ARQ Framework Improves LLM Reasoning with Generated Stepping Stones
A new framework called ARQ (Asking the Right Questions) enhances large language models' reasoning by generating intermediate stepping stones such as simplifications, alternative framings, or subproblems. Research shows these questions are transferable and significantly aid LLMs of varying capabilities in solving complex tasks like math and coding. The approach frames stepping stone generation as a post-training task, demonstrating that good questions can be systematically produced to improve reasoning performance.
Key facts
- ARQ introduces a question generator to the default reasoning pipeline.
- Stepping stones include simplifications, alternative framings, or subproblems.
- Good stepping stone questions are transferable across LLMs.
- The framework improves reasoning in math and coding tasks.
- Stepping stone generation is framed as a post-training task.
- Research is published on arXiv with ID 2602.19069.
Entities
Institutions
- arXiv