Research map / Probabilistic and scientific modeling / Molecules and proteins
Discovering Non-monotonic Autoregressive Orderings with Variational Inference
Xuanlin Li, Brandon Trabucco, Dong Huk Park, Michael Luo, Sheng Shen, Trevor Darrell, Yang Gao
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 |
|---|---|---|---|
| replace fixed generation order with learned VOI order | Transformer-InDIGO on MS-COCO · MS-COCO | CIDEr: 95.6 -> 100.6 | Improved |
| replace fixed generation order with learned VOI order | Transformer-InDIGO on Django · Django | BLEU: 42.6 -> 45.9 | Improved |
| replace fixed generation order with learned VOI order | Transformer-InDIGO on Gigaword · Gigaword | R-1: 35.6 -> 36.6 | Improved |
| replace L2R order with learned VOI order | Transformer-InDIGO on WMT16 Ro-En · WMT16 Ro-En | BLEU: 32.7 -> 32.9 | Improved |
| use VOI instead of the strongest reported baseline | Transformer-InDIGO on WMT16 Ro-En · WMT16 Ro-En | TER: 49.0 -> 49.3 | Hurt |
| use Gumbel-Matching instead of Plackett-Luce (ablation) | VOI encoder order recovery · MS-COCO order recovery | Normalized Levenshtein Distance: 0.62 -> 0.25 | Improved |
| use Transformer-XL relative positional encoding instead of sinusoid encoding (ablation) | VOI encoder with Gumbel-Matching · MS-COCO order recovery | Normalized Levenshtein Distance: 0.40 -> 0.25 | Improved |
| increase K from 2 to 10 sampled latents (ablation) | VOI encoder order recovery · MS-COCO order recovery | Normalized Levenshtein Distance: 0.31 -> 0.21 | Improved |
| increase K from 10 to 20 sampled latents (ablation) | VOI encoder order recovery · MS-COCO order recovery | Normalized Levenshtein Distance: 0.21 -> 0.21 | No clear effect |
| fine-tune the decoder with the encoder fixed (ablation) | VOI on MS-COCO · MS-COCO | performance | Hurt |
| fine-tune the decoder with the encoder fixed (ablation) | VOI on Gigaword · Gigaword | ROUGE: around 1.5 to 2.0 | Improved |
| fine-tune the decoder with the encoder fixed (ablation) | VOI on WMT16 Ro-En · WMT16 Ro-En | BLEU: around 1.5 points | Improved |
Closest papers by meaning
- Learning-Order Autoregressive Models with Application to Molecular Graph Generation (ICML 2025 · similarity 0.88)
- Order-Independence Without Fine Tuning (NeurIPS 2024 · similarity 0.87)
- Beyond Autoregression: Fast LLMs via Self-Distillation Through Time (ICLR 2025 · similarity 0.87)
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models (ICLR 2025 · similarity 0.85)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.85)
- Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions (ICML 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.
