Don’t Answer Bench (DNA Bench)
收藏资源简介:
DNA Bench是一个由ServiceNow创建的新型基准数据集,旨在评估大型语言模型在理解复杂推理触发因素并避免不必要推理方面的能力。该数据集包含150个对抗性设计的提示,对人类来说简单易懂,但对许多最近突出的推理型LLM来说却难以应对。数据集分为五个类别,每个类别针对不同的挑战,反映真实世界的失败模式。数据集通过手动设计和使用OpenAI GPT-4o生成提示的方式创建,旨在揭示LLM在处理具有误导性的简单问题时过度推理的倾向。
DNA Bench is a novel benchmark dataset developed by ServiceNow, designed to evaluate the capabilities of large language models (LLMs) in comprehending complex reasoning triggers and avoiding unnecessary reasoning. This dataset comprises 150 adversarially crafted prompts that are straightforward and comprehensible to humans, yet pose significant challenges to many prominent recent reasoning-oriented LLMs. The dataset is divided into five categories, each targeting distinct challenges that mirror real-world failure patterns. Constructed through both manual curation and prompt generation utilizing OpenAI GPT-4o, this dataset aims to expose the propensity of LLMs to engage in over-reasoning when presented with deceptively simple problems.

- 1DNA Bench: When Silence is Smarter -- Benchmarking Over-Reasoning in Reasoning LLMsServiceNow · 2025年



