meituan-longcat/R-HORIZON-Websearch
收藏资源简介:
R-HORIZON 是一个旨在激发大型推理模型(LRM)中长跨度推理行为的新方法,通过查询组合将孤立的问题转化为复杂的多步骤推理场景。它揭示了即使在面对跨长推理跨度的相互依赖问题时,最先进的 LRM 也会遭受显著的性能下降。R-HORIZON 基准包括来自数学、代码生成和代理应用领域的 6 个代表性数据集,并通过使用长跨度推理数据使强化学习具有验证奖励 (RLVR)。该数据集的结构包括输入问题、实例 ID、原始实例 ID、目标答案、问题数量以及所选变量等信息。
R-HORIZON is a novel method designed to stimulate long-horizon reasoning behaviors in Large Reasoning Models (LRMs) through query composition. It transforms isolated problems into complex multi-step reasoning scenarios, revealing that even the most advanced LRMs suffer significant performance degradation when facing interdependent problems that span long reasoning horizons. The R-HORIZON Benchmark comprises 6 representative datasets from mathematics, code generation, and agent applications, and enables reinforcement learning with verified rewards (RLVR) using long-horizon reasoning data. The dataset structure includes input, instance ID, origin instance IDs, target, number of problems, and selected variables.



