Large Reasoning Models (LRMs) extend beyond traditional LLMs by incorporating a planning and verification phase before generating responses. Unlike LLMs that predict tokens sequentially based on statistical patterns, LRMs sketch out plans, evaluate options, and double-check calculations internally before outputting answers. This chain-of-thought approach enables better performance on complex tasks like debugging, multi-step math problems, and logical reasoning. LRMs are built by fine-tuning pre-trained LLMs on curated datasets with reasoning examples, then using reinforcement learning (RLHF or process reward models) to optimize logical coherence. The tradeoff is higher computational cost, increased latency, and more expensive inference, making LRMs ideal for complex reasoning tasks but potentially overkill for simple queries.
Sort: