Researchers have developed a new metric, k*, to measure the sparsity of features required to understand task information within language model representations. This metric was applied to 19 language models from five different labs across 14 core tasks, revealing that most models consistently require similar feature counts for readable tasks. The study also found that instruction tuning selectively reduces k*, and that successful learning of targets can be achieved with single-feature readouts, contrasting with controls that require tens of features. AI
IMPACT Provides a new method for analyzing and understanding the internal organization of information within language models.
RANK_REASON Academic paper detailing a new metric and its application to language models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →