Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Distribution Transformers: Fast Approximate Bayesian Inference With On-The-Fly Prior Adaptation
George Whittle, Juliusz Ziomek, Jacob Henry Rawling, Michael A Osborne
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add prior loss term to the training objective (ablation) | Distribution Transformer · Inverse-gamma (wide meta-prior), prior KL | expected prior KL-divergence: 0.3020 -> 0.0004 | Improved |
| add prior loss term to the training objective (ablation) | Distribution Transformer · Inverse-gamma (wide meta-prior), posterior KL | expected posterior KL-divergence: 0.0004 -> 0.0003 | Improved |
| increase GMM components from 2 to 5 in the DT | Distribution Transformer · Inverse-gamma (narrow meta-prior) | expected KL-divergence: 0.0044 -> 0.0004 | Improved |
| increase GMM components from 2 to 5 in the DT | Distribution Transformer · Inverse-gamma (wide meta-prior) | expected KL-divergence: 0.0058 -> 0.0003 | Improved |
| test on unseen shifted meta-prior instead of training meta-prior (ablation) | DT-5 · Meta-prior misspecification | expected KL-divergence: 0.0003 -> 0.0054 | Hurt |
| halve GMM components at inference time (10 -> 5) (ablation) | DT trained with 10 components · 8-dim linear Gaussian | expected posterior KL-divergence: 0.0873 -> 0.0985 | Hurt |
| double GMM components at inference time (10 -> 20) (ablation) | DT trained with 10 components · 8-dim linear Gaussian | expected posterior KL-divergence: 0.0873 -> 0.0946 | Hurt |
| use 2 components at inference time instead of 10 (ablation) | DT trained with 10 components · 8-dim linear Gaussian | expected posterior KL-divergence: 0.0873 -> 0.1494 | Hurt |
| replace EKF with Distribution Transformer for filtering | DT · Bayesian sensor fusion | expected NLL: 95.9 -> -0.197 | Improved |
| replace particle filter with Distribution Transformer | DT · Bayesian sensor fusion | iteration time: 0.818 -> 0.017 | Improved |
| use DT-5 rather than a PFN with 5000 Riemann buckets | DT-5 · Inverse-gamma (wide meta-prior) | expected KL-divergence: 0.2935 -> 0.0003 | Improved |
| use DT instead of ACE for GP hyperposterior | DT · GP hyperposterior | expected NLL: 0.35 -> 0.31 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TabPFNv2 — TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second (ICLR 2023) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Can Transformers Learn Full Bayesian Inference in Context? (ICML 2025 · similarity 0.87)
- What does a Bayes-filtered transformer believe? A predictive Monte Carlo approach (NeurIPS 2026 · similarity 0.86)
- Transformers on Markov data: Constant depth suffices (NeurIPS 2024 · similarity 0.86)
- Model-Informed Flows for Bayesian Inference (NeurIPS 2025 · similarity 0.85)
- Generalized Variational Inference via Optimal Transport (AAAI 2024 · similarity 0.85)
- To discretize continually: Mean shift interacting particle systems for Bayesian inference (NeurIPS 2026 · 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.
