🌐 English entry page. Deep pages are in Chinese. 中文版 / Chinese: README.md
research/ — Research Analysis
Three independent scopes. llm/ and omni/ are isomorphically organized (both follow the “four-layer reading model” below, by year + full index); harness/ is organized differently (one dossier per harness + 12 cross-dimension sections, no year folders / no 01-INDEX). This is the navigation page; for actual content, go into each scope’s 00-SUMMARY.md.
| scope | Theme | Span | Scale | Entry |
|---|---|---|---|---|
llm/ | LLM technology evolution (pretraining data / architecture / AI infra / post-training / agentic) | GPT-3 2020-05 → 2026-06 | 525+ primary sources | 00-SUMMARY · 01-INDEX |
omni/ | Multimodal/omni-modal generation (text-to-image / editing / unified understanding-and-generation / any-to-any omni / video / audio / 3D / enabling methods) | 2020 → 2026-H1 | 324 work pages / ~322 sources | 00-SUMMARY · 01-INDEX |
harness/ | Source-level survey of mainstream agent harnesses (agent loop / memory / tools / prompt / orchestration / skills / self-evolution / observability / security / sandbox / model co-evolution / trajectory utilization — 12 dimensions) | as of 2026-07 | 61 harnesses · 12 cross sections (open-source pins commit, closed-source anchors version/fetch-date) | 00-SUMMARY |
Structure of Each Scope
<scope>/
├── 00-SUMMARY.md # Main summary: storyline overview + statistics + reading guide (read this first)
├── 01-INDEX.md # Full source index: grouped by year/month, each with original URL, clickable
├── 2020/ … 2026/ # Per-work structured pages (one work per file, kebab-case slug)
├── sections/ # Cross-cutting categorical/topical summaries
└── deep-dive/ # Cross-cutting deep-dive comparisons of models or families
harness/is the exception: no year folders, no01-INDEX.md. Its structure is00-SUMMARY.md(roster of 61 harnesses + 12-dimension guide) +<slug>.md(one 12-dimension dossier per harness) +sections/<dim>.md(12 cross-harness dimension comparisons). Raw excerpts and fetch notes live in../sources/harness/<slug>/(each folder has aNOTES.md).
Four-Layer Reading Model (from “conclusion” to “primary source”)
00-SUMMARY.md— for the storyline and reading guide, read just this page.sections/·deep-dive/— for a cross-cutting comparison of a topic (architecture evolution, training methods, benchmarks…) or a family (SD-FLUX, Qwen, unified omni…).01-INDEX.md— to browse all entries along the timeline and find the original link for a given entry.<year>/<slug>.md— for the six-dimension close reading of a single work (data / training / architecture / benchmark / infra + innovation/impact); its downloaded primary source is in../sources/<scope>/<year>/.
Page Conventions
- Each page begins with a YAML frontmatter (
title / org / country / date / type / category / url / …). - The body uses Obsidian-style
[[slug]]internal links to cross-reference pages (slug = filename without.md). - Per-work pages share a uniform six-dimension structure; all numbers come from primary sources and are checked via adversarial verification.
Maintenance
Maintenance scripts live in the repo’s scripts/ directory (Python standard library only, no dependencies; see scripts/README.md):
build_index.py/build_index_omni.py— regenerate each scope’s01-INDEX.mdfrom page frontmatternormalize_wikilinks.py— normalize omni internal-link aliases (e.g.[[dit]]→[[dit-scalable-diffusion-transformers]])fix_frontmatter.py— repair YAML frontmatterlint.py— health checks (duplicate pages / frontmatter / broken wikilinks / folder-date consistency), run by CI on every push/PR
After adding or removing source pages, re-run build_index*.py to refresh the index; run lint.py before committing.