Concise Reasoning: Making LLMs More Efficient Through Self-Training

We recently published Self-Training Elicits Concise Reasoning in Large Language Models, where we fine-tune LLMs to generate more concise reasoning paths without sacrificing accuracy. We would like to create a Hugging Face Space to showcase our models and demonstrate their efficiency.

Objectives:

• Deploy an interactive demo to demonstrate concise reasoning outputs from our models

Impact:

Verbose outputs from LLMs increase inference costs and limit deployment in resource-constrained settings. Our approach reduces token usage significantly, making LLM-based applications more cost-effective and sustainable—especially for tasks involving chain-of-thought reasoning.

We kindly request access to GPU resources to host this demo and make our research easily accessible to the community.