---
license: mit
tags:
  - icml2026-repro
  - gflownet
  - reproduction
---

# ST-GFN reproduction code (ICML 2026 #18502, OpenReview QQRCqb2PDj)

Independent reimplementation of **Spectral Flow Matching: Stabilizing Stochastic
GFlowNets via Frequency-Domain Regularization** (no author code was released).

| file | purpose |
|---|---|
| `envs.py` | BitSequence, HyperGrid, TicTacToe, SingleCell proxy environments |
| `models.py` | RFF map, GFlowNet backbone, RND/ICM, autocorrelated intrinsic reward |
| `train.py` | ST-GFN + 10 baselines, training loop, metrics |
| `run_suite.py` | parallel experiment runner |
| `analyze.py` | aggregation, 95% CIs, Welch tests, Cohen's d |
| `diagnose_spectral.py` | structural diagnostics of the two spectral terms |
| `hf_job_scaled.py` | scaled experiment for Hugging Face GPU Jobs |

Reproduce locally:

```bash
python run_suite.py --envs hypergrid --iters 5000 --seeds 0,1,2,3,4 --out outputs/main
python analyze.py --out outputs/main
```

Full write-up: see the Trackio logbook linked from the reproduction collection.