When Can Large Reasoning Models Save Thinking? Mechanistic Analysis of Behavioral Divergence in Reasoning
This work addresses inefficiencies in large reasoning models for AI practitioners, but it is incremental as it analyzes existing mechanisms without proposing new methods.
The study investigated how reinforcement learning-trained large reasoning models behave when prompted to save thinking, identifying three distinct thinking modes and showing that no thinking reduces output length but sacrifices accuracy, while explicit and implicit thinking maintain accuracy with reduced length.
Large reasoning models (LRMs) have significantly advanced performance on complex tasks, yet their tendency to overthink introduces inefficiencies. This study investigates the internal mechanisms of reinforcement learning (RL)-trained LRMs when prompted to save thinking, revealing three distinct thinking modes: no thinking (NT), explicit thinking (ET), and implicit thinking (IT). Through comprehensive analysis of confidence in thinking termination, attention from thinking to generation, and attentional focus on input sections, we uncover key factors influencing the reasoning behaviors. We further find that NT reduces output length at the cost of accuracy, while ET and IT maintain accuracy with reduced response length. Our findings expose fundamental inconsistencies in RL-optimized LRMs, necessitating adaptive improvements for reliable efficiency.