Strategy-Induct: Task-Level Instruction Generation from Questions Alone
Strategy-Induct is a framework that generates task-level instructions for large language models using only example questions, without requiring labeled answers. It first prompts the model to produce explicit reasoning strategies for each question, forming (strategy, question) pairs, then induces a task instruction from these pairs. Experiments across multiple tasks and model scales show it outperforms state-of-the-art methods in question-only settings. The approach also benefits from jointly using LLMs and large language models.
Key facts
- Strategy-Induct derives task-level instructions solely from example questions.
- No labeled answers are required.
- The framework forms (strategy, question) pairs to induce instructions.
- Outperforms state-of-the-art methods in question-only settings.
- Jointly utilizes LLMs and large language models.
- Tested across multiple tasks and model scales.
- Addresses limitation of prior work relying on input-output pairs.
- Published on arXiv with ID 2605.20924.
Entities
Institutions
- arXiv