Researchers have introduced a novel method for auditing language models called Reference Feature Atlases. This approach involves training a sparse feature library on a reference panel of models, which can then be reused to interpret the internal features of new target models by fitting only a linear decoder. This technique provides two distinct views: one that maps the target model onto the interpreted features of the panel, offering a stable coordinate system, and another that identifies features the atlas fails to reconstruct, signaling aspects outside the reference panel. Experiments on Mistral and Qwen-2.5 models demonstrated the residual channel's ability to control injected mechanisms and reveal panel-relative phenomena like political framing. AI
IMPACT Provides a standardized framework for understanding and comparing the internal workings of different language models.
RANK_REASON The cluster contains an academic paper detailing a new methodology for auditing language models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
