Research map / Reinforcement learning / RL from feedback
Teacher Forcing Recovers Reward Functions 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 |
|---|---|---|---|
| semi-supervised RL with induced reward on non-parallel data | T5-Base · DailyDialog (+OpenSubtitles) test | BLEU2: 8.96 -> 11.02 | Helped |
| semi-supervised RL with induced reward on non-parallel data | T5-Base · DailyDialog (+OpenSubtitles) test | BLEU4: 3.69 -> 4.30 | Helped |
| induced reward instead of metric-based R-Regression reward | T5-Base RL · DailyDialog (+OpenSubtitles) test | BLEU2: 10.34 -> 11.02 | Helped |
| induced reward instead of metric-based R-Regression reward | T5-Base RL · DailyDialog (+OpenSubtitles) test | BLEU4: 4.18 -> 4.30 | Helped |
| add noisy parallel OpenSubtitles data in full supervision | T5-Base · DailyDialog (+OpenSubtitles) test | BLEU2: 8.96 -> 8.75 | Hurt |
| RL-based semi-supervision instead of Self-Training | T5-Base · DailyDialog (+OpenSubtitles) test | BLEU2: 9.10 -> 11.02 | Helped |
| induced reward on non-parallel Quora data | T5-Base · Quora paraphrase generation test | iBLEU: 23.27 -> 23.78 | Helped |
| Self-Training on non-parallel Quora data | T5-Base · Quora paraphrase generation test | Self-BLEU: 44.77 -> 48.02 | Hurt |
| use dense step-wise induced reward instead of sparse reward (ablation) | Induced-R · DailyDialog (+OpenSubtitles) test | BLEU2: 9.75 -> 11.02 | Helped |
| use dense step-wise induced reward instead of sparse reward (ablation) | Induced-R · DailyDialog (+OpenSubtitles) test | BLEU4: 3.99 -> 4.30 | Helped |
| use dense step-wise induced reward instead of sparse reward (ablation) | Induced-R · Quora paraphrase generation test | iBLEU: 23.63 -> 23.78 | Helped |
| synchronizing period k=1 (on-policy) vs k=5000 (ablation) | policy gradient · DailyDialog (+OpenSubtitles) test | BLEU4 | Hurt |
Closest papers by meaning
- Towards Cost-Effective Reward Guided Text Generation (ICML 2025 · similarity 0.90)
- Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback (NeurIPS 2025 · similarity 0.85)
- PoEM: Predicting New RL Outcomes from Existing Policies (NeurIPS 2026 · similarity 0.85)
- Successor Features for Efficient Multi-Subject Controlled Text Generation (ICML 2024 · similarity 0.85)
- Subject-driven Text-to-Image Generation via Preference-based Reinforcement Learning (NeurIPS 2024 · similarity 0.85)
- Text2Reward: Reward Shaping with Language Models for Reinforcement Learning (ICLR 2024 · 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.