Technology is the core driver of Ant's future creation.
This paper presents a lightweight, high‑efficiency framework called Triple Alignment with Rationale Generation (TAG) that transforms knowledge‑based visual question answering into a contrastive learning task, dramatically reducing trainable parameters while achieving state‑of‑the‑art performance on major KVQA benchmarks.
Three award‑winning papers—OceanBase’s unitized database architecture for billion‑scale map services, a video‑diffusion‑based dynamic typography system that animates text semantically, and the FinBench LDBC financial graph benchmark—are examined, highlighting their design, experimental results, and impact on industry applications.
The paper presents CMA+DB, a hierarchical multi‑agent framework that automatically tunes database parameters across diverse workloads by combining classification‑based collaboration, layered training, and joint action selection, achieving superior performance, faster convergence, and strong generalization compared with existing tuning methods.
This article highlights two cutting‑edge works from Ant Group’s research team—HAWK, a fixed‑word key decomposition switching accelerator that overcomes hardware challenges for FHE, and FHEFusion, a compiler framework that introduces operator fusion to dramatically speed CKKS‑based DNN inference—showcasing their designs, optimizations, and experimental gains.
AEnvironment is an open‑source, unified environment platform for Agentic Reinforcement Learning that abstracts all resources as services via the MCP protocol, enabling trillion‑scale model training, rapid app generation, benchmark integration, and seamless deployment through a high‑performance ASandbox runtime.
This article explains how the open‑source AReaL framework boosts large‑scale reinforcement learning by separating agent execution from training logic, introducing a decoupled Agentic RL service and a Single‑Controller architecture that improves data flow, fault tolerance, and GPU utilization.
FinEval‑KR, a new EMNLP2025 evaluation framework co‑authored by Shanghai University of Finance and Economics and Ant Group, separates knowledge coverage from logical reasoning to reveal why financial LLMs often hallucinate on calculation tasks, introduces KS, RS, and CS metrics, and ranks 18 state‑of‑the‑art models on a rigorously curated finance dataset.
AState is a general‑purpose state data management system for reinforcement‑learning tasks that tackles low IO efficiency, slow weight synchronization, and state‑recovery challenges, achieving sub‑10‑second weight sync for trillion‑parameter models through a three‑layer architecture, zero‑redundancy transfers, and hardware‑aware co‑design, with the code openly available on GitHub.
The article explains the design, implementation, and performance of the AMem NCCL‑Plugin, a lightweight extension to NVIDIA's NCCL that enables transparent offloading and rapid recovery of GPU memory during reinforcement‑learning training of trillion‑parameter models, detailing its architecture, APIs, benchmarks, installation steps, and integration guidelines.
Awex is a high‑performance Python framework that synchronizes training and inference weights for trillion‑parameter reinforcement‑learning models in seconds, using unified conversion, metadata management, and NCCL/RDMA transfer plans, dramatically reducing RL training latency and supporting diverse parallel strategies.
