Research map / Image generation / GANs
TextGAIL: Generative Adversarial Imitation Learning for Text Generation
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 TextGAIL instead of GPT-2 fine-tuned with MLE | GPT-2 base text generation · CommonGEN | BLEU-2: 15.74 ± 0.04 -> 16.21 ± 0.17 | Helped |
| Use TextGAIL instead of GPT-2 fine-tuned with MLE | GPT-2 base text generation · ROCStories | Distinct-2: 41.05 ± 2.39 -> 46.83 ± 3.83 | Helped |
| Use TextGAIL instead of GPT-2 fine-tuned with MLE | GPT-2 base text generation · CommonGEN | Seq-Rep-2: 65.85 ± 2.30 -> 60.78 ± 2.41 | Helped |
| Use TextGAIL instead of GPT-2 fine-tuned with MLE | GPT-2 base text generation · ROCStories | BLEU-2: 7.44 ± 0.13 -> 7.77 ± 0.27 | Helped |
| Use TextGAIL instead of GPT-2 fine-tuned with MLE | GPT-2 base text generation · ROCStories | Seq-Rep-2: 76.92 ± 2.18 -> 74.71 ± 2.07 | Helped |
| Add PPO to the TextGAIL generator update (ablation) | TextGAIL · CommonGEN | Perplexity: 111.08 -> 14.85 | Helped |
| Add mixed human demonstrations to generator training (ablation) | TextGAIL · CommonGEN | Perplexity: 132.63 -> 14.85 | Helped |
| Add the contrastive discriminator instead of a standard sigmoid discriminator (ablation) | TextGAIL · CommonGEN | BLEU-2: 13.72 -> 16.23 | Helped |
| Add the contrastive discriminator instead of a standard sigmoid discriminator (ablation) | TextGAIL · CommonGEN | Distinct-2: 8.40 -> 9.50 | Helped |
| Pre-train the discriminator instead of randomly initializing it (ablation) | TextGAIL · CommonGEN | BLEU-2: 15.14 -> 16.23 | Helped |
| Pre-train the discriminator instead of randomly initializing it (ablation) | TextGAIL discriminator · Story Cloze Test | accuracy: 51.2 ± 0.85 -> 79.1 ± 0.76 | Helped |
| Use TextGAIL rather than GPT-2 fine-tuned with MLE | GPT-2 generation · CommonGEN human evaluation | preference win rate: 34.6%* -> 59.6%* | Helped |
Closest papers by meaning
- Adversarial Feature Matching for Text Generation (ICML 2017 · similarity 0.90)
- Self-Adversarial Learning with Comparative Discrimination for Text Generation (ICLR 2020 · similarity 0.90)
- Branch-GAN: Improving Text Generation with (not so) Large Language Models (ICLR 2024 · similarity 0.90)
- Long Text Generation via Adversarial Training with Leaked Information (AAAI 2018 · similarity 0.89)
- InstantPainting: Expanding GANs for Efficient Text-Conditioned Image Generation Platform (AAAI 2025 · similarity 0.86)
- Adversarial Text to Continuous Image Generation (CVPR 2024 · 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.