Implement an entropy-guided decoding strategy that optimizes LLM reasoning by monitoring token probabilities. This provides a low-overhead alternative to computationally expensive self-consistency chains.