notesum.ai
Published at November 8Reasoning Robustness of LLMs to Adversarial Typographical Errors
cs.CL
cs.AI
Released Date: November 8, 2024
Authors: Esther Gan1, Yiran Zhao1, Liying Cheng2, Yancan Mao1, Anirudh Goyal3, Kenji Kawaguchi1, Min-Yen Kan1, Michael Shieh1
Aff.: 1National University of Singapore; 2Singapore University of Technology and Design; 3Google DeepMind

| Dataset | Model (#Params) | Ori. | Avg-ATA | ATA-1 | ATA-2 | ATA-4 | ATA-8 |
|---|---|---|---|---|---|---|---|
| GSM8K | Gemma-2B (2.5B) | () | |||||
| Llama2-7B (6.7B) | () | ||||||
| Mistral-7B (7.2B) | () | ||||||
| Gemma-7B (8.5B) | () | ||||||
| BBH | Gemma-2B (2.5B) | () | |||||
| Llama2-7B (6.7B) | () | ||||||
| Mistral-7B (7.2B) | () | ||||||
| Gemma-7B (8.5B) | () | ||||||
| MMLU | Gemma-2B (2.5B) | () | |||||
| Llama2-7B (6.7B) | () | ||||||
| Mistral-7B (7.2B) | () | ||||||
| Gemma-7B (8.5B) | () |