arXiv Submission: Graph-Augmented Recursive Language Models for Personal Knowledge Systems

arXiv Submission: Graph-Augmented Recursive Language Models for Personal Knowledge Systems

Submission Metadata

Abstract

We present Cog-RLM, a graph-augmented recursive language model architecture for personal knowledge systems that achieves 90.3

Submission Files

Compilation Instructions

# Standard pdflatex compilation (3 passes for cross-references)
pdflatex main.tex
pdflatex main.tex
pdflatex main.tex

No bibtex needed — bibliography is inline using `

References

`.

No external figures — all 6 figures are generated via TikZ/pgfplots inline.

arXiv Submission Steps

1. Go to https://arxiv.org/submit
2. Upload files: `main.tex` and `neurips_2024.sty`
3. Fill metadata:
- Title: Graph-Augmented Recursive Language Models for Personal Knowledge Systems
- Authors: Mohamed Diomande
- Abstract: (copy from above)
- Primary: `cs.CL`
- Cross-list: `cs.AI`
- Comments: `38 pages, 6 figures, 103-question evaluation, 59 references`
- License: CC BY 4.0
4. Preview the compiled PDF on arXiv
5. Submit

Package Dependencies (all standard TeX Live)

The paper uses only standard LaTeX packages available in arXiv's TeX Live distribution:
- `neurips_2024` (included as .sty)
- `inputenc`, `fontenc`, `hyperref`, `url`
- `booktabs`, `amsfonts`, `amsmath`, `amssymb`
- `nicefrac`, `microtype`, `graphicx`, `xcolor`
- `algorithm`, `algorithmic`
- `tikz`, `pgfplots`, `subcaption`
- `enumitem`, `multirow`, `pifont`, `longtable`

All are in standard TeX Live 2024/2025 and arXiv's compilation environment.

Paper Structure

1. Introduction — Problem, contributions, Cog-RLM overview
2. Related Work — 7 subsections covering RLMs, RAG, KGs, personal AI, SLMs, memory networks, multi-hop reasoning
3. Architecture — Full formal description with 4 algorithms
4. Evaluation — 103-question Eval Cube, ablation study, error analysis
5. Discussion — Key findings, limitations, future work
6. Reproducibility — Hardware/software requirements, setup guide
7. Conclusion
8. Appendix A — Sample questions and responses (15 examples)
9. Appendix B — System prompts (actual prompts from twin_server_v3.py)
10. Appendix C — Knowledge graph schema (complete 25-node inventory)
11. Appendix D — Reproduction guide
12. Appendix E — Full 103 Eval Cube questions (NEW for arXiv)
13. Appendix F — Expanded 174-question eval overview (NEW for arXiv)

Notes

• The paper uses NeurIPS format with `[preprint]` option (no submission numbers, line numbers, etc.)
• All figures are TikZ-generated — no external image files needed
• The bibliography contains 59 entries, all inline in the .tex file
• This is a single-author paper (Mohamed Diomande, Independent Research)