HRBench: A Unified Framework for Evaluating Thinking-Mode Switching in Hybrid-Reasoning LLMs

HRBench has been launched by researchers as a comprehensive evaluation framework aimed at examining thinking-mode switching in hybrid-reasoning large language models (LLMs). These models enable users to manage reasoning effort, balancing the quality of answers with the cost of inference. HRBench categorizes the design space into two dimensions: three families of switching strategies (prompt-based selection, external routing, speculative execution) and four training approaches (training-free, SFT, offline RL, online RL), resulting in 12 distinct evaluation scenarios. The framework assesses these scenarios across six LLMs, ranging from Qwen3.5-2B to Kimi-K2.5-1.1T, along with five reasoning benchmarks in mathematics, science, and coding, while also reimplementing over 12 notable methods. This initiative seeks to standardize comparisons for adaptive thinking-mode selection strategies, which have faced inconsistent evaluations in the past.