Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Streaming Bayes GFlowNets
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SB-GFlowNets, which extend GFlowNets to streaming Bayesian inference over discrete parameter spaces via a streaming balance loss and a KL-based streaming update, with theoretical error bounds.
Problem
Streaming Bayesian inference over discrete, combinatorial parameter spaces is hard because VI over discrete posteriors relies on continuous relaxations and posterior propagation across data chunks is not supported.
Approach
Use a GFlowNet trained on the previous posterior as the prior for the next chunk; the streaming balance condition and a KL streaming criterion update it using only the newly observed data likelihood, without revisiting past data. Includes theoretical bounds on error accumulation across streaming updates.
Key results
- On phylogenetics with 7 leaves, SB-GFlowNet trained in 1279.68 s vs 2846.88 s for a standard GFlowNet, with comparable TV (relative gain 0.00 ± 0.04).
- In set generation, TB outperforms KL for sparse targets (TV 0.36 vs 0.55 at alpha=0.5), while KL is better at alpha=1.0 (0.13 vs 0.21).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KL minimization instead of TB (ablation) | GFlowNet set generation · Set generation, α=1.00 | TV distance: 0.21 ± 0.06 -> 0.13 ± 0.03 | Helped |
| Use KL minimization instead of TB (ablation) | GFlowNet set generation · Set generation, α=0.75 | TV distance: 0.28 ± 0.10 -> 0.17 ± 0.04 | Helped |
| Use KL minimization instead of TB (ablation) | GFlowNet set generation · Set generation, α=0.50 | TV distance: 0.36 ± 0.24 -> 0.55 ± 0.38 | Hurt |
| Use streaming posterior updates instead of training from scratch | GFlowNet phylogenetic inference · Phylogenetic inference, 7 leaves | training time (seconds per 20k epochs): 2846.88 s -> 1279.68 s | Helped |
| Use streaming posterior updates instead of training from scratch | GFlowNet phylogenetic inference · Phylogenetic inference, 9 leaves | training time (seconds per 20k epochs): 3779.11 s -> 1714.49 s | Helped |
| Use streaming posterior updates instead of training from scratch | GFlowNet phylogenetic inference · Phylogenetic inference, 11 leaves | training time (seconds per 20k epochs): 4821.74 s -> 2303.99 s | Helped |
| Use streaming posterior updates instead of training from scratch | GFlowNet phylogenetic inference · Phylogenetic inference, 7 leaves | relative accuracy gain (TV): 0.00 ± 0.04 | No clear effect |
| Use streaming posterior updates instead of training from scratch | GFlowNet phylogenetic inference · Phylogenetic inference, 9 leaves | relative accuracy gain (TV): -0.02 ± 0.04 | No clear effect |
| Use streaming posterior updates instead of training from scratch | GFlowNet phylogenetic inference · Phylogenetic inference, 11 leaves | relative accuracy gain (TV): 0.00 ± 0.01 | No clear effect |
Surprising findings
- Model accuracy decreases with the number of streaming updates in phylogenetic inference, which the authors attribute mainly to posterior sparsity.
- Permutation invariance of the learned distribution is not guaranteed when learning objectives are imperfectly minimized.
Limitations
- Errors accumulate through repeated streaming updates, potentially causing catastrophic forgetting and requiring retraining from checkpoints.
- Expressiveness of GFlowNet parameterizations is unclear, unlike traditional variational families.
- Updating when the size of the generated object changes (e.g., new species in phylogenetics) remains open.
Experimental setup
Datasets: synthetic streaming data from set generation task synthetic linear preference data simulated JC69 DNA sequences synthetic linear Gaussian SEM data
Benchmarks: set generation (synthetic) linear preference learning (synthetic) online Bayesian phylogenetic inference (synthetic) Bayesian structure learning (synthetic, d=5)
Metrics: total variation distance training time (seconds per 20k epochs) predictive NLL MSE probability of true phylogeny
Baselines: GFlowNet trained from scratch TB loss training KL-based GFlowNet training
Builds on: GFlowNets trajectory balance SDA-Bayes DAG-GFlowNet
Closest papers by meaning
- Embarrassingly Parallel GFlowNets (ICML 2024 · similarity 0.88)
- On Divergence Measures for Training GFlowNets (NeurIPS 2024 · similarity 0.87)
- Model-Informed Flows for Bayesian Inference (NeurIPS 2025 · similarity 0.87)
- Generalized Variational Inference via Optimal Transport (AAAI 2024 · similarity 0.85)
- Inflationary Flows: Calibrated Bayesian Inference with Diffusion-Based Models (NeurIPS 2024 · similarity 0.85)
- Variational Inference with Mixtures of Isotropic Gaussians (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.