ebm-splats 2.0.0

Energy-Based Model with Gaussian Splats on a 640D hypersphere — archived research project

EBM-Splats

Status: Research complete. Phase 1-2 empirical tests finished. Energy-guided generation continued in m2m-energy-fields.

Energy-Based Model with Gaussian Splats on a 640D hypersphere. Explores distributional representations for latent spaces, sampling via Langevin dynamics and Rectified Flow.

Project Phases

Phase 1: EBM + PGLF (April 2026) — Discarded

EBM with Gaussian splats as attractors on S^639 + PGLF (projection over MiniLM with contrastive loss).

Result: PGLF degraded MiniLM on STS-B (-4.7%). Projection over pre-trained embeddings always destroys geometry.

Phase 1 Empirical: Discard Tests (July 2026)

3 empirical tests to discard or confirm alternatives. RTX 3090, real data.

Test	Hypothesis	Result	Verdict
PGLF Grid (14 configs)	Can any config beat MiniLM?	0/14 beat baseline (0.8672)	DISCARDED
OOD Detection	Does EBM energy detect OOD?	AUROC=1.0 but NN=0.999	NO ADVANTAGE
RF vs Langevin	Does RF solve the speed bottleneck?	24-29x faster, better quality	CONFIRMED

Key finding: The argument that "200 Langevin steps per token" is prohibitive no longer applies. Rectified Flow with 1-2 steps produces better samples than Langevin with 200 steps, 24x faster.

Phase 2: Energy-Guided Generation (July 2026)

EBM as a generator that learns its own latent space (not as a layer over pre-trained models), with sampling via Rectified Flow.

Test	Hypothesis	Result	Verdict
Energy-Guided Generation	Can energy manipulation steer generation?	100% topic control at gs=1.0-2.0	CONFIRMED
Concept Composition	Can multiple concepts be combined?	4/4 mechanisms work	CONFIRMED

Composition results:

• Equal blend (A+B): balanced similarity to both topics
• Weighted (70/30): asymmetric control confirmed
• Suppression (A−B): sim_B dropped from 0.44 to −0.29
• Triple (A+B+C): all three topics active (sim > 0.48)

EBM + RF enables semantic arithmetic on the hypersphere.

Repository Structure

├── src/ebm/               # Core EBM modules (geometry, splats, energy, model, etc.)
├── pglf/                  # PGLF (archived — discarded empirically)
├── scripts/               # Training and generation scripts
├── tests/
│   ├── phase1_t11_rf_vs_langevin.py    # RF vs Langevin benchmark
│   ├── phase1_t12_pglf_grid.py         # PGLF grid search
│   ├── phase1_t13_ood_energy.py        # OOD detection test
│   ├── phase2_energy_guided.py         # Energy-guided generation test
│   ├── phase2_composition.py           # Concept composition test
│   └── t*_results.jsonl                # Raw results
├── docs/
│   ├── PHASE1_RESULTS.md  # Full Phase 1 report
│   ├── PHASE2_RESULTS.md  # Phase 2 energy-guided generation report
│   └── ...
└── benchmark_results/     # Previous benchmarks

Detailed Results

See docs/PHASE1_RESULTS.md and docs/PHASE2_RESULTS.md for full empirical test reports.

Tech Stack

• Python, PyTorch (CUDA 12.4, RTX 3090)
• sentence-transformers, HuggingFace datasets
• Rust (M2M integration via HTTP)

License

Apache-2.0