Reproduce “Steering Vectors as a Training Signal in LLM Post-Training” 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

Tiejin Chen, Maunil Vyas, Huaiyuan Yao, Hua Wei

The authors introduce activation-direction methods for supervised distillation and reinforcement learning. Per-sample Slerp steers training toward reference answers, while SV-Rescue ranks wrong rollouts when group-relative rewards provide no signal. On MATH, Qwen2.5-1.5B accuracy rises from 18.13% ± 2.30 to 33.47% ± 1.14 with a Qwen3-8B thinking teacher; reference-based steering can fail on code.

Summary by alphaXiv.

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