BudgetThinker: Empowering Budget-aware LLM Reasoning with Control Tokens

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wen, Hao, Wu, Xinrui, Sun, Yi, Zhang, Feifei, Chen, Liye, Wang, Jie, Liu, Yunxin, Liu, Yunhao, Zhang, Ya-Qin, Li, Yuanchun
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908509286694912
author Wen, Hao
Wu, Xinrui
Sun, Yi
Zhang, Feifei
Chen, Liye
Wang, Jie
Liu, Yunxin
Liu, Yunhao
Zhang, Ya-Qin
Li, Yuanchun
author_facet Wen, Hao
Wu, Xinrui
Sun, Yi
Zhang, Feifei
Chen, Liye
Wang, Jie
Liu, Yunxin
Liu, Yunhao
Zhang, Ya-Qin
Li, Yuanchun
contents Recent advancements in Large Language Models (LLMs) have leveraged increased test-time computation to enhance reasoning capabilities, a strategy that, while effective, incurs significant latency and resource costs, limiting their applicability in real-world time-constrained or cost-sensitive scenarios. This paper introduces BudgetThinker, a novel framework designed to empower LLMs with budget-aware reasoning, enabling precise control over the length of their thought processes. We propose a methodology that periodically inserts special control tokens during inference to continuously inform the model of its remaining token budget. This approach is coupled with a comprehensive two-stage training pipeline, beginning with Supervised Fine-Tuning (SFT) to familiarize the model with budget constraints, followed by a curriculum-based Reinforcement Learning (RL) phase that utilizes a length-aware reward function to optimize for both accuracy and budget adherence. We demonstrate that BudgetThinker significantly surpasses strong baselines in maintaining performance across a variety of reasoning budgets on challenging mathematical benchmarks. Our method provides a scalable and effective solution for developing efficient and controllable LLM reasoning, making advanced models more practical for deployment in resource-constrained and real-time environments.
format Preprint
id arxiv_https___arxiv_org_abs_2508_17196
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle BudgetThinker: Empowering Budget-aware LLM Reasoning with Control Tokens
Wen, Hao
Wu, Xinrui
Sun, Yi
Zhang, Feifei
Chen, Liye
Wang, Jie
Liu, Yunxin
Liu, Yunhao
Zhang, Ya-Qin
Li, Yuanchun
Machine Learning
Artificial Intelligence
Recent advancements in Large Language Models (LLMs) have leveraged increased test-time computation to enhance reasoning capabilities, a strategy that, while effective, incurs significant latency and resource costs, limiting their applicability in real-world time-constrained or cost-sensitive scenarios. This paper introduces BudgetThinker, a novel framework designed to empower LLMs with budget-aware reasoning, enabling precise control over the length of their thought processes. We propose a methodology that periodically inserts special control tokens during inference to continuously inform the model of its remaining token budget. This approach is coupled with a comprehensive two-stage training pipeline, beginning with Supervised Fine-Tuning (SFT) to familiarize the model with budget constraints, followed by a curriculum-based Reinforcement Learning (RL) phase that utilizes a length-aware reward function to optimize for both accuracy and budget adherence. We demonstrate that BudgetThinker significantly surpasses strong baselines in maintaining performance across a variety of reasoning budgets on challenging mathematical benchmarks. Our method provides a scalable and effective solution for developing efficient and controllable LLM reasoning, making advanced models more practical for deployment in resource-constrained and real-time environments.
title BudgetThinker: Empowering Budget-aware LLM Reasoning with Control Tokens
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2508.17196