Research map / Probabilistic and scientific modeling / Variational and generative sampling
Latent Template Induction with Gumbel-CRFs
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 the straight-through Gumbel-CRF estimator instead of REINFORCE-MS-C | Structured text modeling on E2E · E2E text modeling test | NLL: 33.82 -> 33.13 | Helped |
| Add the straight-through hard-forward, relaxed-backward variant (ablation) | Gumbel-CRF · E2E text modeling test | NLL: 35.41 -> 33.13 | Helped |
| Use Gumbel-CRF rather than PM-MRF-ST | Structured text modeling · E2E text modeling test | NLL: 37.03 -> 35.41 | Helped |
| Use Gumbel-CRF rather than REINFORCE | Latent-template data-to-text model · E2E data-to-text | BLEU: 60.41 -> 65.83 | Helped |
| Use Gumbel-CRF rather than REINFORCE | Latent-template paraphrase model · MSCOCO unsupervised paraphrase generation | iB4: 11.20 -> 10.20 | Hurt |
| Use Gumbel-CRF rather than REINFORCE | Structured inference network · E2E text modeling | gradient variance ratio | Helped |
| Use Gumbel-CRF rather than REINFORCE | Training structured latent templates · Training efficiency | Sec. per batch: 1.42 -> 0.48 | Helped |
| Use Gumbel-CRF rather than REINFORCE | Training structured latent templates · Training efficiency | GPU mem: 1.8G -> 1.1G | Helped |
| Use neural latent templates rather than the HSMM template model | Data-to-text generation · E2E data-to-text | BLEU: 55.17 -> 65.83 | Helped |
| Use Gumbel-CRF rather than KV2Seq | Data-to-text generation · E2E data-to-text | BLEU: 74.72 -> 65.83 | Hurt |
| Use the unsupervised latent-template model rather than supervised LBOW | Paraphrase generation · MSCOCO paraphrase generation | B4: 25.27 -> 15.75 | Hurt |
Closest papers by meaning
- Paraphrase Generation with Latent Bag of Words (NeurIPS 2019 · similarity 0.89)
- Variational Template Machine for Data-to-Text Generation (ICLR 2020 · similarity 0.87)
- Toward Controlled Generation of Text (ICML 2017 · similarity 0.87)
- Neural Variational Inference for Text Processing (ICML 2016 · similarity 0.86)
- Compressed and Smooth Latent Space for Text Diffusion Modeling (NeurIPS 2025 · similarity 0.85)
- The Hyperfitting Phenomenon: Sharpening and Stabilizing LLMs for Open-Ended Text Generation (ICLR 2025 · 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.