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.
The article presents Baidu's PLATO open‑domain dialogue system, detailing its evolution from expert‑rule to retrieval‑based and large‑scale generative models, describing its hidden‑variable architecture, major research challenges such as persona stability, long‑term memory, knowledge accuracy, and showcasing real‑world applications and Q&A from a DataFunSummit2022 livestream.
This article introduces Baidu's PLATO open‑domain dialogue technology, explains the evolution from rule‑based to retrieval‑based and large‑scale generative models, discusses major challenges such as persona stability, long‑term memory, knowledge accuracy, and proactive conversation, and showcases real‑world applications and Q&A insights.
The paper presents MME‑CRS, a champion method for DSTC10 open‑domain dialogue evaluation that combines seven diverse metrics—fluency, relevance, topic coherence, engagement, and three specificity measures—using a correlation‑re‑scaling algorithm to weight each metric, achieving state‑of‑the‑art Spearman correlation and top rankings across multiple evaluation dimensions.
