EVENT DOSSIER
On September 7, 2026, arXiv published research results titled SocialBuddy, which is the first intelligent agent search framework specifically designed for social scenarios. The study built the SocialEnv simulation environment containing 200,000 user profiles, 10 million social posts, and 50,000 reasoning trajectories, and designed the SocialPO hybrid granularity optimization framework to address the problem of sparse rewards. The finally established SocialSearch Benchmark was used to quantify and evaluate performance, and experiments showed that the SocialBuddy-35B model outperformed larger-scale frontier LLMs in terms of performance. The relevant code and dataset will be released after the article is accepted.
arXiv:2609.01641v2 发布 SocialBuddy，首个专为社交场景设计的智能体搜索框架。该研究构建了包含 20 万用户画像、1000 万社交帖子及 5 万推理轨迹的 SocialEnv 模拟环境，并设计了 SocialPO 混合粒度优化框架以解决稀疏奖励问题。最终建立的 SocialSearch Benchmark 用于量化评估能力，实验表明 SocialBuddy-35B 显著优于更大规模的 frontier LLMs。代码与数据集将在文章接收后发布。