Ideas Have Genomes: When AI Scientists Can't Tell Who Inherited What

An Awkward Academic Family Reunion

Imagine attending a family reunion of papers — YOLO, YOLOv2, DETR, Faster R-CNN, CLIP, LLaVA, BERT, Transformer, ViT — all doing "object detection" or "vision-language" work, all seemingly related.

But looking alike doesn't mean being kin. YOLOv2 genuinely inherited the "single-stage detection" gene from YOLO, fixing only local flaws (anchor boxes, batch normalization, multi-scale training). DETR also does object detection, but it replaced Faster R-CNN's CNN region proposals entirely with Transformer set prediction — that's "same village, different clan": competition, not inheritance.

The question: Can an AI scientist at this reunion tell who is whose parent?

In July 2026, a team led by Xue Yang at Shanghai Jiao Tong University, together with Tsinghua, CMU, and other institutions, released IdeaGene-Bench (IG-Bench), applying an elegant evolutionary biology framework to give 14 LLM scientists a "family relationship test." The result was stunning: the strongest system (Claude Code + GPT-5.5) scored only 27.3%.

Surface Similarity vs. Genome Inheritance

The paper is titled *Ideas Have Genomes* — and the analogy is not decorative; it is an operational formalization.

Core insight: scientific progress is not topical proximity but mechanism inheritance. Two papers can share a task without inheriting from each other (Faster R-CNN and YOLO both do detection, but region proposals vs. single-stage regression are competing mechanisms). Two papers can be textually distant yet carry the same core mechanism forward (Transformer moving from text to image patches became ViT — adaptive radiation).

Existing AI evaluations — retrieval quality, factuality, writing fluency, novelty — cannot distinguish these cases. A model can retrieve the right papers yet completely miss which parts are "inherited mechanisms" versus "fixed defects."

Idea Genome: The Hereditary Unit Smaller Than a Paper

The IdeaGene framework's first step decomposes each paper into a set of Idea Genome objects, each with:

• Role type: niche (problem niche), mechanism (inheritable method/design), observation (motivating empirical observation), limitation (defect or bottleneck), delta (fix or design change relative to predecessors), claim (asserted result)
• Content description
• Evidence pointer: to specific paragraphs, figures, or equations in the paper
• Optional constraints

Four key constraints: typed (not unstructured summaries), evidenced (no free-floating inference), minimally self-contained (can be independently inherited/mutated/lost), and lineage-relevant (changing it alters the "valid descendant" judgment).

Like genes in biology: the paper is not the hereditary unit — the idea genome is.

GenomeDiff: A Scientific git diff

Given a predecessor \(p_s\) and successor \(p_t\), GenomeDiff \(\Delta_{s \to t}\) records each genome object's fate:

• Inherited: preserved as-is
• Mutated: locally modified but core retained
• Lost: disappeared
• Novel: newly appearing in the successor
• External: imported from another lineage

Each diff also records the primary transition driver — the genome object whose change best explains the transition.

Six Evolutionary Dynamics

GenomeDiff patterns are classified into six dynamics. The first four are genuine lineage relations; the last two are not:

| Dynamic | Lineage? | Criterion | Classic case | |---------|----------|-----------|--------------| | Mutation | Yes | Driver mechanism inherited or locally modified, niche unchanged | YOLO → YOLOv2: single-stage detection kept; anchors/BN/multi-scale fix local flaws | | Adaptive Radiation | Yes | Driver mechanism retained, migrated to new task/domain | Transformer → ViT: self-attention from text sequences to image patches | | Hybridization | Yes | Successor imports driver objects from two or more lineages | CLIP vision encoder + instruction-tuned LLM → LLaVA | | Speciation | Yes | Same niche, but predecessor's driver mechanism replaced by a new one | Faster R-CNN → DETR: CNN region proposals replaced by Transformer set prediction | | Niche Competition | No | Same niche, no driver inheritance | Faster R-CNN vs. YOLO: competing mechanisms on detection | | Isolation | No | Neither shared niche nor driver inheritance | BERT vs. YOLO |

Ambiguous cases are resolved by a fixed priority rule: hybridization > speciation > niche competition; adaptive radiation > mutation.

The taxonomy is not a complete theory of science history but an auditable evaluation contract — turning the vague question "is this paper a descendant of that one?" into a checkable judgment.

IG-Bench: 1,961 Gold Lineages

• 1,961 gold lineage trajectories across 10 scientific fields (NLP, computer vision, multimodal learning, biology, chemistry, physics, materials, medicine, math)
• 1,085 curated Idea Genome objects
• 920 pairwise GenomeDiff records
• Each trajectory contains 3–7 papers, expanded via citation links, semantic search, and domain curation

Evaluation has two parts:

IG-Exam (closed-book): 42 task types, 1,029 instances, testing four capability tiers:

• T1 Genome abstraction (extract genomes from a single paper)
• T2 Inheritance tracking (track genome fates across papers)
• T3 Evolutionary reasoning (infer dynamics type)
• T4 Lineage verification (judge whether lineage claims are self-consistent)

IG-Arena (open-ended generation): systems write research proposals under three information settings — Question-only, Library, Lineage — scored by a Population-Evolution Score (PES) with three sub-dimensions:

• Heredity: did it inherit the correct parental genome objects?
• Variation: does it vary meaningfully from neighbors?
• Selection: does it offer selection value for future research?

