Tag
Discusses a new paper from Tsinghua/Zhipu AI on asynchronous reinforcement learning for agents, and notes that their previous GLM-5.2 model uses a critic instead of GRPO.
Tsinghua's SAO algorithm addresses stability and off-policy drift in asynchronous reinforcement learning for LLMs, achieving consistent improvements over GRPO on agentic coding and reasoning benchmarks, and is used to train the GLM-5.2 model.
Tsinghua team proposes EurekAgent, arguing that the bottleneck in autonomous scientific research is environment engineering rather than smarter Agents. By engineering four dimensions—permissions, artifacts, budgets, and human-AI collaboration—they achieve SOTA on several mathematical and kernel engineering tasks, discovering a new optimal arrangement for 26 circles for under $11.
The article reviews the early development history of large models in China, pointing out that BAAI supported the earliest Qingyuan CPM (2020) and Wudao 1.0 (2021), and corrects the claim that Huawei Pangu was the first domestic large model.
Tsinghua University NLP Lab open-sourced 269 projects on GitHub, covering large model training, knowledge graphs, Prompt learning, parameter fine-tuning, and more, including well-known projects such as OpenPrompt, OpenNRE, OpenKE, UltraChat, and OpenDelta. Suitable for AI researchers and application developers.