Reproduce “Towards Multi-Value Alignment via Value Localization in LLMs” 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

Xueqi Ma, Yanbei Jiang, Xingjun Ma, James Bailey, Sarah M. Erfani

The authors identify attention heads linked to six value and style dimensions, then steer behavior by shifting selected activations at inference without changing model parameters. Masking value-associated heads affects responses more than masking a number of random heads. On Qwen3-4B, multi-value steering improves all three scores over the original model and training baselines, though results vary and harmlessness declines in one setting.

Summary by alphaXiv.

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