One Diacritic Shifts LLM Cross-Script Output from 47% to 94%
rayanpal_ · reddit · 2026-08-25
A study reveals that adding a single diacritic to a Hebrew word in the system prompt shifts the model's output accuracy on a specific task from 47% to 94% when processing Arabic input. Using a frozen protocol, the experiment compared "dotted" vs. "undotted" Hebrew conditions, resulting in a 47 percentage-point difference in exact artifacts generated. The author provides the full prompts, a published paper (via Zenodo), and a GitHub repository for reproducibility.
Related event: One Diacritic Doubles GPT Instruction-Following Rate(3 posts)→
More from Research
- Goodfire's Nick Cammarata: neural networks are "already clean" in a possible interpretability breakthrough — tszzl · 2026-10-15
- tszzl: interpretability breakthrough feels like my understanding of neural nets jumped five years — tszzl · 2026-10-15
- LightSuite: fast cell detection and whole-brain atlas registration on a standard workstation — hisspikeness · 2026-10-15
- Pathway CSO pitches BDH as post-Transformer architecture uniting memory, reasoning, adaptation — schwentker · 2026-10-15
- AI-designed molecular scissors cut Alzheimer's proteins — but is it really AI? — gerardsans · 2026-10-15
- EmbodiedEvo tops RoboDojo at 38.95% success by letting robots inherit failed practice — thetripathi58 · 2026-10-15