A new research paper proposes a novel perspective on the internal workings of large language models (LLMs), suggesting that temporal correlations in transformer activations can be explained by "latent persistence." This concept posits that abstract latent variables, encoded persistently across layers, govern long stretches of input data. The study found that this persistence grows with depth in LLMs, correlating with the magnitude of autocorrelations and implying that deeper features endure for longer context spans. Experiments involving ablation of these persistent features confirmed their role in influencing autocorrelations. AI
IMPACT Proposes a new theoretical framework for understanding LLM internal dynamics, potentially guiding future model development.
RANK_REASON Research paper published on arXiv detailing a new theoretical concept for LLM internal workings. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →