Summary of Reasoning Robustness Of Llms to Adversarial Typographical Errors, by Esther Gan et al.
Reasoning Robustness of LLMs to Adversarial Typographical Errors
by Esther Gan, Yiran Zhao, Liying Cheng, Yancan Mao, Anirudh Goyal, Kenji Kawaguchi, Min-Yen Kan, Michael Shieh
First submitted to arxiv on: 8 Nov 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper investigates the robustness of Large Language Models (LLMs) to typographical errors in Chain-of-Thought (CoT) prompting. The authors develop an Adversarial Typo Attack algorithm that iteratively introduces typos to words important to the query, showing LLMs are sensitive to minimal changes. Notably, a single character edit reduces Mistral-7B-Instruct’s accuracy by 5.1% on GSM8K. To evaluate larger, closed-source models, the authors create the R2ATA benchmark, which applies the algorithm to open-source LLMs and demonstrates transferability, causing notable performance drops across multiple super large and closed-source LLMs. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large Language Models can make mistakes when given bad typing. Researchers tested how well these models could handle errors by making small changes to words that are important to understanding a question. They found that even a single mistake can greatly reduce the model’s accuracy. To see if this would happen with other, more powerful models, they created a benchmark that tests their ability to handle mistakes. |
Keywords
» Artificial intelligence » Prompting » Transferability
