Independent research
Mechanistic interpretability, sparse representations, activation interventions, and the geometry of language-model internals.
Open source: vollib.org
Email: larry@source1.jp
Paper pages include abstracts, citation details, and source documents.
A confirmatory study measuring holonomy in Gemma 2 2B. Active SAE feature planes carried less measured holonomy than matched mixed-feature controls in the preregistered setup.
Abstract and linksA worked Gemma Scope case study using a fully crossed factorial design to measure how prompt factors and transformer layer reorganize SAE-derived feature responses.
Abstract and PDFA boundary study separating answer-token control from portable factual transfer across optimized payloads, donor activation transplants, and aligned patching.
Abstract and PDFA two-dimensional linear interface that exposes analogy, classification, and approximately orthogonal semantic directions in frozen Qwen hidden states.
Abstract and PDF