跳转至

文章背景与核心概要

本文探讨了大语言模型（LLMs）在评估序列证据时所表现出的位置偏差（如首因效应和近因效应），并将其与人类的判断模式进行了对比。研究发现，LLMs 在证据呈现过程中进行信念更新的时机与人类行为存在显著差异。

研究的核心在于揭示“何时更新信念”（即在证据呈现过程中实时更新，还是仅在证据结束时更新）如何影响位置偏差的产生。作者发现，随着模型代际的更迭，这些位置偏差在较新的模型中反而变得更加明显，这为理解大模型的认知局限性提供了重要视角。

查询时机导致大语言模型与人类在位置偏差上的截然相反

摘要

Query Timing Produces Opposite Positional Biases Between LLMs and Humans

总结

Summary

This paper investigates positional biases (such as primacy and recency effects) in Large Language Models (LLMs) compared to human judgment when evaluating sequential evidence. The authors find that LLMs diverge from human behavior regarding when belief updates occur during evidence presentation, and note that these positional biases have become more pronounced in newer model generations.

论文元数据

  • arXiv ID: arXiv:2608.12387 [cs.CL]
  • 学科领域: 计算与语言 (cs.CL)；人工智能 (cs.AI)
  • 提交日期: 2026年8月1日
  • 作者: Jasin Cekinmez, Addison J. Wu, Thomas L. Griffiths
  • 奖项: Entropic Award (前三名论文), ICBINB @ ICLR 2026
  • arXiv ID: arXiv:2608.12387 [cs.CL]
  • Subject Areas: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
  • Submitted On: August 1, 2026
  • Authors: Jasin Cekinmez, Addison J. Wu, Thomas L. Griffiths
  • Recognition: Entropic Award (Top 3 Paper), ICBINB @ ICLR 2026

摘要

大语言模型（LLMs）中已记录了诸如近因效应和首因效应等位置偏差，但这些模型进行评估的潜在机制仍知之甚少。在人类对证据的判断中，首因效应和近因效应均有观察记录，但近期研究表明，听者“何时”更新其信念——是在证据呈现过程中，还是仅在结束时——会影响此类效应的存在。我们调查了 LLMs 是否存在类似的现象，并发现其与人类行为存在分歧。与前代模型相比，这些偏差在较新的模型中更为加剧。

Abstract

Positional biases such as recency and primacy effects have been documented in large language models (LLMs), yet the underlying mechanism by which these models make their evaluations remains poorly understood. Both primacy and recency biases have been observed in human judgments in response to evidence, but recent work suggest that when the listener updates their beliefs—during the presentation of evidence or only at the end—influences the presence of such effects. We investigate whether a similar phenomenon holds for LLMs, finding divergence from human behavior. These biases are more exacerbated in newer models compared to their predecessors.

访问链接与资源