Reproduce “Internal Value Alignment in Large Language Models through Controlled Value Vector Activation” 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
Haoran Jin, Meng Li, Xiting Wang, Zhihao Xu, Minlie Huang, Yantao Jia, Defu Lian
A framework named ConVA directly aligns Large Language Models' internal representations with human values, moving beyond behavioral alignment. This approach achieved a 29.6% average relative improvement in value control success rates across 10 basic values while maintaining output fluency above 97% and preserving general model capabilities.
Key insights
- Accurate identification of unbiased internal value representations in LLMs relies critically on carefully curated, context-controlled datasets, which mitigate confounding contextual biases.
- Effective internal value alignment can be achieved without degrading model fluency or general capabilities by dynamically adjusting perturbation strength and gating value activation based on input relevance.
- LLMs exhibit an internal value structure partially consistent with human value theories, yet some opposing human values are encoded similarly, indicating potential for inherent ethical inconsistencies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.