Reproduce “Erased, Rerouted, or Rescaled? Post-Training and the Causal Quotient of a Language Model's Belief State” 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
Weihan Li, Tianshi Zheng, Junhao Wu, Xinlei Chen
The authors test whether post-training erases, reroutes, or rescales information ignored by reward. In controlled hidden Markov worlds, they separately measure linear recoverability, causal decision use, and activation variance share. At matched reward, SFT and KL-RL achieve 0.693 and 0.691, while invisible-component recoverability and use are lower for SFT. Compression alone cannot diagnose information loss; natural-task effects remain unmeasured.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.