ConciseHint: Boosting Efficient Reasoning via Continuous Concise Hints during Generation

2025-06-23

AI Summary

Large reasoning models (LRMs) often produce excessively long chains of thought for complex tasks, severely degrading inference efficiency. To address this, we propose a **dynamic conciseness-guidance mechanism operating during inference**: at each token-generation step, adaptive textual prompts—either manually crafted or trained on concise reasoning data—are injected, with prompt strength dynamically modulated based on real-time complexity estimation to constrain reasoning length. Unlike conventional “optimize-then-infer” paradigms, our approach enables online, generation-time conciseness intervention—a first in LRM inference. Evaluated on GSM8K using Qwen-3 4B, our method reduces average reasoning length by 65% while preserving near-original accuracy, yielding substantial gains in both efficiency and practical utility.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Recent advancements in large reasoning models (LRMs) like DeepSeek-R1 and OpenAI o1 series have achieved notable performance enhancements on complex reasoning tasks by scaling up the generation length by Chain-of-Thought (CoT). However, an emerging issue is their inclination to produce excessively verbose reasoning processes, leading to the inefficiency problem. Existing literature on improving efficiency mainly adheres to the before-reasoning paradigms such as prompting and reasoning or fine-tuning and reasoning, but ignores the promising direction of directly encouraging the model to speak concisely by intervening during the generation of reasoning. In order to fill the blank, we propose a framework dubbed ConciseHint, which continuously encourages the reasoning model to speak concisely by injecting the textual hint (manually designed or trained on the concise data) during the token generation of the reasoning process. Besides, ConciseHint is adaptive to the complexity of the query by adaptively adjusting the hint intensity, which ensures it will not undermine model performance. Experiments on the state-of-the-art LRMs, including DeepSeek-R1 and Qwen-3 series, demonstrate that our method can effectively produce concise reasoning processes while maintaining performance well. For instance, we achieve a reduction ratio of 65% for the reasoning length on GSM8K benchmark with Qwen-3 4B with nearly no accuracy loss.

Problem

• Reducing verbose reasoning in large models
• Intervening during generation for conciseness
• Adapting hint intensity to query complexity

Innovation

• Continuous concise hints during token generation
• Adaptive hint intensity based on query complexity
• Textual hints from manual design or concise data