Reproduce “Retraining Seeks Stable Signals” 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

Moritz Hardt

This research introduces the "stable signal principle" to analyze retraining dynamics in performative prediction, demonstrating that a model-independent component in the target can guide robust convergence even under strong self-influence. The work shows that regularization or normalization can stabilize learning processes, offering theoretical explanations for the stability of training systems, including language models using self-generated data.

Key insights

  • The presence of even a small, model-independent "stable signal" ensures robust convergence of retraining dynamics to the direction of this signal, irrespective of the model's self-influence strength.
  • Regularization serves a novel function in performative systems by explicitly controlling feedback strength, making otherwise non-contractive retraining processes stable and convergent.
  • The key parameters of performative feedback and the stable signal component can be identified from just two consecutive retraining steps, allowing direct computation of the unregularized performatively stable point.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.