AIML-TUDA/SLR-Bench
收藏资源简介:
SLR-Bench是一个可扩展的、全自动的基准测试,旨在通过归纳逻辑编程(ILP)任务系统地评估和训练大型语言模型(LLM)的逻辑推理能力。它提供了一个包含20个复杂度级别的课程,分为4个广泛的难度等级(基础、简单、中等、困难)。每个任务都包含一个自然语言提示、一个可执行的验证程序和一个潜在的地面真实规则。SLR-Bench可以用于评估和训练各种LLM,包括GPT-4o、Llama-3、Gemini和DeepSeek-R1。
SLR-Bench is a scalable, fully-automated benchmark designed to systematically evaluate and train Large Language Models (LLMs) in logical reasoning via inductive logic programming (ILP) tasks. It presents LLMs with open-ended logic problems of progressively increasing difficulty, assesses their solutions via deterministic symbolic evaluation, and supports both curriculum learning and systematic measurement of reasoning performance. The benchmark consists of over 19,000 tasks, each with a natural language prompt, an executable validation program, and a latent ground-truth rule. It is licensed under CC BY 4.0.




