Reproduce “À la recherche du sens perdu: your favourite LLM might have more to say than you can understand” 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
K. O. T. Erziev
A comprehensive investigation reveals that leading language models can understand instructions encoded in visually incomprehensible Unicode sequences, with some models achieving 40% success rates in bypassing content filters through these hidden communication channels, raising significant concerns about current safety mechanisms.
Key insights
- LLMs can consistently understand instructions encoded in visually incomprehensible characters
- Capability varies significantly between models, with newer architectures showing stronger abilities
- Phenomenon extends beyond simple tokenization effects
- Current safety measures may have fundamental limitations in addressing these vulnerabilities
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.