跳转至

文章背景与核心概要

随着大语言模型（LLMs）的广泛应用，诸如自适应检索增强生成（RAG）和ReAct风格的智能体等多轮推理机制被越来越多地用于处理复杂查询。然而，这些系统面临着一个核心困境：如何确定最佳的停止时机。当前的方法大多依赖启发式规则或固定预算，缺乏形式化保证，往往导致过高的延迟和成本，或者过早地产出不准确的结论。

本文引入了 MiCP（基于共形预测的多轮大语言模型），这是首个将共形预测（Conformal Prediction）应用于多轮推理的框架。通过在不同轮次之间动态分配错误预算，MiCP 使得模型能够在保持严格、形式化覆盖保证的同时实现提前停止。在多跳和单跳问答基准上的实验结果表明，MiCP 在确保预测可靠性的同时，有效降低了推理成本和轮次消耗。

Adaptive Stopping for Multi-Turn LLM Reasoning

Authors: Xiaofan Zhou, Huy Nguyen, Bo Yu, Chenxi Liu, Lu Cheng
arXiv ID: 2604.01413
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Submission Date: April 1, 2026 (Last revised: August 14, 2026)

Adaptive Stopping for Multi-Turn LLM Reasoning

Authors: Xiaofan Zhou, Huy Nguyen, Bo Yu, Chenxi Liu, Lu Cheng
arXiv ID: 2604.01413
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Submission Date: April 1, 2026 (Last revised: August 14, 2026)

Summary

大语言模型（LLMs）越来越多地利用多轮推理（例如自适应检索增强生成（RAG）和ReAct风格的智能体）来处理复杂查询。然而，这些系统面临着一个关键困境：确定最佳停止点。当前的方法依赖于启发式规则或固定预算，缺乏形式化保证，往往导致过高的延迟/成本或过早得出不准确的结论。

本文推出了 MiCP（基于共形预测的多轮大语言模型，Multi-Turn Language Models with Conformal Prediction），这是首个将共形预测应用于多轮推理的框架。通过在不同轮次之间动态分配错误预算，MiCP 使得模型能够在维持严谨、形式化覆盖保证的同时进行早期停止。在多跳和单跳问答基准测试上的实验结果证明，MiCP 在确保预测可靠性的同时，有效地降低了推理成本和轮次计数。

Summary

Large Language Models (LLMs) increasingly utilize multi-turn reasoning—such as adaptive Retrieval-Augmented Generation (RAG) and ReAct-style agents—to tackle complex queries. However, these systems face a critical dilemma: determining the optimal point to stop. Current methods rely on heuristic rules or fixed budgets, which lack formal guarantees and often result in either excessive latency/cost or premature, inaccurate conclusions.

This paper introduces MiCP (Multi-Turn Language Models with Conformal Prediction), the first framework to apply Conformal Prediction to multi-turn reasoning. By dynamically allocating error budgets across different turns, MiCP enables models to stop early while maintaining rigorous, formal coverage guarantees. Experimental results on multi-hop and single-hop QA benchmarks demonstrate that MiCP effectively reduces inference costs and turn counts while ensuring prediction reliability.

Key Contributions

  • 形式化覆盖保证： 与启发式方法不同，MiCP 提供了统计学保证，确保最终预测包含正确答案。
  • 自适应错误预算： 该框架在推理轮次智能分配错误预算，兼顾效率和准确性进行优化。
  • 效率指标： 作者引入了一种新的评估指标，能够同时评估覆盖的有效性与推理过程的效率。
  • 广泛适用性： 证明了其在自适应 RAG 和 ReAct 风格智能体架构中的有效性。

Key Contributions

  • Formal Coverage Guarantees: Unlike heuristic approaches, MiCP provides statistical guarantees that the final prediction contains the correct answer.
  • Adaptive Error Budgeting: The framework intelligently distributes error budgets across reasoning turns, optimizing for both efficiency and accuracy.
  • Efficiency Metrics: The authors introduce a novel evaluation metric that jointly assesses the validity of coverage and the efficiency of the reasoning process.
  • Broad Applicability: Demonstrated effectiveness in both adaptive RAG and ReAct-style agent architectures.

Access & Resources

license icon

Access & Resources

license icon

Submission History

  • v1： 2026年4月1日
  • v2： 2026年4月5日
  • v3： 2026年8月14日（当前版本）

Submission History

  • v1: 1 Apr 2026
  • v2: 5 Apr 2026
  • v3: 14 Aug 2026 (Current)