Large reasoning models (LRMs) have recently achieved significant progress in complex reasoning tasks, aided by reinforcement learning with verifiable rewards. However, LRMs often suffer from overthinking, expending excessive computation on simple problems and reducing efficiency. Existing efficient reasoning methods typically require accurate task assessment to preset token budgets or select reasoning modes, which limits their flexibility and reliability. In this work, we revisit the essence of overthinking and identify that encouraging effective steps while penalizing ineffective ones is key to its solution. To this end, we propose a novel rule-based verifiable stepwise reward mechanism (VSRM), which assigns rewards based on the performance of intermediate states in the reasoning trajectory. This approach is intuitive and naturally fits the step-by-step nature of reasoning tasks. We conduct extensive experiments on standard mathematical reasoning benchmarks, including AIME24 and AIME25, by integrating VSRM with PPO and Reinforce++. Results show that our method achieves substantial output length reduction while maintaining original reasoning performance, striking an optimal balance between efficiency and accuracy. Further analysis of overthinking frequency and pass@k score before and after training demonstrates that our approach indeed effectively suppresses ineffective steps and encourages effective reasoning, fundamentally alleviating the overthinking problem.

The efficient utilization of underground spaces is a crucial strategy for mitigating land scarcity and expanding habitable environments. 3D Gaussian Splatting (3DGS) has emerged as a key enabler for enhancing robotic perception and spatial digitalization in underground spaces due to its unique advantages. However, illumination variation, sensor noise, and geometric degradation of underground spaces significantly degrade the localization accuracy, which compromises the robustness of existing Simultaneous Localization and Mapping (SLAM) systems. Therefore, we propose an RGB-D perception-enhanced 3DGS SLAM method. First, a multi-dimensional data enhancement and correction pipeline is introduced, integrating Multi-Scale Retinex with Color Restoration (MSRCR), Side Window Filtering (SWF), and adaptive gamma correction in the Hue-Intensity-Saturation (HIS) color space to improve low-light visual fidelity, while a depth completion network is developed to perform hole-filling in depth data. Second, a keyframe selection method based on the hybrid metric is proposed, which incorporates consistency constraints and multi-view overlap analysis to balance computational efficiency and representational completeness. Finally, a dual-constraint Gaussian management strategy is introduced, integrating an opacity threshold and observation frequency to filter out invalid Gaussian ellipsoids. At the same time, loop closure detection ensures global trajectory consistency and mapping accuracy. To validate the proposed method, experiments were conducted in typical underground spaces, including coal mine tunnels and underground parking garages, using a custom-designed underground mobile robot platform. The results demonstrate that, compared to state-of-the-art methods, the proposed method achieves a 13.8% improvement in Peak Signal-to-Noise Ratio (PSNR) over the best benchmark, while also achieving competitive trajectory accuracy and computational efficiency. These findings provide strong support for the development of digital twin systems in underground spaces.