May 12, 2026 Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Models 开源 SAE 套件,覆盖 Qwen3/3.5 系列 7 个模型变体的 14 组 SAE。展示 SAE 特征超越事后分析,成为模型开发实用接口:推理时引导、评估分析、数据工作流、后训练优化四大方向。 Sparse Autoencoders Mechanistic Interpretability LLM Steering Feature Analysis Benchmark Redundancy Toxicity Classification Safety Data Synthesis Reinforcement Learning