Research map / Reinforcement learning / Policy gradient and actor-critic
Entropy-preserving reinforcement learning
Aleksei Petrenko, Ben Lipkin, Kevin Chen, Erik Wijmans, Marco Francis Cusumano-Towner, Raja Giryes, Philipp Kraehenbuehl
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add adaptive entropy control (REPO-R) on top of GRPO | GRPO · AppWorld and AIME | test accuracy | Improved |
| add adaptive asymmetric clipping (ADAPO) on top of DAPO | DAPO · AppWorld and AIME | test accuracy | Improved |
| use FP16 instead of BF16 training with log-prob rounding fix (ablation) | DAPO · AppWorld (Qwen-3-8B) | entropy dynamics | Improved |
| switch to FP16 training for strictly on-policy RLOO (ablation) | RLOO · AppWorld (Qwen-3-32B) | task goal completion: 79% Test Normal, 71% Test Challenge | Improved |
| use PPO-style off-policy multi-epoch updates (GRPO) instead of strictly on-policy RLOO | RLOO · AppWorld/AIME | entropy retention: GRPO reduces entropy by nearly 90% | Hurt |
| use asymmetric clipping (DAPO) and sequence-level clipping (GSPO) | GRPO/LOOP · AppWorld/AIME | entropy retention | Improved |
| train on entropy-preserving algorithms before sequential training on a new environment | GRPO · Sequential AIME to AppWorld and back | test accuracy (TGC) | Improved |
Closest papers by meaning
- SCOPE-RL: Stable and Quantitative Control of Policy Entropy in RL Post-Training (NeurIPS 2026 · similarity 0.90)
- Convergence Theorems for Entropy-Regularized and Distributional Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Fine-Tuning Discrete Diffusion Models with Policy Gradient Methods (NeurIPS 2025 · similarity 0.89)
- Follow the Winners: Conservative Policy Improvement with the Cross-Entropy Method for Critic-Free RFT (NeurIPS 2026 · similarity 0.89)
- Reasoning with Exploration: An Entropy Perspective (AAAI 2026 · similarity 0.89)
- Stabilizing Policy Optimization via Logits Convexity (NeurIPS 2026 · similarity 0.89)
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.
