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

TL;DR

This work tackles Lazy Reasoning in large reasoning models when performing complex multi-turn tool use by highlighting a lack of sub-task decomposition. It introduces D-CORE, a two-stage framework combining self-distillation to enforce explicit task decomposition and a diversity-aware reinforcement learning objective (DA-GRPO) to preserve reflective, diverse reasoning. Empirical results on BFCLv3 and tau-bench show state-of-the-art performance for 8B and 14B models, with 8B achieving 77.7% and 14B 79.3% overall accuracy, while the 14B model remains far smaller than the prior state-of-the-art. The approach demonstrates robust generalization to unseen tasks (ACEBench, tau2, BFCLv4-agentic) and indicates that structured reasoning and decomposition can dramatically improve tool-use proficiency in LRMs, enabling smaller models to rival larger ones. Future work will extend to multimodal settings and explore additional RL strategies for efficient reasoning.

Abstract

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~(\underline{\textbf{D}}ecomposing tasks and \underline{\textbf{Co}}mposing \underline{\textbf{Re}}asoning 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\%, outperforming 70B models despite being 5$\times$ smaller. The source code is available at https://github.com/alibaba/EfficientAI.

Figures (14)

• Figure 1: Comparison of baseline and D-CORE trained LRMs in complex tool use scenarios. Baseline LRMs exhibit "Lazy Reasoning" with repetitive reflection and incorrect answers, while D-CORE trained LRMs decompose tasks into executable subtasks.
• Figure 2: Comparison of LRM Qwen3 vs. instruct LLM xLAM2 performance on BFCLv3 Parallel, Irrelevance , Multi-turn and $\tau$-bench task.
• Figure 3: (a) Distribution of four behavioral categories across rollout samples. (b) Ratio of lazy reasoning in different tasks.(c) Accuracy degradation caused by subtask composition.(d) Accuracy improvement through manual task decomposition.
• Figure 4: Overview of D-CORE self-distillation stage. Through self-distillation, a LRM with task decomposition and subtask execution capabilities is acquired.
• Figure 5: (a) Effect of self-distillation on reward variance. SD: self-distillation. (b) Top 10 high-entropy tokens for Qwen3-8B-SD.

Theorems & Definitions (7)

• Theorem 3.1: Prevention of Learning Stagnation
• Theorem 3.2: Entropy Reduction Property of DA-GRPO
• Theorem 1.1: Prevention of Learning Stagnation
• proof
• Theorem 1.2: Entropy Reduction Property of DA-GRPO
• proof
• Remark 1.3