Saved in:
Bibliographic Details
Main Authors: Kang, Zilin, Liao, Chonghua, Xu, Tingqiang, Xu, Huazhe
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2510.08549
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911201198342144
author Kang, Zilin
Liao, Chonghua
Xu, Tingqiang
Xu, Huazhe
author_facet Kang, Zilin
Liao, Chonghua
Xu, Tingqiang
Xu, Huazhe
contents We propose ERA, a new paradigm that constrains the sampling entropy above given thresholds by applying specially designed activations to the outputs of models. Our approach demonstrates broad effectiveness across different domains: 1) for large language models(LLMs), boosting the AIME 2025 score for Qwen2.5-Math-7B by 37.4%; 2) for continuous control reinforcement learning agents, improving performance by more than 30% over strong baselines such as SAC on the challenging HumanoidBench; 3) for image classification, enhancing ImageNet top-1 accuracy by 0.69% for ResNet-50. These gains are achieved with a computational overhead of less than 7%. Our work validates output activation as a powerful tool for entropy control, opening a new direction for designing simpler and more robust algorithms.
format Preprint
id arxiv_https___arxiv_org_abs_2510_08549
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Entropy Regularizing Activation: Boosting Continuous Control, Large Language Models, and Image Classification with Activation as Entropy Constraints
Kang, Zilin
Liao, Chonghua
Xu, Tingqiang
Xu, Huazhe
Machine Learning
We propose ERA, a new paradigm that constrains the sampling entropy above given thresholds by applying specially designed activations to the outputs of models. Our approach demonstrates broad effectiveness across different domains: 1) for large language models(LLMs), boosting the AIME 2025 score for Qwen2.5-Math-7B by 37.4%; 2) for continuous control reinforcement learning agents, improving performance by more than 30% over strong baselines such as SAC on the challenging HumanoidBench; 3) for image classification, enhancing ImageNet top-1 accuracy by 0.69% for ResNet-50. These gains are achieved with a computational overhead of less than 7%. Our work validates output activation as a powerful tool for entropy control, opening a new direction for designing simpler and more robust algorithms.
title Entropy Regularizing Activation: Boosting Continuous Control, Large Language Models, and Image Classification with Activation as Entropy Constraints
topic Machine Learning
url https://arxiv.org/abs/2510.08549