Reproduce “All Code, No Thought: Current Language Models Struggle to Reason in Ciphered Language” 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
Shiyuan Guo, Henry Sleight, Fabien Roger
Current large language models demonstrate a substantial drop in accuracy when performing complex reasoning tasks in ciphered language, even while often retaining the ability to translate such ciphers back into plain English. This asymmetry suggests that ciphered reasoning is currently an ineffective tactic for evading Chain-of-Thought monitoring for computational tasks.
Key insights
- A fundamental asymmetry exists where LLMs can fluently translate ciphered text but significantly struggle to perform complex reasoning *within* that ciphered language.
- A model's ability to reason in a specific cipher strongly correlates with the cipher's prevalence in its pretraining data, especially for structure-preserving ciphers.
- Improving ciphered reasoning capability scales slowly and expensively with additional fine-tuning data and parameters, contrasting with the much faster saturation of cipher translation ability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.