Ask, Condition or Abstain: Reinforcement Learning for Missing-Premise Reasoning
2026-08-17 • Computation and Language
Computation and Language
AI summaryⓘ
The authors study how AI models can handle questions that don’t include all the necessary information to give a clear answer. They created a new training method called Ask-Condition-Abstain Reinforcement Learning (ACA-RL), which teaches models to ask for missing information, guess carefully based on what’s unknown, or say they can’t answer if there’s no helpful reply. They also made a new test set called the Missing-Premise Benchmark (MPB) with different types of problems that need extra info. Their approach improves AI models’ ability to deal with uncertain questions while still doing well on normal, fully detailed problems.
Reinforcement LearningMissing PremiseReasoning ModelsStructured RewardQwen3LLaMABenchmarkUncertainty HandlingNatural Language ProcessingTask Underdetermination
Authors
Yongqi Tong, Zhenyu Zhang, Zimi Liu, Kewei Fu, Mingli Song, Haofei Zhang, Junshao Zhang, Hong Zhu, Jiang-Ming Yang, Xin Zhang, Jianshe Li
Abstract
Answer-only reinforcement learning (RL) trains reasoning models to solve fully specified problems, but many realistic queries omit a premise needed for a unique answer. In this setting, the useful response is not always refusal: the model should ask for the missing premise, condition its answer on the unknown quantity, or abstain when no informative conditional response is available. We present \emph{Ask-Condition-Abstain Reinforcement Learning} (ACA-RL), a data-augmented RL framework for this setting. Its reasoning-graph-guided pipeline converts well-posed problems into missing-premise training instances with localized gap annotations; ACA-RL then trains on these instances with a structured reward over five observable response behaviors. We also introduce the \emph{Missing-Premise Benchmark} (MPB), a 274-instance human-verified benchmark spanning mathematical, logical, and real-world word problems. Across Qwen3 and Llama models, ACA-RL consistently improves on MPB while preserving competitive performance on well-posed reasoning tasks. Together with the released code, MPB, and training data, this work supports a new mission for NLP evaluation: measuring whether models can recognize when a task is underdetermined and handle uncertainty, not only whether they can answer fully specified questions.