SelfBudgeter Introduces Adaptive Token Allocation to Trim LLM Reasoning Length

Global: SelfBudgeter: Adaptive Token Allocation for Efficient LLM Reasoning

Researchers Zheng Li, Qingxiu Dong, Jingyuan Ma, Di Zhang, Kai Jia, and Zhifang Sui presented a new self‑adaptive reasoning strategy called SelfBudgeter in a paper submitted to arXiv on 16 May 2025 and revised through 9 Jan 2026. The approach targets large language models (LLMs) that often consume excessive tokens during reasoning, aiming to reduce resource waste and user latency while preserving answer quality.

Background

Current LLM reasoning pipelines generate token sequences that can be disproportionately long for relatively simple queries. This inefficiency leads to higher computational costs and longer response times, which can hinder real‑time applications and increase operational expenses.

Methodology

SelfBudgeter first trains the model to self‑estimate the necessary reasoning budget based on the incoming query. It then applies a budget‑guided guided policy optimization (GPRO) reinforcement‑learning framework that conditions generation on the predicted budget, allowing the model to truncate or expand its output dynamically.

Performance Results

Experimental evaluation on mathematical reasoning benchmarks shows that SelfBudgeter achieves an average response‑length compression of 61 % while maintaining comparable accuracy to baseline models. The authors note that the compression is achieved without statistically significant drops in task performance.

User Controls

The system provides users with a preview of the expected generation length, enabling them to decide whether to proceed or halt the process. Additionally, users may set explicit token budgets before inference, granting direct control over reasoning depth.

Implications

By reducing token consumption, SelfBudgeter could lower inference costs for cloud‑based LLM services and improve responsiveness in interactive applications. The adaptive budgeting mechanism also offers a transparent way for end‑users to manage computational resources, which may be valuable for deployment in constrained environments.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.