跳转至

文章背景与核心概要

随着具备思考能力的现代大语言模型和视觉语言模型（LLM/VLM）的普及，模型在输出最终答案前会生成复杂的推理轨迹。然而，这些看似深思熟虑的推理过程是否真正与答案的正确性相关？本文探讨了思考模型中推理行为与正确性之间的复杂关系，揭示了当前推理导向训练中的深层缺陷。

研究团队引入了“行为提升”（Behavioral Lift）这一新指标，用来衡量当推理轨迹中存在或缺少某种特定行为时，答案正确性的变化情况。通过对 15 个模型、6 个基准测试中的 15,282 条带注释的推理轨迹进行分析，论文发现了著名的“放大-提升鸿沟”（Amplification-Lift Gap）：虽然推理导向的训练大幅放大了自我纠错、假设检验和不确定性确认等特征（达 3-7 倍），但这些行为与正确性的关联往往较弱甚至呈负相关。相反，真正高提升的行为（如置信度校准、知识对齐和自我意识）却几乎没有被放大。

这一发现表明，当前的推理导向训练往往更倾向于优化表层的“审慎外观”，而非真正具有预测性的推理行为。该研究为未来开发能奖励校准与扎实推理的流程级优化目标提供了重要的理论动机。

Amplified Does Not Mean Predictive: Reasoning Behaviors in Thinking Models

arXiv ID: arXiv:2608.13760 [cs.CL]
Published: August 13, 2026 (Published in COLM 2026)
Authors: Jean de Dieu Nyandwi, Leena Mathur, Yonatan Bisk, Robert Hawkins, Graham Neubig

📌 Summary

本文研究了现代思考模型（包括纯文本大语言模型和视觉语言模型）中推理行为与正确性之间的关系。作者引入了行为提升（Behavioral Lift）指标，用于衡量当推理轨迹中存在特定行为与不存在该行为时，正确性发生的变化。

通过分析跨越 15 个模型和 6 个基准测试的 15,282 个带注释的轨迹，研究揭示了放大-提升鸿沟（Amplification-Lift Gap）：尽管面向推理的训练大幅放大了自我纠错、假设检验和不确定性承认等特征（3-7倍），但这些行为通常与正确性的关联较弱或呈负相关。相反，提升度最高的行为——如置信度校准、知识对齐和自我意识——却几乎没有被放大。这些研究结果表明，当前的面向推理的训练优化的只是表层的审慎外观，而非真正具预测性的推理行为。

This paper investigates the relationship between reasoning behaviors and correctness in modern thinking models (both text-only Large Language Models and Vision-Language Models). The authors introduce Behavioral Lift—a metric measuring how correctness changes when a specific behavior is present versus absent in a reasoning trace.

Analyzing 15,282 annotated traces across 15 models and 6 benchmarks, the study reveals the Amplification-Lift Gap: while reasoning-oriented training heavily amplifies traits like self-correction, hypothesis testing, and uncertainty acknowledgment (by 3–7\(\times\)), these behaviors are often weakly or negatively associated with correctness. Conversely, the highest-lift behaviors—such as confidence calibration, knowledge alignment, and self-awareness—are barely amplified. The findings suggest that current reasoning-oriented training optimizes for surface-level deliberative appearance rather than genuinely predictive reasoning behaviors.

🧭 Metadata & References

📄 Abstract

推理模型中哪些推理行为与正确答案相关？面向推理的训练是否放大了这些行为？这一区分至关重要，因为面向推理的训练可能会让推理轨迹看起来更具深思熟虑，却并未放大与模型正确性最紧密相关的行为。

我们通过行为提升（Behavioral Lift）来量化这种不匹配。该指标用于衡量当模型推理轨迹中存在某行为与不存在该行为时，正确性的变化程度。在涵盖文本和视觉语言推理的 15 个模型和 6 个基准测试中，我们使用一套同时适用于 LLM 和 VLM 轨迹的核心行为分类法，对 15,282 条轨迹进行了标注。

我们发现了放大-提升鸿沟（Amplification-Lift Gap）的证据：思考模型极大地放大了自我纠错、假设检验和不确定性承认，而提升度最高的行为其实是置信度校准、知识对齐和自我意识。

  • 置信度校准是两种模态中正确性的最强正向信号之一，但几乎没有被放大。
  • 不确定性承认被放大了 3-7 倍，但与正确性的关联却很弱或呈负相关。

最终，我们发现面向推理的训练并没有优先放大高提升（High-Lift）行为，这促使我们需要采用过程级的目标，来奖励经过校准和扎实的推理，而不仅仅是表层形式。

Which reasoning behaviors are associated with correct answers in reasoning models, and does reasoning-oriented training amplify those behaviors? This distinction is important because reasoning-oriented training can make traces look more deliberative without amplifying the behaviors most tied to model correctness.

We quantify this mismatch with Behavioral Lift, a metric that measures how much correctness changes when a behavior is present versus absent in a model's reasoning trace. Across 15 models and 6 benchmarks spanning text-only and vision-language reasoning, we annotate 15,282 traces with a taxonomy whose core behaviors are defined for both LLM and VLM traces.

We find evidence for an Amplification-Lift Gap, in which thinking models strongly amplify self-correction, hypothesis testing, and uncertainty acknowledgment, while the highest-lift behaviors are confidence calibration, knowledge alignment, and self-awareness.

  • Confidence calibration is among the strongest positive signals of correctness in both modalities, yet is barely amplified.
  • Uncertainty acknowledgment is amplified by 3–7\(\times\), yet is weakly or negatively associated with correctness.

Ultimately, we find that reasoning-oriented training does not preferentially amplify the highest-Lift behaviors, motivating process-level objectives that reward calibrated and grounded reasoning rather than surface form alone.