跳到正文
arXiv cs.AI· Jasin Cekinmez, Addison J. Wu, Thomas L. Griffiths·· 4 小时前AI 评分34

查询时机让 LLM 与人类产生相反的位置偏差

Query Timing Produces Opposite Positional Biases Between LLMs and Humans

AI 导读

研究发现,查询时机会使 LLM 与人类表现出相反的位置偏差:人类在证据呈现过程中更新信念时更易出现近因效应,而 LLM 的行为与之背离。该偏差在较新模型中比前代模型更明显。论文入选 EMNLP 2026 Findings,并获 ICBINB @ ICLR 2026 熵奖(Top 3 论文)。

正文 · 原文

View PDF HTML (experimental)

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 pronounced in newer models compared to their predecessors.
Comments: EMNLP 2026 Findings; Entropic Award (Top 3 Paper), ICBINB @ ICLR 2026
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Cite as: arXiv:2608.12387 [cs.CL]
  (or arXiv:2608.12387v2 [cs.CL] for this version)
  https://doi.org/10.48550/arXiv.2608.12387

arXiv-issued DOI via DataCite

Submission history

From: Addison J. Wu [view email]
[v1] Sat, 1 Aug 2026 03:23:52 UTC (327 KB)
[v2] Tue, 17 Nov 2026 04:46:51 UTC (770 KB)

来源:arXiv cs.AI · arxiv.org