Ask, Condition or Abstain: Reinforcement Learning for Missing-Premise Reasoning

TL;DR - ACA-RL trains reasoning models to handle underspecified questions by asking for missing information, giving conditional answers, or abstaining. It improves missing-premise reasoning while retaining competitive performance on well-posed tasks.

• Generates training examples by removing premises from well-posed problems and annotating the resulting gaps.
• Uses structured rewards covering five observable response behaviors.
• Introduces MPB, a 274-instance, human-verified benchmark spanning mathematical, logical, and real-world problems.
• Consistently improves Qwen3 and Llama models on MPB; code, benchmark, and training data are released.