The Report Card for 14 AI Scientists

Systems tested: direct LLMs (GPT-5.5, Claude Opus 4.7, Qwen3.6, Gemini-3.1, Kimi-K2, DeepSeek-V4, GLM-5.1, MiniMax-M2.7), research agents (AI Scientist v2, CoI-Agent), and CLI toolchains (Codex, Claude Code).

Finding 1: Lineage reasoning is a compositional challenge

The best IG-Exam score was only 27.3% (Claude Code + GPT-5.5); the best direct LLM was 23.1% (GPT-5.5).

| Capability tier | Best score | |---|---| | T1 Genome abstraction | 34.4% (Gemini-3.1) | | T2 Inheritance tracking | 37.9% (Claude Code + GPT-5.5) | | T3 Evolutionary reasoning | 25.3% (Claude Code + GPT-5.5) | | T4 Lineage verification | 17.4% (Qwen3.6) |

The compositional burden grows from T1 to T4. T4 requires simultaneously maintaining parent identity, genome compatibility, driver consistency, and evidence validity — multiple constraints must hold at once.

Errors are rarely simple retrieval misses. Models may find the correct parent paper but mislabel the dynamics, or infer the right driver genome but misjudge its fate. Precise scoring exposes exactly the consistency failures that downstream generation needs.

Finding 2: Tool scaffolding helps retrieval, not reasoning

CLI toolchains substantially improved T2 (GPT-5.5: 25.7% → 37.9%) via iterative tool calls for cross-paper retrieval. But gains were marginal on T3 and nearly vanished on T4. Research agents overlapped with direct GPT-5.5 on most dimensions — and even scored worse on generation (PES), suggesting multi-step pipelines can hurt generation coherence.

Tool scaffolding amplifies retrieval-dependent abilities while leaving the compositional bottleneck untouched.

Finding 3: Plausibility > lineage coherence

PES decomposition reveals a deeper issue:

• Variation barely changes across settings (82.7–84.7) — systems don't lack novelty
• Selection barely changes (79.7–82.2) — systems don't lack "sounds useful"
• Heredity is the only discriminating dimension (61.9 → 84.2)

Question-only Mutation mode: Heredity 61.9, PES 69.2. Lineage-setting Hybridization: Heredity 84.2, PES 83.6.

The gain from Question to Lineage is not "more innovative" but "better grounded." Systems naturally excel at producing plausible-sounding things but fail to preserve correct parent mechanisms or defect-fix relations.

What This Means for AI Research

1. Retrieval ≠ understanding

Retrieval-augmented auto-research has a structural blind spot: retrieval finds co-occurring papers, not inherited mechanisms. Without genome-level alignment, systems cannot distinguish niche competition from adaptive radiation.

2. Compositional consistency is the real bottleneck

T4's 17.4% is not about intelligence but about maintaining multiple constraints simultaneously — consistent with prior "Knowing-Using Gap" findings. Auto-research systems need compositional verification modules, not just better retrieval.

3. The gap between "sounds right" and "is right"

Systems overproduce plausibility rather than lineage coherence. An idea can be high-variation and low-heredity — sounds novel, inherits nothing. In peer review, that's fatal: "this method is disconnected from related work."

4. Evolutionary biology is a rich mine for AI evaluation

When Darwin wrote "descent with modification" in 1859, he described biological evolution. 167 years later, the same vocabulary evaluates whether AI understands the evolution of scientific ideas. The six dynamics provide auditable criteria, formalizing a once-vague scholarly judgment.

My Take: The Granularity Problem

My biggest question after reading: the granularity of Idea Genome objects.

The paper's criterion — "split if two components can independently disappear or transfer in follow-up work; merge if one is unintelligible without the other" — is a functional-dependency test that is hard to apply. Is multi-head attention one genome, or are "QK matrix multiplication" and "multi-head splitting" two? If follow-up work swaps the similarity metric but keeps multi-head splitting, is that mutation or speciation?

Biology faced the same problem — "gene" evolved from Mendel's hereditary factor to coding sequence to modern regulatory-element-plus-coding-region definitions. Idea Genome likely needs similar refinement.

Still, as a starting point, IG-Bench has done something important: it turned "can AI understand the inheritance structure of scientific ideas?" from philosophical discussion into a quantifiable evaluation. The 27.3% score shows how far we are from a true AI scientist — but at least now we know what to measure.

---

Paper: arxiv.org/abs/2607.08758

Code and data: github.com/VisionXLab/IdeasHaveGenomes

Project page: visionxlab.github.io/IdeasHaveGenomes