HRBench: Benchmarking and Understanding Thinking-Mode Switch Strategies in Hybrid-Reasoning LLMs

HRBench is a new benchmark for studying thinking-mode switching in hybrid-reasoning LLMs.

It evaluates three strategy families: prompt-based selection, external routing, and speculative execution.

The study spans 12 settings, 6 models, and 5 benchmarks, with over 12 prior methods reimplemented.

Results show different strategies win in different efficiency-accuracy trade-offs, depending on training, scale, and task domain.