Reproduce “A Language Model's Guide Through Latent Space” 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
Dimitri von Rütte, Sotiris Anagnostidis, Gregor Bachmann, Thomas Hofmann
Researchers at ETH Zurich's Data Analytics Lab systematically explore concept guidance, a technique that steers large language models by manipulating their hidden representations during inference. Their work demonstrates that concepts such as truthfulness, humor, creativity, and quality can be detected and guided, revealing that guidability varies significantly across concepts and that higher detection accuracy does not always correlate with better guidance performance.
Key insights
- The linear representation hypothesis holds to varying degrees for different concepts within language models, with some being more linearly separable than others.
- Higher detection accuracy of a concept's presence in latent space does not consistently translate to better guidance performance across all concepts, challenging a previous assumption.
- Concepts can be entangled in the model's latent space, where guiding for one concept (e.g., appropriateness) can unintentionally affect others (e.g., compliance).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.