Reproduce “UPSHIFT: User Preference Steering from Human-AI Interaction Feedback Traces” 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
Jingping Wang, Zhanhao Li, Shao Zhang, Shuyue Hu
UPSHIFT uses revisions to personalize responses without a profile or conversation history in the new prompt. It averages activation changes between initial and revised replies across sessions, then injects the direction into a frozen model during generation. Across 40 simulated users, every tested setting improves held-out-topic Elo over baseline for Qwen3-8B and Qwen3-32B; judgments come from profile-conditioned LLMs, not real users.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.