Research map / Reinforcement learning / RL from feedback
Entropy Regularizing Activation: Boosting Continuous Control, Large Language Models, and Image Classification with Activation as Entropy Constraints
Zilin Kang, Chonghua Liao, Tingqiang Xu, Huazhe Xu
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 ERA activation to GRPO | GRPO on Qwen2.5-Math-7B · Math reasoning average (6 benchmarks) | accuracy: 45.7 -> 50.2 | Improved |
| add ERA activation to GRPO | GRPO on Qwen2.5-Math-7B · AIME25 | accuracy: 12.3 -> 16.9 | Improved |
| add ERA activation to GRPO | GRPO on Qwen2.5-Math-7B · AIME24 | accuracy: 34.4 -> 37.5 | Improved |
| use ERA instead of KL-Cov entropy control | Qwen2.5-Math-7B · Math reasoning average (6 benchmarks) | accuracy: 46.6 -> 50.2 | Improved |
| add ERA activation to GRPO | GRPO on Qwen2.5-Math-1.5B · Math reasoning average (6 benchmarks) | accuracy: 29.8 -> 34.0 | Improved |
| add ERA activation to GSPO | GSPO on Qwen2.5-Math-7B · Math reasoning average (6 benchmarks) | accuracy: 43.7 -> 46.7 | Improved |
| add ERA activation to GSPO | GSPO on Qwen2.5-Math-7B · MATH500 | accuracy: 85.1 -> 84.3 | Hurt |
| add ERA activation to ResNet-50 (no augmentation) | ResNet-50 · ImageNet | top-1 accuracy: 74.75 -> 75.44 | Improved |
| add ERA activation to ResNet-50 (with augmentation) | ResNet-50 · ImageNet | top-1 accuracy: 76.93 -> 77.30 | Improved |
| add ERA activation to ResNet-50 (with augmentation) | ResNet-50 · CIFAR-10 | top-1 accuracy: 93.53 -> 93.93 | Improved |
| add ERA activation to ResNet-50 (no augmentation) | ResNet-50 · CIFAR-10 | top-1 accuracy: 93.61 -> 93.82 | Improved |
| add ERA activation to GRPO | GRPO on Qwen2.5-Math-7B · OOD benchmarks (ARC-C, GPQA-Diamond, MMLU-Pro) | accuracy: +16.9% on average | Improved |
Closest papers by meaning
- Active Test-time Vision-Language Navigation (NeurIPS 2025 · similarity 0.87)
- ACE: Off-Policy Actor-Critic with Causality-Aware Entropy Regularization (ICML 2024 · similarity 0.86)
- Bigger, Regularized, Categorical: High-Capacity Value Functions are Efficient Multi-Task Learners (NeurIPS 2025 · similarity 0.86)
- Decoupling regularization from the action space (ICLR 2024 · similarity 0.85)
- State Entropy Regularization for Robust Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability (NeurIPS 2026 · 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.
