Research map / Probabilistic and scientific modeling / Variational and generative sampling
Training and Inference on Any-Order Autoregressive Models the Right Way
Andy Shih, Dorsa Sadigh, Stefano Ermon
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| train with MAC mask distribution instead of ARDM uniform-cardinality masks | ARDM · Text8 | joint bpd: 1.48 -> 1.40 | Improved |
| train with MAC instead of ARDM | ARDM · Text8 | marginal bpd: 1.12 -> 1.09 | Improved |
| train with MAC instead of ARDM | ARDM · ImageNet32 | joint bpd: 3.60 -> 3.58 | Improved |
| train with MAC instead of ARDM | ARDM · ImageNet32 | marginal bpd: 2.10 -> 2.08 | Improved |
| train with MAC instead of ARDM | ARDM · CIFAR10 | joint bpd: 2.86 -> 2.81 | Improved |
| train with MAC instead of ARDM | ARDM · CIFAR10 | marginal bpd: 1.84 -> 1.81 | Improved |
| add deterministic decomposition protocol only (insight A), no edge weighting (ablation) | ARDM · Text8 | mask distribution ablation | Improved |
| add edge-weighting with random decomposition only (insight B) (ablation) | ARDM · Text8 | mask distribution ablation | Improved |
| combine both insights (w-MAC protocol) (ablation) | ARDM · Text8 | mask distribution ablation | Improved |
| add cardinality reweighing heuristic (ablation) | MAC · Text8 | mask distribution ablation | Improved |
| build MAC on ACE instead of ACE original masks | ACE · gas conditional | log-likelihood: 9.64 -> 9.77 | Improved |
| build MAC on ACE instead of ACE original masks | ACE · hepmass conditional | log-likelihood: -3.86 -> -3.05 | 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.
- ACFlow — ACFlow: Flow Models for Arbitrary Conditional Likelihoods (ICML 2020) · better on 5 of 5 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
- Unlocking Any-Order Generation in Pretrained Autoregressive Image Models (NeurIPS 2026 · similarity 0.87)
- Architectural and Inferential Inductive Biases for Exchangeable Sequence Modeling (NeurIPS 2025 · similarity 0.85)
- Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching (ICLR 2025 · similarity 0.84)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (NeurIPS 2025 · similarity 0.84)
- Learning-Order Autoregressive Models with Application to Molecular Graph Generation (ICML 2025 · similarity 0.84)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.83)
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.
