Query timing flips positional bias between LLMs and humans, arXiv study finds
A preprint reports that large language models diverge from humans in when recency and primacy biases appear, depending on when beliefs are updated during evidence presentation.
In plain words
When judging evidence, people and language models can favor items at the start (primacy) or end (recency) of a list. The authors tested whether the timing of belief updates shapes these biases, and found that language models behave differently from humans in this respect.
What was announced
The arXiv preprint examines positional biases such as recency and primacy effects, which have been documented in large language models, while noting that the mechanism behind how these models reach evaluations remains poorly understood. Drawing on prior work suggesting that whether a listener updates beliefs during evidence presentation or only at the end influences such effects in human judgments, the authors investigated whether the same pattern holds for LLMs. They report finding divergence from human behavior, with query timing producing opposite positional biases between the two.
The abstract also states that these biases are more pronounced in newer models compared to their predecessors, though the excerpt cuts off before detailing the comparison or the experimental setup. The paper is available as a v2 replacement announcement on arXiv, and the findings are limited to what the abstract discloses; no independent evaluation of the results is included in this source.
What it does not show
- The source is an abstract only; the full paper's methods, models tested, and detailed results are not available in the supplied passage.
- The claim that biases are more pronounced in newer models is truncated mid-sentence in the source, so the precise comparison is not fully stated.
- No independent confirmation of the findings is provided; the evidence reflects the authors' own reported results.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.