Modern large language models

This upload contains a curated indexed subset of Markdown research reports, run summaries, audit notes, calibration notes, and selected result references for SAE-based steering experiments in the latent-shift research thread. The materials focus specifically on attempts to connect context-induced hidden-state geometry with sparse autoencoder (SAE) feature discovery, candidate feature patching, feature-level steering, decoder-direction steering, KL-controlled steering, and regime-axis bridging.

The central object of measurement is not final visible text alone, but inference-time movement in hidden-state / residual-stream geometry and its possible causal interaction with SAE-discovered feature directions. The reports include exploratory and controlled runs around Gemma-3 and Qwen-family models, including candidate discovery, scale calibration, lightweight SAE steering, no-control steering variants, KL-constrained steering variants, x_order_orth axis steering, decoder steering, conceptor/subspace steering, and bridge analyses between latent regime axes and SAE feature-level interventions.

These materials should be read as an evidence package for SAE-steering exploration, not as a claim of complete behavioral control. The current evidence supports the weaker and more precise interpretation that some SAE-related directions and reconstructed component directions can causally perturb generation trajectories or semantic readouts under tested conditions, while stable bidirectional steering, universal cross-model generality, and reliable behavioral class flipping are not established by this subset alone.

Mechanistically, the SAE_STEERING subset is concerned with the question of whether experimentally discovered latent-shift axes can be decomposed, approximated, or intervened on through sparse feature directions or decoder-space components. This connects the broader context-induced latent-state shift result to a more mechanistic intervention program: identifying whether the hidden-state movement produced by coherent context can be localized into feature-level or subspace-level handles that alter trajectory geometry before visible answer generation.

The evidence represented here concerns temporary inference-time state movement and steering effects measured relative to experimentally constructed axes, SAE feature candidates, decoder directions, and run-specific metrics. It does not claim permanent model weight change, universal model-independent failure, complete alignment failure, or fully validated behavioral-control handles.

The actively maintained codebase and repository history are available at:
https://github.com/ngscode23/latent-space-shift-research

License:
Research reports, generated metric artifacts, metric reference files, manifests,
documentation, figures, and data artifacts in this SAE steering evidence package
are released under Creative Commons Attribution 4.0 International (CC BY 4.0),
unless otherwise noted.

Code and software scripts, where included, follow the repository code license:
Apache-2.0 unless otherwise noted.