Saved in:
Bibliographic Details
Main Authors: Yang, Ming, Li, Xiaofan, Ma, Zhiyuan, Shi, Dengliang, Du, Jintao, Cheng, Yu, Zheng, Weiguo
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2509.25240
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918150875906048
author Yang, Ming
Li, Xiaofan
Ma, Zhiyuan
Shi, Dengliang
Du, Jintao
Cheng, Yu
Zheng, Weiguo
author_facet Yang, Ming
Li, Xiaofan
Ma, Zhiyuan
Shi, Dengliang
Du, Jintao
Cheng, Yu
Zheng, Weiguo
contents Recent curriculum reinforcement learning for large language models (LLMs) typically rely on difficulty-based annotations for data filtering and ordering. However, such methods suffer from local optimization, where continual training on simple samples in the early steps can cause the policy to lose its exploration. We propose a novel schema, namely Hamiltonian curiosity augmented large language model reinforcement (HAMMER), that transfers diversity metrics, commonly used in dataset evaluation, into the dynamic reinforcement learning procedure, where training samples are ordered via a minimum-semantic Hamiltonian path making the initial training retrain more exploration. From a theoretical perspective of generalization bounds, diversity-driven ordering facilitates stable convergence. Empirical evaluations indicate that HAMMER stimulates model "curiosity" and consistently achieves a 3% to 4% average accuracy gain across diverse inference benchmark.
format Preprint
id arxiv_https___arxiv_org_abs_2509_25240
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle HAMMER: Hamiltonian Curiosity Augmented Large Language Model Reinforcement
Yang, Ming
Li, Xiaofan
Ma, Zhiyuan
Shi, Dengliang
Du, Jintao
Cheng, Yu
Zheng, Weiguo
Machine Learning
Artificial Intelligence
Computation and Language
I.2.7
Recent curriculum reinforcement learning for large language models (LLMs) typically rely on difficulty-based annotations for data filtering and ordering. However, such methods suffer from local optimization, where continual training on simple samples in the early steps can cause the policy to lose its exploration. We propose a novel schema, namely Hamiltonian curiosity augmented large language model reinforcement (HAMMER), that transfers diversity metrics, commonly used in dataset evaluation, into the dynamic reinforcement learning procedure, where training samples are ordered via a minimum-semantic Hamiltonian path making the initial training retrain more exploration. From a theoretical perspective of generalization bounds, diversity-driven ordering facilitates stable convergence. Empirical evaluations indicate that HAMMER stimulates model "curiosity" and consistently achieves a 3% to 4% average accuracy gain across diverse inference benchmark.
title HAMMER: Hamiltonian Curiosity Augmented Large Language Model Reinforcement
topic Machine Learning
Artificial Intelligence
Computation and Language
I.2.7
url https://arxiv.org/abs/2509.25240