D-CORE: Incentivizing Task Decomposition in Large Reasoning Models for Complex Tool Use

Effective tool use and reasoning are essential capabilities for large reasoning models (LRMs) to address complex real-world problems. Through empirical analysis, we identify that current LRMs lack the capability of sub-task decomposition in complex tool use scenarios, leading to Lazy Reasoning. To address this, we propose a two-stage training framework D-CORE (Decomposing tasks and Composing Reasoning processes) that first incentivize the LRMs' task decomposition reasoning capability via self-distillation, followed by diversity-aware reinforcement learning (RL) to restore LRMs' reflective reasoning capability. D-CORE achieves robust tool-use improvements across diverse benchmarks and model scales. Experiments on BFCLv3 demonstrate superiority of our method: D-CORE-8B reaches 77.7\% accuracy, surpassing the best-performing 8B model by 5.7\%. Meanwhile, D-CORE-14B establishes a new state-of-the-art at 79.3\%. Code is available at https://github.com/alibaba/EfficientAI; the model is available at https://huggingface.co/bowiehsu/D-CORE-8B; and the data are available at https://huggingface.co/datasets/bowiehsu/D-CORE-Self-Distillation-50K.

AI assistants that can use software tools — booking flights, processing refunds, or managing files — often stumble when a task requires multiple steps. We discovered that today's advanced reasoning models fall into a pattern we call "Lazy Reasoning": instead of breaking a complex request into a clear plan, they loop through unproductive trial-and-error, wasting computation without solving the problem. To fix this, we developed D-CORE, a two-stage training method. First, the model teaches itself to decompose hard tasks into manageable subtasks, without needing a more powerful external model as a tutor. Second, a new algorithm called DA-GRPO restores the model's ability to reflect and adapt by using the uncertainty in its own word choices as a learning signal. The result: a model with only 14 billion parameters outperforms models five times its size on realistic multi-step tool-use benchmarks, cutting errors caused by Lazy Reasoning from 45% to 6%. This work brings reliable, cost-efficient AI agents closer to real-world deployment in customer service, enterprise automation, and beyond.