Microsoft Research announces multiple new advances in AI and robotics, covering cultural sentiment analysis, robot assembly learning, and more

ME News Report, April 7th (UTC+8), Microsoft Research recently released several latest advancements in artificial intelligence and robotics. In sentiment analysis, the research team proposed a new diagnostic framework that treats emotions as culturally embedded concepts, testing large language models’ understanding ability in low-resource, culturally nuanced contexts by combining manual annotation, sentiment reversal counterfactual examples, and metric-based interpretability assessments. The article points out that the findings show GPT-4 series models have more stable and interpretable reasoning, while open-source models like Gemma-3-27B and Mistral-7B are more sensitive to ambiguity and subtle emotional changes. In robotics, the study introduced the OmniReset framework, which can teach complex assembly tasks through reinforcement learning in simulation and perform zero-shot transfer to the physical world. Regarding AI agents, the team developed the ATLAS framework, enabling small models to handle long-sequence, multi-step tasks through reinforcement fine-tuning. The VeriStruct project extends AI-assisted formal verification to complete data structure modules in the Rust language. Additionally, Microsoft Research will participate in CHI 2026 with 26 papers, 5 workshops, and a social impact award. (Source: InFoQ)