🔄 Geometric Memory — Reproduction Report

ICML 2026 Paper #22 · arXiv:2510.26745 · Reproducer: Luis Poveda · 2026-07-16

Coverage

✅ Pass: 31/44 (70%) 🟡 Partial: 8/44 (18%) ❌ Fail: 5/44 (11%)

Category	Pass	Partial	Fail
Setup	3	1	1
Claim 1 — Path-star	0	3	2
Claim 2 — Contradiction	4	0	0
Claim 3 — Geometry	5	0	0
Claim 4 — Spectral bias	5	0	0
Robustness	3	0	1
Publishing	7	2	1
Deliverables	4	2	0
Total	31	8	5

❌ Failed Requirements

1. 1.1 — Full-scale Transformer (50k nodes, d=10000): not run. d=1000 pilot canceled (credits exhausted).
2. 1.2 — Mamba variant: not run. mamba-ssm unavailable locally; HF credits exhausted.
3. 0.2 — Conda environment: used Hermes venv + pip instead.
4. 5.4 — Path length l=8: not tested (credits exhausted).
5. 6.7 — Posterly poster: unavailable. Replaced with matplotlib A0 PDF.

Results by Claim

Claim 1 — Path-star 🟡 Partial

d=10 (N=1111): ~60% accuracy vs paper's ~100%. d=1000 pilot: canceled at epoch 1,231/10,000 — train 100%, test 0%. Mamba: not tested.

Claim 2 — Associative contradiction ✅ Pass

l=2..6 sweep: 0% path accuracy across ALL arms (geometric and frozen). Hardest-token: geometric 100%, frozen 33.33%.

Claim 3 — Embedding geometry ✅ Pass

Heatmaps, UMAP projections, and Node2Vec comparison completed. Unfrozen models develop geometric structure; frozen do not.

Claim 4 — Spectral bias ✅ Pass

Node2Vec d=100: 53.1% top-1 edge recovery, 100% top-5. Procrustes residual 0.4158→0.6570 (no monotonic improvement). Transformer: no spectral alignment (Procrustes 0.0068).

Resources

• 📦 HF Dataset — 7 figures, poster PDF, remediation results, source code
• 📎 Appendix (Vercel) — full report with all 4 claims, figures, analysis scripts
• 📄 Status report (markdown)
• 📋 Original reproduction plan
• 🛠️ Deviation remediation plan
• 🐍 Claim 3 analysis script
• 🐍 Claim 4 analysis script