ConciseHint: Boosting Efficient Reasoning via Continuous Concise Hints during Generation

TL;DR

This work addresses the inefficiency of verbose reasoning in large reasoning models by proposing ConciseHint, an in-reasoning intervention that injects concise hints during the generation of intermediate reasoning steps. Hints can be manual or learned, and their intensity is adaptively controlled by the query’s complexity through $ au_k = \alpha + \beta \cdot l_k$, with the insertion position dynamically chosen as $p = \tau_k \cdot \min( (\tau_k-\alpha)/1024, 0.8 )$ to balance compute and accuracy. A learning-augmented variant, ConciseHint-T, enables further gains by interpolating embeddings between the original and optimized concise hints via $E_{interp} = \gamma E_{optim} + (1-\gamma) E_{ori}$. Experiments on GSM8K, AIME24, and GPQA-Diamond with DeepSeek-R1 and Qwen-3 series demonstrate substantial token reductions while maintaining accuracy, and show that ConciseHint can serve as a flexible plugin with compatibility to existing efficiency methods. The results highlight the potential of in-reasoning guidance to push the efficiency frontier in LRMs for complex reasoning tasks.

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, a critical issue is their tendency 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 learnable hints (manually designed or learned on concise data) during the generation of the reasoning. 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 while maintaining the performance well. Moreover, we show that ConciseHint is flexible and can be seamlessly integrated with existing methods to further push the upper bound of the efficiency.

Figures (8)

• Figure 1: Previous works mainly enhance conciseness before the actual reasoning (i.e., adding the control prompt or optimizing the model via SFT/RL), while we focus on intervening during the reasoning process to encourage conciseness, i.e., in-reasoning intervention. ConciseHint achieves this goal by continuously injecting learnable hints during the generation.
• Figure 2: The illustration of ConciseHint(-T) framework. Upon obtaining $\tau_k$, the LRM generates the next $\tau_k$ tokens, injects the hint, and updates $l_k$ and $\tau_k$ in sequence, repeating this cycle until the reasoning is finished. The corresponding pseudo-code is shown in \ref{['algo:concise_hint']}. There are two ways of obtaining the hint. Firstly, we can manually design the text with expertise and prior knowledge. Secondly, we can train the hint embeddings on concise reasoning data with SFT in a Next-token Prediction (NTP) way, which can further enhance the efficiency and acquire the controllability.
• Figure 3: Controllability curves obtained by adjusting $\gamma$ on Qwen3-1.7B. Different scattered points represent different $\gamma$ values.
• Figure 4: Ablation study of $\beta$ on GSM8K, AIME24, and GPQA-Diamond with Qwen3-4B. The black line denotes the accuracy of the original reasoning.
• Figure 5: Visualization of cost analysis for hint injection. Green and orange represent cached and non-cached tokens, respectively. The rectangles and triangles represent common and injected hint tokens, respectively.