Thinking Deeper Where It Counts: Step-Aware Reasoning Optimization for Cloud-Based LLMs

Chain-of-thought (CoT) prompting has empowered large language models (LLMs) to tackle complex reasoning tasks, yet it often leads to excessive verbosity—even on trivial problems. Prior solutions optimize overall CoT length but overlook the fact that not all reasoning steps are created equal: some are pivotal, others merely filler. We propose the Key Focus Reward Model (KFRM): a fine-grained reinforcement learning framework that teaches models where to think harder and when to say less. Applied to DeepSeekMath-Instruct 7B, KFRM cuts average CoT length while boosting accuracy on GSM8K to $83.15 \%$, outperforming baselines in efficiency and interpretability. These improvements demonstrate that thinking less can mean thinking better—so long as models think deeply where it counts.