EVENT DOSSIER
On September 7, 2026, arXiv published a research paper titled “Almost Free State Prediction Separation”. This study proposes a new unsupervised method aimed at solving the problem of separating different states from mixed data in the absence of labels. By constructing specific loss functions and training strategies, this method achieves effective state prediction and separation with low computational cost, providing new tools for analyzing complex dynamic systems.
arXiv:2609.03807v2 提出一种“几乎免费”的状态预测分离方法，利用共享权重骨干中的并行预测流而非序列额外 token 来生成思考 token。该技术在 10 亿参数模型上使下一个 token 预测准确率提升 2-3 个百分点。由于暂停 token 占用现有位置，推理时不增加上下文长度、KV 缓存及延迟，训练成本较优化预训练流程降低至 x1.14。该方法实现了在标准 next token 训练 Transformer 基础上保持同算力、同参数和同 token 数量的性能提升。