Reference Feature Atlases for Auditing Language Models
A novel approach for auditing language models employs a reference feature atlas, which is a sparse library of features trained once on a reference panel and subsequently applied to new models via a linear decoder fitting. This method offers two perspectives: the atlas channel interprets the target based on established panel features, providing a consistent coordinate framework across different models; meanwhile, the residual channel captures features that the atlas fails to reproduce, explicitly signaling audits "outside the reference panel." Researchers developed leave-one-out atlases across five instruction-tuned models with 7-9 billion parameters and evaluated held-out Mistral and Qwen targets. The residual channel enabled perfect runtime control of three controlled LoRA hidden objectives, while matched controls remained unaffected, identifying the planted objective as the leading feature. The paper can be accessed on arXiv.
Key facts
- Reference feature atlas is a sparse feature library trained once on a reference panel.
- New models attach by fitting only a linear decoder.
- Atlas channel provides stable coordinate system across models.
- Residual channel learns features from what the atlas fails to reconstruct.
- Leave-one-out atlases trained over five 7-9B instruction-tuned models.
- Held-out targets: Mistral and Qwen.
- Three controlled LoRA hidden objectives injected into both targets.
- Residual channel makes planted mechanism perfectly controllable at runtime.
Entities
—