Stopping Agents: An Intelligent Decision-Making System for Solving Optimal Stopping Problems Using Large Language Models

Stopping Agents is an intelligent decision-making system framework developed collaboratively by researchers from Cornell University, Harvard Business School, and Columbia University. It uses generative large language models to solve optimal stopping problems, and can make real-time judgments in conversational scenarios on whether to continue waiting for more information or terminate the conversation immediately. Its practical application value has been verified in sales conversation scenarios.

The optimal stopping problem is a classic challenge in decision theory. The core contradiction lies in the fact that waiting may yield better decision-making basis, but it comes with time and opportunity costs. It exists widely in real life, such as job seekers choosing offers, investors entering or exiting markets, and salespeople deciding when to end a conversation.

The core innovation lies in combining LLMs with optimal stopping theory, enabling it to understand conversational semantics, dynamically evaluate the value of information, and make real-time decisions. Unlike traditional methods, it does not require preset probability distributions or cost functions. The technical architecture consists of three components: the conversation observation module (encoding conversational semantic representations), the decision generation engine (dynamically outputting wait/quit), and the value evaluation and feedback loop (cost-benefit assessment and learning strategies).

The project provides a demo application for sales conversations. By inputting real-time conversation text streams, it analyzes and advises salespeople whether to continue or end the conversation. The related paper "Learning When to Quit in Sales Conversations" is published on arXiv (arXiv:2511.01181). The research significance includes expanding the boundaries of decision science, providing an implementable solution, and demonstrating the methodology of combining AI with operations research.

Its advantages include strong semantic understanding, good scalability, and solid academic endorsement. Limitations include reliance on LLM APIs (cost issues), need for domain-specific data, and insufficient decision interpretability. This system is a microcosm of AI empowering human decision-making, helping to make wise stop/continue choices in more scenarios.

Stopping Agents represents the direction of AI-assisted decision-making—decision enhancement rather than replacing humans. Such systems will be applied in more fields in the future. For developers, it is an example of combining AI with classic problems; for researchers, it demonstrates the potential of interdisciplinary (computer science + operations research + business) collaboration.