Reproduce “Stored in Optimizer State, Valued by Later Training: A Causal Account of Subliminal Trait Transfer” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qinyang Xu
This research offers a causal explanation for subliminal trait transfer in deep learning, demonstrating that optimizer states physically transport subtle source perturbations while subsequent training dynamically determines the behavioral value of these stored influences. The framework precisely predicts how varying future training paths assign different behavioral outcomes to the same underlying ancestry.
Key insights
- Optimizer moments (e.g., first moment in AdamW, velocity in momentum SGD) serve as primary, delayed-release carriers for subliminal source perturbations, consistently transporting these influences.
- The behavioral value (sign and magnitude) of stored ancestry is not an intrinsic property but is dynamically assigned by later training, separating physical transport from behavioral manifestation.
- The full-horizon costate predictor derived from adjoint sensitivity analysis accurately anticipates route-dependent behavioral outcomes without requiring fitted parameters, providing a strong diagnostic tool.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.