LLM Research Lineage

A multidimensional atlas of the research ideas that compose modern language models

---

Navigate a shared research timeline, inspect individual techniques, and reveal the component lineages composing a complete model.

Open the live atlas

What is LLM Research Lineage?

LLM Research Lineage is an interactive map of the research decisions inside modern large language models.

It is not a model-release timeline or an architecture catalog. The atlas places the complete LLM development pipeline on one shared coordinate system:

data -> tokenization -> attention -> position -> normalization -> FFN/MoE
     -> pretraining objective -> training systems -> post-training
     -> reasoning -> inference and serving

Time runs vertically and pipeline stages run horizontally. Direct technical descent, conceptual influence, set-aside branches, revivals, implementation advances, model releases, and complete model architectures remain visible together.

Selecting a model reveals an important fact that ordinary family trees obscure:

A modern language model does not have one lineage. It recombines many independent research lineages into a single system.

Explore the atlas

Open index.html directly in a browser. The compiled visualization has no runtime dependencies and works through file://.

You can also use any static server:

npx serve .

The interface supports:

• one unified canvas containing all component lineages;
• complete model-architecture highlighting;
• research-pressure filters spanning the whole pipeline;
• density-aware year, quarter, and month resolution;
• direct-descent and influence edge semantics;
• model release context;
• set-aside and revived research branches;
• search, zoom, fit-to-width, and drag-to-pan navigation;
• detailed motivation, impact, tradeoff, and adoption records.

Current coverage

Dimension	Coverage
Pipeline stages	11
Curated techniques and developments	150+
Model projections	71
Timeline	2014-2026
Research-pressure paths	6
Runtime dependencies	0

Coverage is deliberately curated rather than exhaustive. Relationship quality is more important than raw node count.

Contributor-friendly database

The browser does not treat a monolithic JavaScript file as its source of truth. Contributors edit a normalized JSON database:

database/
|-- manifest.json
|-- families.json
|-- components/
|   |-- 01-attention.json
|   |-- 02-positional-encoding.json
|   `-- ...
|-- models/
|   |-- releases.json
|   |-- picks.json
|   |-- fingerprints.json
|   `-- metadata.json

Each component is independently reviewable. Model releases and model-to-technique mappings are maintained separately from the research graph.

The build process compiles this database into dist/atlas-data.js, which is checked in so the atlas still works without a server. index.html loads that generated bundle directly. In short:

contributors edit database/**/*.json
              | npm run build:data
              v
      dist/atlas-data.js
              | loaded by
              v
          index.html

There are intentionally two layers: database/ is human-maintained source data; dist/ is machine-generated frontend input. Never edit dist/atlas-data.js manually.

Development

Requirements

• Node.js 18 or newer
• Git
• No package installation is required

Build the browser data

npm run build:data

Validate everything

npm test

The validation pipeline checks:

• canonical JSON parsing;
• required component and node fields;
• component-local duplicate IDs;
• lifecycle values and lane bounds;
• dangling research relationships;
• model-to-technique references;
• application JavaScript syntax;
• duplicate and missing DOM targets.

GitHub Actions runs the same validation for every push to main and every pull request.

Repository structure

.
|-- index.html                    # unified visualization and interactions
|-- database/                     # canonical contributor-facing database
|-- dist/
|   `-- atlas-data.js             # generated browser bundle; do not edit
|-- docs/
|   |-- DATA_MODEL.md             # entities, fields and relationships
|   `-- EDITORIAL_POLICY.md       # evidence and curation standards
|-- scripts/
|   |-- build-data.js             # database -> browser bundle
|   `-- validate-data.js          # integrity and syntax checks
|-- CONTRIBUTING.md
|-- CODE_OF_CONDUCT.md
|-- LICENSE                       # MIT license for software
|-- LICENSE-DATA.md               # CC BY 4.0 terms for curated data
`-- package.json

Contributing

Corrections and additions are welcome. Start with CONTRIBUTING.md and the editorial policy.

A useful data contribution identifies:

1. the exact component, technique, relationship, or model mapping;
2. the proposed record or correction;
3. a primary technical source;
4. the exact section, table, configuration field, or page;
5. whether the evidence is documented, supported, or inferred;
6. relevant scope limitations or tradeoffs.

The project does not accept automatically generated bulk graphs without human review and curation.

Design principles

1. One view, not a dashboard. Complexity is preserved spatially instead of divided into disconnected pages.
2. Models are recombinations. A model combines choices from many research lineages.
3. Time resolution follows research density. Dense periods expand to quarter or month resolution.
4. Influence is not inheritance. These relationships use different visual and data semantics.
5. Unknown remains unknown. Closed-model internals are visibly inferred or omitted.
6. Negative results matter. Set-aside branches and later revivals remain part of the record.
7. Every improvement has a cost. Bottlenecks, gains, and tradeoffs belong together.

Documentation

• Data model
• Editorial and evidence policy

Project status

The atlas is currently in pre-release preparation. The visualization, normalized database, deterministic build system, validation workflow, licenses, and contribution templates are functional.

Before the first public release, the project still needs:

• a final primary-source audit of recent records;
• production URL and social-preview metadata;
• accessibility and multi-browser review.

Licensing

Software in this repository is licensed under the MIT License. The curated records under database/ are licensed separately under CC BY 4.0.

---

Built by LLMs Research as an open reference for understanding how modern language models came to be.