New Framework Tells LLMs When to Stop Multi-Turn Reasoning Tasks

When language models tackle complex questions through multi-turn reasoning—such as searching for information multiple times or taking sequential actions—they face a thorny problem: how do they know when to stop? Quitting too early risks a wrong answer, but continuing unnecessarily wastes compute and time.

Researchers have now developed MiCP (Multi-Turn Language Models with Conformal Prediction), a framework that gives LLMs a principled way to decide when to halt multi-turn reasoning tasks while maintaining formal accuracy guarantees.

The Challenge of Knowing When to Stop#

Today's multi-turn LLM systems—including adaptive retrieval-augmented generation (RAG) and ReAct-style agents—iterate through cycles of information retrieval, reasoning, or action-taking to solve difficult questions. Most existing systems rely on ad-hoc stopping rules (like "stop after three tries") or fixed turn budgets, neither of which guarantees the final answer is actually correct.

This uncertainty becomes costly in high-stakes fields such as healthcare and finance, where incorrect decisions due to premature stopping could have serious consequences, yet unnecessary reasoning cycles add latency and expense.

MiCP's Approach#

The new framework adapts conformal prediction (CP)—a statistical technique that provides formal coverage guarantees—to handle the complexities of multi-turn pipelines. Rather than trying to guarantee accuracy in a single pass, MiCP distributes error budgets across each turn of reasoning. This allows the model to stop as soon as it has sufficient confidence that its answer is correct, while still maintaining an overall coverage promise.

When tested on single-hop and multi-hop question-answering benchmarks using both adaptive RAG and ReAct agents, MiCP achieved its target coverage guarantees while reducing the number of turns, computational cost, and the size of prediction sets.

Why It Matters#

The framework introduces a new evaluation metric that jointly measures both whether coverage guarantees hold and how efficiently the system answers questions. By balancing these two dimensions, MiCP addresses a real gap in deployed LLM systems: the need for provable correctness guarantees without sacrificing practical efficiency.