GPT-5.4 quirk: one diacritic changes output rate by 47%
rayanpal_ · reddit · 2026-08-25
A study on GPT-5.4 found that adding or removing diacritics on specific Hebrew and Arabic characters in the system prompt drastically changes compliance rates, jumping from 47.3% to 94.3%. This phenomenon reveals extreme sensitivity to specific character features, potentially linked to internal tokenization or safety mechanisms.
Related event: One Diacritic Doubles GPT Instruction-Following Rate(3 posts)→
More from Models
- Reasonable post-trains Nemotron 3.5 on 3,000 Verus proofs, lifting pass@1 ~30 points to 73.3% — fhuszar · 2026-10-16
- NVIDIA's Nemotron 3.5 Super VL hits CoreWeave: 12B-active multimodal MoE at $0.25/M input tokens — wandb · 2026-10-16
- Opus 5 might be a very deep model or use looping, commentator speculates — scaling01 · 2026-10-16
- GPT-6.1 Sol audit finds 6% of arXiv papers had false main theorems, down to 1.5% recently — nitarshan · 2026-10-16
- Nvidia-backed post-training pushes Nemotron 3.5 VL past frontier on dermatology diagnosis — LoganGrasby · 2026-10-16
- Gemini 4 Argon scores 97.5% on ARC-AGI-3, second lab above 95% — patience_cave · 2026-10-16