Research map / Probabilistic and scientific modeling / Variational and generative sampling
Scaling Structured Inference with Randomization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use randomized top-state selection with importance-weighted tail samples instead of deterministic TopK truncation | Structured inference for linear chains and hypertrees · N=10000; dense, intermediate, and long-tail distributions | MSE: Linear-chain log Z: 2.134 -> 0.004 (dense), 2.013 -> 0.003 (intermediate), 1.647 -> 0.003 (long-tail); hypertree log Z: | Helped |
| Scale the structured VAE latent space from N=100 to N=2000 using RDP | Structured VAE on MSCOCO · MSCOCO test | negative log likelihood: 39.59 ± 0.08 -> 39.48 ± 0.14 | Helped |
| Use RDP rather than TopK for the structured VAE inference network | Structured VAE, N=2000 · MSCOCO test | negative log likelihood: 39.84 ± 0.31 -> 39.48 ± 0.14 | Helped |
| Combine local emission weights and global state-norm weights in the proposal | Randomized Forward · Dense, intermediate, and long-tail distributions | MSE: Dense: 0.655 -> 0.028 (Uniform); 2.908 -> 0.028 (Local); 0.453 -> 0.028 (Global); intermediate: 11.05 -> 0.017; 9.171 -> | Helped |
| Replace the combined local + global proposal with a uniform proposal (ablation) | Randomized Forward · Dense, intermediate, and long-tail distributions | MSE: 0.028 -> 0.655 (dense); 0.017 -> 11.05 (intermediate); 0.022 -> 0.160 (long-tail) | Hurt |
| Replace the combined local + global proposal with a local-only proposal (ablation) | Randomized Forward · Dense, intermediate, and long-tail distributions | MSE: 0.028 -> 2.908 (dense); 0.017 -> 9.171 (intermediate); 0.022 -> 0.481 (long-tail) | Hurt |
| Replace the combined local + global proposal with a global-only proposal (ablation) | Randomized Forward · Dense, intermediate, and long-tail distributions | MSE: 0.028 -> 0.453 (dense); 0.017 -> 0.096 (intermediate); 0.022 -> 0.100 (long-tail) | Hurt |
| Increase the RDP computation budget from 1%N to 20%N (ablation) | Randomized inference · N=10000; linear-chain log Z, hypertree log Z, and linear-chain entropy | MSE: Linear-chain log Z: 0.078 -> 0.004, 0.616 -> 0.003, 0.734 -> 0.003; hypertree log Z: 3.376 -> 0.148, 5.012 -> 0.246, 7.2 | Helped |
| Use deterministic TopK truncation without tail sampling (ablation) | Structured VAE inference network, N=2000 · MSCOCO test | negative log likelihood: 39.48 ± 0.14 -> 39.84 ± 0.31 | Hurt |
| Use TopK truncation instead of RDP for the structured VAE with N=100 (ablation) | Structured VAE inference network · MSCOCO test | negative log likelihood: 39.59 ± 0.08 -> 39.76 ± 0.11 | Hurt |
Closest papers by meaning
- Differentiable Dynamic Programming for Structured Prediction and Attention (ICML 2018 · similarity 0.89)
- Variational Message Passing with Structured Inference Networks (ICLR 2018 · similarity 0.86)
- d$^2$p: Structured Soft Attention Is All You Need (ICML 2026 · similarity 0.86)
- TEST-TIME SCALING IN DIFFUSION LLMS VIA HIDDEN SEMI-AUTOREGRESSIVE EXPERTS (ICLR 2026 · similarity 0.84)
- Deep Linear Network Training Dynamics from Random Initialization: Data, Width, Depth, and Hyperparameter Transfer (ICML 2025 · similarity 0.84)
- Entropy-Gated Latent Recursion (NeurIPS 2026 · similarity 0.84)
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.