Research map / Reinforcement learning / Policy gradient and actor-critic
An Actor-Critic Algorithm for Sequence Prediction
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 actor-critic plus log-likelihood instead of log-likelihood training | Spelling correction · Spelling correction, four L and η settings | CER: 17.81 -> 16.65; 18.4 -> 17.1; 38.12 -> 34.6; 40.87 -> 36.36 | Helped |
| Use actor-critic training | IWSLT 2014 convolutional encoder · Greedy decoding | BLEU: 19.33 -> 21.66 | Helped |
| Use actor-critic training | IWSLT 2014 convolutional encoder · Beam decoding | BLEU: 21.46 -> 22.45 | Helped |
| Use actor-critic plus log-likelihood instead of log-likelihood training | IWSLT 2014 bidirectional GRU encoder · Greedy decoding | BLEU: +1.7 BLEU | Helped |
| Use actor-critic plus log-likelihood instead of log-likelihood training | IWSLT 2014 bidirectional GRU encoder · Beam decoding | BLEU: +1.0 BLEU | Helped |
| Use actor-critic plus log-likelihood instead of log-likelihood training | WMT14 English-French · Greedy decoding | BLEU: +1.5 BLEU | Helped |
| Use actor-critic plus log-likelihood instead of log-likelihood training | WMT14 English-French · Beam decoding | BLEU: +0.4 BLEU | Helped |
| Add log-likelihood gradient to actor-critic training | Spelling correction AC · Spelling correction, four L and η settings | CER: 17.24 -> 16.65; 17.31 -> 17.1; 35.89 -> 34.6; 37.0 -> 36.36 | Helped |
| Increase critic update speed γφ to 0.01 (ablation) | Actor-critic · IWSLT 2014 ablation, greedy, first 10 epochs | valid BLEU: 23.16 -> 23.03 | Hurt |
| Increase critic update speed γφ to 0.1 (ablation) | Actor-critic · IWSLT 2014 ablation, greedy, first 10 epochs | valid BLEU: 23.16 -> 22.80 | Hurt |
| Increase critic update speed γφ to 1 (ablation) | Actor-critic · IWSLT 2014 ablation, greedy, first 10 epochs | valid BLEU: 23.16 -> 8.14 | Hurt |
| Increase actor update speed γθ to 1 (ablation) | Actor-critic · IWSLT 2014 ablation, greedy, first 10 epochs | valid BLEU: 23.16 -> 22.88 | Hurt |
Closest papers by meaning
- Natural Language Actor-Critic: Scalable Off-Policy Learning in Language Space (NeurIPS 2026 · similarity 0.89)
- MA-RLHF: Reinforcement Learning from Human Feedback with Macro Actions (ICLR 2025 · similarity 0.86)
- Enhancing Decision-Making of Large Language Models via Actor-Critic (ICML 2025 · similarity 0.86)
- ESRL: Efficient Sampling-Based Reinforcement Learning for Sequence Generation (AAAI 2024 · similarity 0.86)
- Sequence to Sequence Reward Modeling: Improving RLHF by Language Feedback (AAAI 2025 · similarity 0.86)
- Teaching Language Models to Critique via Reinforcement Learning (ICML 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.