Reproduce “VERS: Steering Reasoning Models by Editing Their Activation Explanations” 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
Ihor Vitenko
VERS reduces generated tokens by turning edits to text explanations of internal states into a fixed steering direction applied at paragraph breaks during reasoning. On 1,319 GSM8K test questions, it cuts mean generated tokens 45.8%, with accuracy changing by −0.82 percentage points [−1.42, −0.21], meeting the −2-point non-inferiority margin. On MATH-500, it saves 31.3% but accuracy falls 3.05 points.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.