LTMG: Less Thought, More Gain—A Token Efficient Reasoning Framework for Large Language Models

ACL ARR 2025 May Submission7305 Authors

Abstract:

Large reasoning models (LRMs) demonstrate impressive logical capabilities and are able to solve complex problems through multi-step deduction. However, they often generate unnecessarily long chains of thought that significantly increase inference cost without yielding proportional gains in accuracy.
We present an efficient logical reasoning framework that uses a dynamic router to instantly assess the difficulty of the question.
For simple questions, the router dispatches them to a reinforcement learning length-control model that enforces strict control of the output length through learned policies during inference; for complex questions, they are dispatches to LRMs employing customized prompt strategies for different type of questions that ensures computational resources are invested only where they are truly valuable.
Comprehensive experiments on arithmetic, logical, and commonsense benchmarks show that our framework reduces token usage, outperforming prior reasoning systems and demonstrating that efficiency stems from thinking just enough and only when necessary.

Paper Type: Long

Research Area: Efficient/Low-Resource Methods for NLP

Research Area Keywords: LLM Efficiency; parameter-efficient training; prompt strategy

Contribution Types: Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 7305