Research map / Reinforcement learning / Policy gradient and actor-critic
Cold-Start Reinforcement Learning with Softmax Policy Gradient
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 |
|---|---|---|---|
| replace MLE with SPG (p_drop = 0.4) | headline generation · Gigaword-10K | ROUGE-L F1: 35.2 $\pm$ 0.3 -> 37.8 $\pm$ 0.2 | Helped |
| replace RAML with SPG (p_drop = 0.8) | headline generation · DUC-2004 | ROUGE-L F1: 23.1 $\pm$ 0.6 -> 24.6 $\pm$ 0.5 | Helped |
| replace RAML with SPG (p_drop = 0.6) | image caption generation · MSCOCO C40 | CIDEr: 0.97 -> 1.01 | Helped |
| set p_drop to 0.4 rather than 0.2 (ablation) | SPG headline generation · Gigaword-10K | ROUGE-L F1: 36.6 $\pm$ 0.2 -> 37.8 $\pm$ 0.2 | Helped |
| vary p_drop across 0.4–0.8 (ablation) | SPG headline generation · DUC-2004 | ROUGE-L F1: 24.3 $\pm$ 0.5; 24.1 $\pm$ 0.5; 24.6 $\pm$ 0.5 | No clear effect |
| set p_drop to 0.6 rather than 0.2 (ablation) | SPG image caption generation · MSCOCO Validation-4K | CIDEr: 1.001 -> 1.033 | Helped |
| set p_drop to 0.8 rather than 0.6 (ablation) | SPG image caption generation · MSCOCO Validation-4K | ROUGE-L: 38.2 $\pm$ 0.1 -> 37.7 $\pm$ 0.1 | Hurt |
| add the DUP reward to penalize consecutive repeated tokens (ablation) | SPG headline generation · Gigaword | ROUGE-L: +0.1 points | Helped |
| use naive policy gradient without variance reduction or warm start | policy-gradient training · sequence generation | training success | Hurt |
| apply naive policy-gradient refinement to an MLE checkpoint | pre-trained MLE model · sequence generation | task performance | No clear effect |
Closest papers by meaning
- Prioritized Soft Q-Decomposition for Lexicographic Reinforcement Learning (ICLR 2024 · similarity 0.86)
- Oracle-Efficient Reinforcement Learning for Max Value Ensembles (NeurIPS 2024 · similarity 0.85)
- Zeroth-Order Policy Gradient for Reinforcement Learning from Human Feedback without Reward Inference (ICLR 2025 · similarity 0.85)
- VESPO: Variational Sequence-level Soft Policy Optimization for Off-Policy LLM Training (NeurIPS 2026 · similarity 0.85)
- Towards General-Purpose Model-Free Reinforcement Learning (ICLR 2025 · similarity 0.85)
- ReMax: A Simple, Effective, and Efficient Reinforcement Learning Method for Aligning Large Language Models (ICML 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.