Martingale Foresight Sampling

Martingale Foresight Sampling (MFS) is an advanced decoding strategy for Large Language Models (LLMs) designed to address the limitations of standard autoregressive generation, which often struggles to identify globally optimal reasoning paths due to its token-by-token, greedy nature. MFS reframes the LLM decoding problem as the identification of an optimal stochastic process, where the quality of a reasoning path evolves over time. By applying principles from Martingale theory, MFS replaces the ad-hoc heuristics common in other foresight sampling methods with a rigorous probabilistic framework. This allows for principled valuation of future steps, intelligent pruning of suboptimal search branches, and adaptive termination of exploration. This approach is crucial for researchers and ML engineers developing LLMs for complex reasoning tasks, planning, and multi-step problem-solving, where global coherence and optimality are paramount.