When Simple Problems Wear Complex Costumes: Improving Efficiency in LRM's Adaptive Reasoning

Recent Large Reasoning Models (LRMs) have demonstrated powerful multi-step problem-solving capabilities but often suffer from inefficiency due to an ``overthinking phenomenon", where they apply complex reasoning to simple tasks, resulting in unnecessary computational cost and latency. While adaptive reasoning models that can switch between generating explicit reasoning and producing direct answers offer a potential solution, their effectiveness is compromised by a critical flaw: they are often misled by superficial linguistic complexity, mistaking verbosely phrased simple problems for complex ones. To address this, we propose a two-stage training framework to create a more robust adaptive reasoner. The first stage uses supervised fine-tuning with augmented data—presenting simple problems in both concise and redundant forms—to teach the model to ignore superficial verbosity. Subsequently, a reinforcement learning phase utilizes Group Relative Policy Optimization (GRPO) with a custom reward function to refine the model's adaptive policy, ensuring it selects a reasoning mode based on true task complexity rather than surface-level cues. The resulting model reduces computational overhead without sacrificing accuracy and demonstrates improved robustness to misleading linguistic cues.

Modern AI systems designed for complex problem-solving often "overthink" — they apply elaborate, step-by-step reasoning even to simple questions, wasting time and computational resources. Recent approaches have tried to address this by giving AI models the ability to choose between careful reasoning and quick, direct answers. However, we discovered that these models are easily fooled by how a question is worded: if a simple question is phrased in a long or complicated way, the model mistakenly treats it as a hard problem and activates unnecessary reasoning. To solve this, we developed a training method called AdaReasoningSwitch (AdaRS) that teaches AI models to look past surface-level wording and focus on the actual difficulty of a problem. Our approach works in two stages: first, we train the model on examples where simple problems are presented in both concise and wordy forms, so it learns that verbose language does not mean a harder problem; second, we use a reward-based learning process to further sharpen the model's ability to make this distinction. The result is an AI system that reasons carefully only when it truly needs to, responding quickly to easy questions regardless of how they are phrased, and thereby using computational resources more efficiently.