Skip to content

Translation in progress Some entries still include text in the original language.

论文追踪· arXiv:2607.17117· Haoyan Luo, Mateo Espinosa Zarlenga, Mateja Jamnik·Indexed by this site · Originally published

Persistent SAE:为语言模型特征学习特定时间尺度

Persistent Sparse Autoencoders: Learning Feature-Specific Timescales in Language Model Representations

Quick read

Interpretability

Based on 论文 PDF (AI 生成); refer to the original

Placeholder

Questions
Placeholder
Methods
Placeholder
Experiments and results
Placeholder
Limitations and future work
Placeholder
AI summary研究者提出 Persistent Sparse Autoencoders(Persistent SAEs),在标准 SAE 基础上为每个特征学习一个持久性系数,仅靠重构目标就能学到特征特定的时间尺度。实验显示,Persistent SAEs 在保持有竞争力的重构质量的同时学到一组时间尺度谱:短时间尺度(快)特征保持局部可解释性,长时间尺度(慢)特征则累积用于识别当前上下文的信息。在提示注入监控案例研究中,慢特征能保留与注入相关的信号,并在长上下文中保持因果有效性。作者认为这为通过持久稀疏特征解释和监控语言模型提供了新途径。

研究者提出 Persistent Sparse Autoencoders(Persistent SAEs),在标准 SAE 基础上为每个特征学习一个持久性系数,仅靠重构目标就能学到特征特定的时间尺度。实验显示,Persistent SAEs 在保持有竞争力的重构质量的同时学到一组时间尺度谱:短时间尺度(快)特征保持局部可解释性,长时间尺度(慢)特征则累积用于识别当前上下文的信息。在提示注入监控案例研究中,慢特征能保留与注入相关的信号,并在长上下文中保持因果有效性。作者认为这为通过持久稀疏特征解释和监控语言模型提供了新途径。

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    Placeholder

    Placeholder

  2. What related research exists?

    Placeholder

    Placeholder

  3. How does the paper address this problem?

    Placeholder

    Placeholder

  4. What experiments did the paper conduct?

    Placeholder

    Placeholder

  5. What are some points that can be further explored?

    Placeholder

    Placeholder

  6. Summarize the main content of the paper

    Placeholder

    Placeholder

Read originalarxiv.org