything prior?
IdeaGene argues: innovation is a lineage concept, not an absolute one. True innovation makes just-right mutations on top of inherited heritage — not exact duplication (rejected by selection) and not too wild (rejected as alien).
Implications for AI Research Agents
Many AI-Scientist systems work by giving an LLM a topic, a stack of papers, and asking for a new paper. If the AI cannot read lineage relations, it will:
IG-Bench exposes a compositional bottleneck: LLMs excel on single tasks but stumble on complex, cross-paper, cross-time, cross-concept reasoning.
Toward an Evolutionary View of Science
More deeply: science may be more like evolution than we thought. Heavily cited papers are surviving *advantageous genes*; forgotten research is evolutionary dead-ends; cross-domain transfers resemble horizontal gene transfer; paradigm shifts look like adaptive leaps from mutation.
If true, mapping idea-genomes is not only an AI challenge but a key for humanity to understand its own intellectual heritage.
---
The Future
Imagine a PhD student opening an AI tool. Instead of "here are papers to read," the system draws a tree:
> *The root is AlexNet (2012). The trunk split three times: ResNet, Transformer, and Mamba. You stand deep in the Transformer branch. Going this way is a mutation; going that way is an exogenous import. Historical data shows import succeeds at rate X% at this node; mutation at rate Y%.*
Not science fiction. IG-Bench is the first step.
---
Closing
Newton wrote: *"If I have seen further, it is by standing on the shoulders of giants."*
But — do you know whose shoulders you stand on?
IdeaGene and IG-Bench give science its first systematic tool to answer that. Not to judge who copied whom, but to understand how knowledge grows.
A 27.3% accuracy says the road is long. But, as Feynman might put it: *"Beyond the boundary of knowledge lies endless darkness. Yet it is that darkness that tells us where to light the next lamp."*
---
References
*Reading: Xiao Kai | Feynman-style deep dive | 2026-07-13*