The article analyzes RLVER, a reinforcement‑learning framework that integrates a user simulator as both environment and reward source, overcomes three major RL challenges, and elevates the Qwen2.5‑7B model’s Sentient‑Benchmark score from 13.3 to 79.2, rivaling GPT‑4o and Gemini 2.5 Pro.
This article details Huolala’s self‑developed Text‑to‑Speech system, outlining its architecture, the challenges of latency, naturalness, and language support, and the innovative solutions—including streaming synthesis, emotion modeling, and transfer‑learning‑based multilingual capabilities—that deliver more flexible and realistic voice interactions.
This article reviews Alibaba's digital‑human (virtual avatar) research over the past few years, covering the product’s evolution, a six‑stage pipeline for building digital humans, solutions to key challenges in realism, multimodal interaction, and the open‑source MMTK algorithm library.
This article reviews Alibaba's XiaoMi digital human technology, covering its evolution since 2019, a six‑stage pipeline for building avatars, methods to enhance emotional, textual, vocal, and motion expressiveness, and approaches for improving long‑term interactive capabilities such as controllable script generation, multimodal QA, sign‑language translation, and intelligent behavior decision, culminating in the release of the MMTK multimodal algorithm library.
