Results for “interpretability”

Analyzes evaluation results by requiring a stated hypothesis before examining data, then compares expectations to actual result files to prevent post-hoc rationalization.

Audits hidden assumptions in decisions or beliefs by classifying them into fact, convention, belief, or interest-driven, ranking by fragility and impact, then rebuilding conclusions from verified premises. Supports English and Chinese.

Train and analyze Sparse Autoencoders (SAEs) using SAELens to decompose neural network activations into interpretable features for mechanistic interpretability research.