Reproduce “Two Routes to Rotation: Structural Sources of Oscillatory Misalignment in RLHF-Shaped Feedback Loops” 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
Atharva Anup Wasnik
The authors analyze a scalar, deterministic RLHF feedback-loop model with separate states for a human standard, reward model, and policy behavior, plus delayed human recalibration. With the reward-model stage retained, complex modes occur when square roots of the three rates satisfy a strict triangle inequality. Simulations show signed disagreement changing sign; this is not evidence about deployed systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.