Reproduce “SteerDistill: Quantizing LLM Agents via Closed-Loop Teacher Steering” 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
Genming Cui, Xiangyu Li, Zixu Hao, Xin Ding, Yunxin Liu, Ting Cao
SteerDistill trains quantized agents to correct failures using a full-precision teacher, then distills verified corrections for use without guidance. It aligns failed rollouts with references, guides student revisions at divergences, and locally checks transitions. Under W4A16, Qwen3-8B averages 53.18 across five benchmarks, versus QeRL’s 51.00 and BF16’s 53.54; this does not establish recovery in all settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.