Reproduce “Written Once, Read Differently How Language Models Own Answers Steer Their Next Ones” 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
Marcus Armstrong
The authors test whether a model’s forced answer to one yes/no question influences its answer to a different question about the same entity. Across 22 models, they find that shared answer directions predict and causally affect this coupling: ablating the direction leaves at most 15% in instruction-tuned models. How later computation reads it varies by model and prompt; tests use simulated binary surveys.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.