TL;DR: Too Long, Do Re-weighting for Efficient LLM Reasoning Compression
This addresses the challenge of efficient LLM reasoning for researchers and practitioners, but it appears incremental as it builds on existing CoT and re-weighting techniques.
The paper tackles the problem of inefficient language reasoning in LLMs during inference with long outputs by proposing a dynamic ratio-based training pipeline that balances weights between System-1 and System-2 data, resulting in a nearly 40% reduction in output tokens while maintaining reasoning accuracy.
Large Language Models (LLMs) have recently achieved remarkable progress by leveraging Reinforcement Learning and extended Chain-of-Thought (CoT) techniques. However, the challenge of performing efficient language reasoning--especially during inference with extremely long outputs--has drawn increasing attention from the research community. In this work, we propose a dynamic ratio-based training pipeline that does not rely on sophisticated data annotations or interpolation between multiple models. We continuously balance the weights between the model's System-1 and System-2 data to eliminate redundant reasoning processes while preserving the model's reasoning capability. We validate our approach across models on DeepSeek-R1-Distill-7B and DeepSeek-R1-Distill-14B and on a diverse set of benchmarks with varying difficulty levels. Our method significantly reduces the number of output tokens by nearly 40% while maintaining the accuracy of the reasoning. Our code and data will be available soon.