The State of LLM Reasoning Model Inference

Comprehensive technical survey of LLM reasoning model advances since DeepSeek R1, focused on inference-time compute scaling methods. Covers chain-of-thought prompting, majority voting, beam search, and the s1 paper's "budget forcing" via "Wait" tokens — a technique where appending special tokens causes models to self-verify and extend reasoning before finalizing answers. Provides a useful taxonomy distinguishing inference-time scaling (no weight changes) from training-time approaches like RL and distillation, with comparisons across all four categories.