Research map / Probabilistic and scientific modeling / Variational and generative sampling
Mutual Information Constraints for Monte-Carlo Objectives to Prevent Posterior Collapse Especially in Language Modelling
Gábor Melis, András György, Phil Blunsom
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 |
|---|---|---|---|
| use 16 samples instead of 1 with KL objective | DReG · PTB test, hs=128 | perplexity: 92.5 -> 91.7 | Improved |
| use Renyi objective instead of KL with 16 samples (ablation) | DReG · PTB test, hs=128 | perplexity: 91.7 -> 91.6 | Improved |
| use 16 samples instead of 1 | VIMCO · PTB test, hs=128 | perplexity: 94.6 -> 92.1 | Improved |
| use Renyi objective instead of KL with 16 samples (ablation) | VIMCO · PTB test, hs=128 | perplexity: 92.1 -> 92.5 | Hurt |
| use 16 samples instead of 1 with KL objective | DReG · PTB test, hs=256 | perplexity: 83.9 -> 82.8 | Improved |
| use Renyi objective instead of KL with 16 samples (ablation) | DReG · PTB test, hs=256 | perplexity: 82.8 -> 82.2 | Improved |
| use 16 samples instead of 1 | VIMCO · PTB test, hs=256 | perplexity: 84.7 -> 83.4 | Improved |
| use Renyi objective instead of KL with 16 samples (ablation) | VIMCO · PTB test, hs=256 | perplexity: 83.4 -> 83.1 | Improved |
| use multi-sample MI-constrained objectives to raise latent usage | DReG, hs=128 · PTB test | per-token KL: 0.047 -> 0.131 | Improved |
| use multi-sample objectives to avoid collapse with discrete latents | VIMCO, hs=128 · PTB test | per-token KL: 0.000 -> 0.164 | Improved |
| use same dropout mask for all latent samples of a sentence (ablation) | DReG/VIMCO multi-sample · PTB validation | NLL vs latent usage Pareto curve | Improved |
| replace dropout with L2 regularization (ablation) | DReG/VIMCO multi-sample · PTB validation | best NLL | Hurt |
Closest papers by meaning
- PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runs (ICLR 2025 · similarity 0.86)
- XTC: Head-Aware Sampling by Excluding Top Choices (NeurIPS 2026 · similarity 0.86)
- Better Estimation of the Kullback--Leibler Divergence Between Language Models (NeurIPS 2025 · similarity 0.86)
- Beyond Vanilla Variational Autoencoders: Detecting Posterior Collapse in Conditional and Hierarchical Variational Autoencoders (ICLR 2024 · similarity 0.86)
- A Probabilistic Perspective on Unlearning and Alignment for Large Language Models (ICLR 2025 · similarity 0.86)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.86)
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.
