NIRC is based on the National Key Laboratory of Network and Switching Technology at Beijing University of Posts and Telecommunications. It has built a technology matrix across four AI domains—intelligent cloud networking, natural language processing, computer vision, and machine learning systems—dedicated to solving real‑world problems, creating top‑tier systems, publishing high‑impact papers, and contributing significantly to the rapid advancement of China's network technology.
veRoCE is a ByteDance‑engineered RDMA protocol that retains RoCEv2 compatibility while adding multi‑path transmission, native out‑of‑order delivery, efficient loss detection and selective retransmission, allowing high‑performance clusters to operate over lossy Ethernet without PFC.
RecFlow, a new inference engine from Beijing University of Posts and Telecommunications and Meituan, tackles the resource mismatch of DLRM models by coordinating embedding and DNN operators at the intra‑SM level and introducing interference‑aware adaptive scheduling and incremental batching, achieving up to 9.34× higher throughput on RTX 3090.
DiffNBR introduces a dual‑path diffusion framework combined with an information‑bottleneck mechanism to jointly model spatial co‑occurrence and temporal evolution in next‑basket recommendation, achieving state‑of‑the‑art performance and effectively disentangling repetitive and exploratory purchase patterns.
At EMNLP 2025, the BUPT NIRC team presented a paper that introduces the ARR metric to quantitatively separate latent reasoning from factual shortcuts in LLMs, using Logit Lens and Attention Knockout to reveal distinct internal pathways and shares their conference experience.
The report summarizes NeurIPS 2025 in San Diego, highlighting four NIRC papers on noise‑robust 3D human pose estimation, LVLM video‑anomaly understanding, and hand‑object reconstruction, and discusses broader industry trends such as feed‑forward generation and large‑scale pre‑training showcased by leading AI companies.
The paper presents Atlas, a native three‑layer distributed verification system that replaces centralized tools with switch, region, and center adapters, achieving sub‑20 ms validation for thousands of nodes and up to 1500× speedup over EPVerifier, while supporting incremental updates and preserving scalability.
UniCodebook introduces a unified 2D‑3D discrete prior that combines continuous and discrete representations, enabling calibration‑free multiview 3D human pose estimation with superior noise robustness and higher accuracy, as demonstrated by state‑of‑the‑art results on Human3.6M and MPI‑INF‑3DHP.
The article analyzes the severe inference bottlenecks of large language models, CNNs, and recommendation systems and presents a suite of research‑driven accelerations—including token‑level pipeline parallelism (HPipe), KV‑cache clustering (ClusterAttn), quantization (QoKV), heterogeneous edge frameworks (DeepZoning, PICO), delay‑aware edge‑cloud scheduling (DECC), and operator choreography (RACE)—validated on real‑world industrial workloads.
This paper introduces SeDi, a semantics‑ and distribution‑aware cross‑tokenizer knowledge distillation framework that aligns teacher and student token spaces via bipartite graph components and top‑K re‑encoding, achieving state‑of‑the‑art performance and lower exposure bias on multiple LLM benchmarks.
MCP (Model Context Protocol) is an open standard that lets AI applications connect to external systems through a unified client‑server model, exposing Tools, Resources, and Prompts, while addressing security, permission, and audit concerns to make large‑model deployments more reusable and controllable.
