Optional Stopping Theory

Gold definitionUpdated Apr 2, 2026

Definition

Optional Stopping Theory provides conditions under which the expected value of a stochastic process at a random stopping time equals its initial expected value. It is used to design principled rules for terminating a process, such as pruning suboptimal search paths in LLM decoding.

At a glance

Executive summary

Optional Stopping Theory is a mathematical concept that helps determine when to stop a process that unfolds over time, especially when dealing with uncertainty. It provides a rigorous way to make stopping decisions, ensuring that the expected outcome remains consistent. In AI, it's used to efficiently prune less promising search paths in complex tasks like language model reasoning.

TL;DR

It's a math rule that tells you the best time to stop a random process without messing up its average outcome, used in AI to smartly cut off bad search options.

Key points

Provides conditions for stopping a stochastic process at a random time while preserving expected values.
Solves the problem of replacing ad-hoc heuristics with principled, probabilistic rules for terminating sequential processes or pruning search spaces.
Used by researchers in probability theory, sequential analysis, financial mathematics, and increasingly, AI/ML engineers for tasks like LLM decoding.
Offers a theoretically grounded approach to stopping/pruning, unlike heuristic-based methods that lack formal guarantees.
Emerging application in AI for designing robust and efficient search strategies in complex generative models like LLMs.

Use cases

Principled pruning of suboptimal reasoning paths during complex multi-step generation in LLMs, as seen in Martingale Foresight Sampling.
Deciding when to stop a clinical trial early if a drug's efficacy or lack thereof becomes statistically clear, saving time and resources.
Determining the optimal time to exercise an American option or liquidate an asset in financial markets to maximize profit or minimize loss.
Stopping an inspection process in quality control once enough evidence has accumulated to conclude a batch of products meets or fails standards.
Designing stopping criteria for exploration phases or episode termination in reinforcement learning environments with uncertain rewards.

Also known as

Optional Stopping Theorem, Doob's Optional Stopping Theorem