RIFT：重排序指令遵循测试平台评估单步多指令提示结构

Large Language Models (LLMs) are increasingly relied upon for complex workflows, yet their ability to maintain the flow of instructions remains underexplored. Existing benchmarks conflate task complexity with structural ordering, making it difficult to isolate the impact of prompt topology on performance. RIFT, the Reordered Instruction Following Testbed, is introduced to assess instruction following by disentangling structure from content. RIFT systematically investigates the influence of instruction order, inter-instruction dependencies, and instruction complexity on LLM execution by creating various instruction sequence variants through reordering and rephrasing. The testbed designs a series of multi-step instruction scenarios spanning data processing, logical reasoning, and text generation, where the order of instructions can be independently adjusted while preserving their semantic content. Through RIFT, researchers can precisely identify the strengths and limitations of LLMs when handling non-standard instruction sequences, nested instruction structures, or instructions with cyclic dependencies. This platform offers a standardized methodology for quantifying the robustness of different LLM architectures and training strategies in confronting complex instruction structures. Evaluation results from RIFT contribute to guiding further LLM development, enabling them to better comprehend and execute intricate instruction sequences provided by humans—sequences that may not always adhere to strict logical order. This ultimately enhances the reliability and efficiency of LLMs in practical applications.