Aware First, Think Less: Dynamic Boundary Self-Awareness Drives Extreme Reasoning Efficiency in Large Language Models (AAAI2026)

---

We provide training scripts for Qwen2 and Qwen3: verl_drsaf/qwen2_7b.sh and verl_drsaf/qwen3_8b.sh. You will need to fill in your training and test datasets, as well as the local path to the base model for training. You can find our prompt in verl_drsaf/verl/utils/dataset/rl_dataset.py.

Models:

• https://huggingface.co/LARG/DRSAF-Qwen3-8B-distill
• https://huggingface.co/LARG/DRSAF-Qwen2.5-7B-distill
• https://huggingface.co/LARG/DRSAF-Qwen2.5-7B-distill-extreme

To reproduce the results, please follow the prompt located in verl_drsaf/verl/utils/dataset/rl_dataset.py. Removing the mandatory self-capability boundary evaluation before inference will moderately improve accuracy but increase token length.

---

Recent advancements in large language models (LLMs) have greatly improved their capabilities on complex reasoning tasks through Long Chain-of-Thought (CoT). However, this approach often results in substantial redundancy, impairing computational efficiency and causing significant delays in real-time applications. To improve the efficiency, current methods often rely on human-defined difficulty priors, which do not align with the LLM’s self-awared difficulty, leading to inefficiencies. In this paper, we introduce the Dynamic Reasoning-Boundary Self-Awareness Framework (DR. SAF), which enables models to dynamically assess and adjust their reasoning depth in response to problem complexity. DR. SAF integrates three key components: Boundary Self-Awareness Alignment, Adaptive Reward Management, and a Boundary Preservation Mechanism. These components allow models to optimize their reasoning processes, balancing efficiency and accuracy without compromising performance. Our experimental results demonstrate that DR. SAF achieves a 49.27% reduction in total response tokens with minimal loss in accuracy. The framework also delivers a 6.59x gain in token efficiency and a 5x reduction in training time, making it well-suited to resource-limited settings. During extreme training, DR. SAF can even surpass traditional instruction-based models in token efficiency with more than 16% accuracy improvement.

---

You can adjust the cfrb_bparameter and the pfrb_b_extparameter in verl_drsaf\verl\utils\reward_score\init.py. Decreasing both parameters will result in a higher degree of compression, while increasing both will better preserve performance.

---

This work is implemented based on VERL. We extend our gratitude to the authors for their excellent contributions. You can adapt our code using VERL's installation procedure. If you find our work inspiring or helpful, please cite:

@article{chen2025aware,
  title={Aware First, Think Less: Dynamic Boundary Self-Awareness Drives Extreme Reasoning Efficiency in Large Language Models},
  author={Chen, Qiguang and Peng, Dengyun and Liu, Jinhao and Su, HuiKang and Guan, Jiannan and Qin, Libo and Che, Wanxiang},
  journal={arXiv preprint arXiv:2508.11582},
  year={2025}
}