OmniLens Interprets LLM Hidden States at Scale.

A new interpretation method called OmniLens has been developed to analyze the hidden states of large language models (LLMs) at unprecedented scale. Traditional "lens" methods, which map intermediate activations to output vocabulary, have been computationally expensive, limiting their application to smaller models and specific components. OmniLens overcomes these limitations through two key scaling techniques. First, it employs low-rank translators, which reduce the number of trainable parameters per lens by up to 98.4%, making parameter growth linear with model width. Second, its Subset-KL training method materializes only selected vocabulary logits, cutting peak training memory by up to 70% or maintaining unbiased gradients for the full Kullback-Leibler divergence. These efficiencies allowed researchers to create a dense ensemble of 482 lenses for a 70B parameter model, providing six times the coverage of previous designs. This comprehensive analysis revealed that the most visible components for a specific behavior are not necessarily the most effective points for intervention, and that optimal intervention points often lie outside previously studied attention heads.