Graph Retrieval-Augmented Generation (GraphRAG) has shown great effectiveness in enhancing the reasoning abilities of Large Language Models (LLMs) by leveraging graph structures for knowledge representation and modeling complex real-world relationships. However, existing GraphRAG methods still face significant bottlenecks when handling complex problems that require multi-hop reasoning, as their query and retrieval phases are largely based on pre-defined heuristics and do not fully utilize the reasoning potentials of LLMs. To address this problem, we propose GraphRAG-R1, an adaptive GraphRAG framework by training LLMs with process-constrained outcome-based reinforcement learning (RL) to enhance the multi-hop reasoning ability. Our method can decompose complex problems, autonomously invoke retrieval tools to acquire necessary information, and perform effective reasoning. Specifically, we utilize a modified version of Group Relative Policy Optimization (GRPO) that supports rollout-with-thinking capability to train the model. Next, we design two process-constrained reward functions. To handle the shallow retrieval problem, we design a Progressive Retrieval Attenuation (PRA) reward to encourage essential retrievals. Then, to handle the over-thinking problem, we design a Cost-Aware F1 (CAF) reward to balance the model performance with computational costs. We further design a phase-dependent training strategy, containing three training stages corresponding to cold start and these two rewards. These stages empower GraphRAG with format following, behavior shaping, and smartness optimization abilities, respectively. Lastly, our method adopts a hybrid graph-textual retrieval to improve the reasoning capacity. Extensive experimental results demonstrate that GraphRAG-R1 significantly boosts LLM capabilities in solving complex reasoning problems compared to state-of-the-art GraphRAG methods on both in-domain and out-of-domain datasets. Furthermore, our framework can be flexibly integrated with various existing retrieval methods, consistently delivering performance improvements.

Light detection and ranging (LiDAR) sensors are crucial for autonomous vehicles to accurately perceive the surrounding environment. However, the sparsity and irregularity of large-scale LiDAR point clouds (LPCs) bring challenges for storage and transmission. Meanwhile, existing works usually adopt insufficient context and bring intolerable computation complexity, especially for high-precision LPC reconstruction. To address these problems, we propose a novel self-attention-based framework for LPC compression and reconstruction in autonomous driving environments. Specifically, our approach employs a robust backbone for octree-based feature extraction, which can be pretrained and easily extended to various tasks, thereby reducing the need for extensive task-specific architectural modifications. The backbone constructs node sequences of octree by nonoverlapping context windows and shares the result of a multihead self-attention (MSA) operation among them. Considering the similarity in features among sibling nodes, we design a locally enhanced module for exploiting sibling features and a positional encoding generator for enhancing the translation invariance of the octree node sequence. During postprocessing, we further propose an offset prediction model to reduce coordinate distortions caused by voxelization. Experimental results indicate that compared to the benchmark geometry-based point cloud compression (GPCC), our approach achieves gains of up to 54.4% for geometry and 6.8% for intensity, while compared to the attention-based baseline, we achieve up to 99% reduction in coding time. We believe that our approach effectively mines the spatial geometric features in LPCs and has low coupling for specific tasks, which will boost the related applications from algorithm optimization to industrial products.

High-quality LiDAR point cloud (LPC) coding is essential for efficiently transmitting and storing the vast amounts of data required for accurate 3D environmental representation. The Octree-based entropy coding framework has emerged as the predominant method, however, previous study usually overly relies on large-scale attention-based context prediction to encode Octree nodes, overlooking the inherent correlational properties of this structure. In this paper, we propose a novel Graph-driven Attention-based Entropy Model (GAEM), which adopts partitioned graph attention mechanisms to uncover contextual dependencies among neighboring nodes. Different from the Cartesian coordinate-based coding mode with higher redundancy, GAEM uses the multi-level spherical Octree to organize point clouds, improving the quality of LPC reconstruction. GAEM combines graph convolution for node feature embedding and grouped-graph attention for exploiting dependency among contexts, which preserves performance in low-computation using localized nodes. Besides, to further increase the receptive field, we design a high-resolution cross-attention module introducing sibling nodes. Experimental results show that our method achieves state-of-the-art performance on the LiDAR benchmark SemanticKITTI and MPEG-specified dataset Ford, compared to all baselines. Compared to the benchmark GPCC, our method achieves gains of up to 53.9% and 53.6% on SemanticKITTI and Ford while compared to the sibling-introduced methods, we achieve up to 42.3% and 44.7% savings in encoding/decoding time. In particular, our GAEM allows for extension to downstream tasks (i.e.,vehicle detection and semantic segmentation), further demonstrating the practicality of the method.

LiDAR point cloud (LPC) compression is an indispensable component for 3D vision tasks, especially for dynamic point clouds. However, the existing methods based on traditional spatial-temporal attention are immature, causing little improvement in inter-frame feature extraction. In this paper, we propose Diverse Attention-based Point Cloud Compression (DAPCC), an LPC compression entropy model combining aggregation embedding modules for temporal point matching and spatial-temporal attention blocks for dynamic Octree node encoding, which can effectively utilize the change information of dynamic point clouds. Specifically, we first introduce aggregation embedding to match the Octree sequences from two sweeps to establish temporal correlation. To effectively capture the feature details, we further design local and global combined attention for the spatial-temporal information of point clouds which can focus on the whole context. Finally, we organize a symmetric MLP module capable of strengthening vital features. We conduct experiments of static and dynamic compression on both indoor/outdoor point cloud benchmark datasets (i.e., ScanNet, SemanticKITTI, and MPEG Common Test Conditions (CTC) Category 3 datasets) and downstream applications (i.e., vehicle detection and semantic segmentation). Compared with the previous state-of-the-art methods, our method achieves up to 14.7% bpp and 45% decoding time savings and adapts to the downstream tasks with almost no impact on performance.

Point cloud compression is an essential technology for efficient storage and transmission of 3D data. Previous methods usually use hierarchical tree data structures for encoding the spatial sparseness of point clouds. However, the node context within the tree is not fully discovered since the feature space among nodes varies significantly. To address this problem, we innovatively represent the LiDAR points in a two-octree structure instead of using traditional single-octree coding, and then design the cross-attention model to capture the hierarchical features between different octrees, of which each octree incorporates a transformer-based deep entropy model and an arithmetic encoder. Besides, we introduce the untied cross-aware position encoding with principal component analysis and different projection matrices, which enhances the correlations over two octrees’ attention feature embeddings. Experimental results show that our method outperforms the previous state-of-the-art works, achieving up to 8.2% Bpp savings on point cloud benchmark datasets with different lasers.

Point cloud compression with a higher compression ratio and tiny loss is essential for efficient data transportation. However, previous methods that depend on 3D convolution or frequent multi-head self-attention operations bring huge computations. To address this problem, we propose an octree-based Transformer compression method called OctFormer, which does not rely on the occupancy information of sibling nodes. Our method uses non-overlapped context windows to construct octree node sequences and share the result of a multi-head self-attention operation among a sequence of nodes. Besides, we introduce a locally-enhance module for exploiting the sibling features and a positional encoding generator for enhancing the translation invariance of the octree node sequence. Compared to the previous state-of-the-art works, our method obtains up to 17% Bpp savings compared to the voxel-context-based baseline and saves an overall 99% coding time compared to the attention-based baseline.