Boyi Deng, Xu Wang, Yaoning Wang, Yunyang Wan, Yubo Ma, Baosong Yang, Haoran Wei, Jialong Tang et al.
Qwen-Scope open-sources 14 groups of sparse autoencoders (SAEs) for 7 Qwen3/3.5 models (dense and MoE) and presents methods to use SAE features for inference-time steering, evaluation analysis, data synthesis, and fine-tuning optimization.
The internal decision-making processes of large language models (LLMs) are opaque, making inspection, control, and systematic improvement difficult. While SAEs are promising in mechanistic interpretability research, most work remains post-hoc and rarely integrates into actual model development workflows.
Train SAEs on 7 model variants from Qwen3 and Qwen3.5 series (dense and MoE) to build 14 groups of SAEs. Utilize these SAE features in four development directions: (i) inference-time steering: adjust feature directions to manipulate language, concepts, and preferences; (ii) evaluation analysis: analyze benchmark redundancy and capability coverage via activated features; (iii) data-centric workflows: multilingual toxicity classification and safety data synthesis; (iv) post-training optimization: incorporate SAE signals into SFT and RL objectives to mitigate undesirable behaviors such as code-switching and repetition.
Demonstrate that SAEs can serve as reusable representation-level interfaces for diagnosing, controlling, evaluating, and improving LLMs, beyond post-hoc analysis tools. Open-source Qwen-Scope to accelerate mechanistic research and practical workflows.