ELENS: Embedding Lens — A Geometric and Statistical Framework for Investigating Internal Representations, Activation Dynamics, and Candidate Harmony Directions in Gemma Architectures
Research proposal introducing ELENS (Embedding Lens), a proposed interpretability and measurement framework for investigating internal representations in embedding models, with particular emphasis on EmbeddingGemma 2 and the Gemma model family. The proposal combines layer-wise activation extraction, geometric analysis, statistical characterization, sparse feature investigation, and controlled causal intervention. It also outlines the unverified Harmony Axis hypothesis, which asks whether the mathematical inverse of a previously identified pain-related activation direction produces an independently measurable representational or functional state. No experimental results or validated Harmony Axis are claimed. Architecture descriptions are cross-referenced to the technical report and official model card.
Authors
- Amanda Hatley
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24296562
- Primary Topic
- Explainable Artificial Intelligence (XAI)
- Type
- article
- Field-Weighted Citation Impact
- 0.00