# v0.9.8 — World Models Category Reinforcement (896 → 900)

**Date:** 2026-10-20
**Scope:** Targeted reinforcement of the thinnest category (World Models & Spatial Intelligence, 8 → 12 methods). Triggered by a completeness audit that found this category disproportionately sparse relative to the project's stated emphasis on spatial intelligence as a core direction.

## Summary

- Library: **896 → 900** methods (+4), 23 categories unchanged.
- **World Models & Spatial Intelligence**: 8 → 12 (+4 new admissions).
- 2 missing arXiv IDs backfilled (Holi-Spatial, Spatial-TTT) — confirmed via arXiv API.
- 2 entries flagged `unverified` (APEIRIA, S²AM3D) — arXiv ID not locatable via arXiv title search; pending manual source check.
- GSMem (arXiv:2603.19137) was found already in the library under "Embodied AI & Robotics" and was **NOT** duplicated; its existing classification is preserved.
- All 4 new admissions are arXiv-verified; method names self-stated in title/abstract per R9 protocol.

## Admitted (4)

| Method | arXiv | Venue | Category |
|--------|-------|-------|----------|
| HY-World 2.0 | 2604.14268 | arXiv 2026 | World Models & Spatial Intelligence |
| F3DGS | 2604.01605 | arXiv 2026 | World Models & Spatial Intelligence |
| 4DGS-WAM | 2608.25956 | arXiv 2026 | World Models & Spatial Intelligence |
| Cloth-Splatting | 2501.01715 | arXiv 2025 | World Models & Spatial Intelligence |

## Method Highlights

- **HY-World 2.0** — Multi-modal world model producing navigable 3DGS worlds from text/image/video via four-stage pipeline (panorama generation + trajectory planning + world expansion). Advances HY-World 1.0.
- **F3DGS** — Federated 3DGS for decentralized multi-agent world modeling; shared geometric scaffold from registered LiDAR enables distributed reconstruction without centralized data aggregation.
- **4DGS-WAM** — Object-centric World Action Model on 4DGS; separately models dynamic objects (policy-predicted future states) and static background; bridges 4DGS reconstruction and future action prediction.
- **Cloth-Splatting** — 3D cloth state estimation via prediction-update framework; action-conditioned dynamics model + 3DGS state updates; mesh-Gaussian coupling for differentiable state-to-image map.

## Backfilled arXiv IDs (2)

| Method | Previous | Now | Verification |
|--------|----------|-----|--------------|
| Holi-Spatial | (empty) | 2603.07660 | arXiv API: title "Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence" confirmed |
| Spatial-TTT | (empty) | 2603.12255 | arXiv API: title "Spatial-TTT: Streaming Visual-based Spatial Intelligence with Test-Time Training" confirmed |

## Flagged Unverified (2)

| Method | Issue | Action |
|--------|-------|--------|
| APEIRIA | arXiv title search "ti:APEIRIA" returned no results | `venue` field marked `unverified`; pending manual source check |
| S²AM3D | arXiv ID not found; description suggests point-cloud segmentation (possible mis-categorization) | `venue` field marked `unverified`; pending manual source check |

## Deduplication Note

During candidate selection, **GSMem** [arXiv:2603.19137] ("3D Gaussian Splatting as Persistent Spatial Memory for Zero-Shot Embodied Exploration and Reasoning") was found already present in the library under the **Embodied AI & Robotics** category. Per the single-source-of-truth deduplication guarantee, it was **not** re-added to World Models. Its existing classification is semantically valid (embodied exploration) and was preserved unchanged.

## Carrier sync

Count-caliber files bumped to **900 methods** and version **v0.9.8**:
- `data/methods.json` (900 methods, count + category_count + generated_at updated)
- `data/categories.json` (World Models count 8 → 12, generated_at updated)
- `3dgs-methods-overview.csv` (900 rows) — regenerated
- `docs/methods.html` (900 entries) — regenerated
- `references/3dgs-methods-overview.md` (new "October 20, 2026" section + Holi-Spatial/Spatial-TTT ID backfill + APEIRIA unverified flag)
- `package.json` (version 0.9.8, description 900 methods)
- `CLAUDE.md` (900 methods)
- `README.md` / `README_CN.md` (900 count)

## Audit Trail

This update originated from a user-initiated completeness audit ("我的文件中3dgs相关的论文全面吗？足够新吗"). The audit found the library comprehensive (896 methods, 23 categories, updated to 2026-10-12) but identified World Models & Spatial Intelligence (8 methods) as the thinnest category relative to the project's spatial-intelligence emphasis, plus 4 entries with missing arXiv IDs in that category. This changelog documents the targeted remediation.