LOGIT PRISMS: DECOMPOSING TRANSFORMER OUTPUTS FOR MECHANISTIC INTERPRETABILITY

Summary from the TLDR AI newsletter

The logit lens method, which helps in understanding transformer models' decision-making, has been enhanced by breaking down logit outputs into contributions from various model components. This approach, which uses "prisms" for residual streams, attention layers, and MLP layers, reveals how these parts influence predictions and provides insights into tasks like factual retrieval and arithmetic performed by the gemma-2b model.