Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Yang, Ivy Yuqian, Zhang, David Yu
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2511.14630
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866918208278102016
author Yang, Ivy Yuqian
Zhang, David Yu
author_facet Yang, Ivy Yuqian
Zhang, David Yu
contents Scientific idea generation and selection requires exploration following a target probability distribution. In contrast, current AI benchmarks have objectively correct answers, and training large language models (LLMs) via reinforcement learning against these benchmarks discourages probabilistic exploration. Here, we conducted systematic experiments requesting LLMs to produce outputs following simple probabilistic distributions, and found that all modern LLMs tested grossly fail to follow the distributions. For example, requesting a binary output of "1" 49% of the time produces an answer of "0" nearly 100% of the time. This step function-like behavior of near-exclusively generating the output with marginally highest probability even overrules even strong in-built LLM biases.
format Preprint
id arxiv_https___arxiv_org_abs_2511_14630
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Failure to Mix: Large language models struggle to answer according to desired probability distributions
Yang, Ivy Yuqian
Zhang, David Yu
Machine Learning
Artificial Intelligence
Scientific idea generation and selection requires exploration following a target probability distribution. In contrast, current AI benchmarks have objectively correct answers, and training large language models (LLMs) via reinforcement learning against these benchmarks discourages probabilistic exploration. Here, we conducted systematic experiments requesting LLMs to produce outputs following simple probabilistic distributions, and found that all modern LLMs tested grossly fail to follow the distributions. For example, requesting a binary output of "1" 49% of the time produces an answer of "0" nearly 100% of the time. This step function-like behavior of near-exclusively generating the output with marginally highest probability even overrules even strong in-built LLM biases.
title Failure to Mix: Large language models struggle to answer according to desired probability distributions
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2511.14630