Reproduce “Oedipus: LLM-enchanced Reasoning CAPTCHA Solver” 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
Gelei Deng, Haoran Ou, Yi Liu, Jie Zhang, Tianwei Zhang, Yang Liu
Researchers from Nanyang Technological University introduced OEDIPUS, a framework that utilizes Large Language Models (LLMs) to solve complex reasoning CAPTCHAs by breaking down tasks using a Domain Specific Language. This approach achieved an average success rate of 63.5% on commercially available CAPTCHAs from 2022 and 44.1% on newer challenges from 2023, indicating an advancement in automated CAPTCHA solving capabilities.
Key insights
- Direct application of state-of-the-art multimodal LLMs is insufficient for reliably solving complex reasoning CAPTCHAs due to limitations in object recognition, multi-step reasoning, and task continuity.
- Structuring LLM-based CAPTCHA solving through a Domain Specific Language (DSL) and breaking down complex tasks into manageable sub-steps significantly improves performance.
- The ongoing development of LLM-enhanced solvers necessitates continued innovation in CAPTCHA design to maintain their effectiveness against automated systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.