Learning the Dynamics of Relational Worlds by Observation
V-JEPA 2 · V-JEPA · I-JEPA · Graph-JEPA (Skenderi, static) · LeJEPA
Graph-JEPA is a method for self-supervised learning of relational temporal data. At a high level, Graph-JEPA predicts the representation of one entity at one timestep, a node-at-time (v, τ), from the representations of the rest of the temporal graph context.
Notably, this approach learns relational dynamics:
- without relying on hand-crafted graph augmentations or pre-specified invariances, which tend to be biased for particular downstream tasks;
- and without having the model reconstruct edge-level details (e.g., bilateral flow magnitudes in a single year), which are dominated by reporting noise rather than signal.
Action-free pretraining yields a substrate evaluated on two probe families; planning is not addressed.
Across 8 datasets in 7 domains, two empirical signatures of learned dynamics emerge and dissociate cleanly. Compression to ~8 effective dimensions appears across three distinct domains; a graph-aware encoder beats a capacity-matched non-graph ablation on next-state prediction across three relation types. The two findings appear and disappear independently across the matrix.
8-dataset matrix.
| dataset | domain | eff_rank | d≈8? | win rate | per-seed p |
|---|---|---|---|---|---|
| BACI Gravity | bilateral commodity trade | 7.03 | ✓ | 95.8% | <10⁻³⁰ all seeds |
| TGBN-Trade | country-pair trade | 7.97 | ✓ | 77.3% | <10⁻²⁶ all seeds |
| ICIO | sectoral input-output | 2.54 | ✗ | 98.5% | <10⁻⁷ all seeds |
| JODIE Reddit u-u | social co-interaction | 11.44 | ✗ | 71.7% | 10⁻⁶⁵ to 2×10⁻³ |
| JODIE Wiki u-u | wiki editing co-interaction | 3.44 | ✗ | 52.6% | 7×10⁻⁵⁹ to 0.95 |
| DBLP | academic coauthorship | 8.31 | ✓ | 38% mean | bimodal |
| Enron | corporate emails | 8.83 | ✓ | 44% | losses w/ tight CI |
| METR-LA | highway traffic | 18.28 (seq=7.07) | ✗ | 8% | p=1 all seeds |
Per-dataset prediction advantage (Δcos = graph − non-graph; 5 seeds, paired Wilcoxon, Bonferroni-corrected over the 8-dataset matrix).
| dataset | Δcos (mean) | 95% CI | win rate | corrected p | Cliff's δ |
|---|---|---|---|---|---|
| BACI Gravity | +0.198 | [+0.177, +0.224] | 95.8% | 1.1×10⁻³⁰ | 1.00 |
| TGBN-Trade | +0.084 | [+0.069, +0.103] | 77.3% | 3.8×10⁻²⁶ | 1.00 |
| ICIO | +0.038 | [+0.029, +0.046] | 98.5% | 6.0×10⁻⁸ | 1.00 |
| JODIE Reddit u-u | +0.105 | [+0.076, +0.130] | 71.7% | 8.3×10⁻⁶⁵ | 1.00 |
| JODIE Wiki u-u | +0.058 | [+0.041, +0.074] | 52.6% | 5.6×10⁻⁵⁸ | 1.00 |
| DBLP | −0.016 | [−0.137, +0.120] | 38.3% | 3.2×10⁻¹²⁸ | −0.20 |
| Enron | −0.046 | [−0.084, −0.006] | 38.1% | 0.92 | −0.60 |
| METR-LA | −0.004 | [−0.005, −0.003] | 7.8% | 1.0 | −1.00 |
| dataset | domain | seeds | results |
|---|---|---|---|
| BACI Gravity | bilateral commodity trade | 5 | paper_results/baci_gravity/ |
| TGBN-Trade | country-pair trade | 5 | paper_results/tgbn_trade/ |
| ICIO | sectoral input-output | 5 | paper_results/icio/ |
| JODIE Reddit u-u | social co-interaction | 5 | paper_results/jodie_reddit_uu/ |
| JODIE Wiki u-u | wiki editing co-interaction | 5 | paper_results/jodie_wikipedia_uu/ |
| DBLP | academic coauthorship | 5 | paper_results/dblp/ |
| Enron | corporate emails | 5 | paper_results/enron/ |
| METR-LA | highway traffic | 5 | paper_results/metrla/ |
Per-dataset, per-seed stamped outputs (git_sha + dataset SHA256 +
deterministic flag) live under paper_results/. For the full
reproduction protocol, see REPRODUCE.md.
- Python ≥ 3.11
- PyTorch ≥ 2.3
- torch-geometric ≥ 2.4
Full dependency list in pyproject.toml. Install via:
pip install -e '.[experiments,dev]'See the LICENSE file for details.
@misc{bodnar2026graphjepa,
title = {Graph-JEPA: Learning the Dynamics of Relational Worlds
by Observation},
author = {Bodnar, Sofia},
year = {2026},
note = {Graph-JEPA-2 codebase}
}