aifollow.news
Back Rohan Paul
Rohan Paul· @rohanpaul_ai · X· · Original publication time AI score54

Max Planck study finds reasoning makes models more likely to err togetherMachine translation

AI introduction

Researchers tested nine Qwen and GPT models on three prediction tasks built from US census surveys, with reasoning on and off. With reasoning on, agreement increased and the share of people all nine models got wrong rose by 2.0 to 2.4 percentage points; among those with the longest reasoning traces, adjusted agreement fell to about 0.31 to 0.39.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

马克斯·普朗克研究所的一篇新论文发现,开启推理功能会让不同的大语言模型更趋于一致,即使答案是错的;而较长的推理轨迹则标示出它们在哪些地方存在分歧。

当几个大语言模型对同一个人作出判断时,一个模型给出的错误答案可能通过换用另一个模型来纠正。但这只有在各模型并非都犯同样错误时才行。

作者使用美国人口普查调查数据构建了3项预测任务,测试了9个Qwen和GPT模型在开启和关闭推理功能时的表现。他们衡量了超出仅凭各模型准确率所能预测的模型间一致程度。

开启推理功能后,模型间的一致程度提高,9个模型全都答错的人所占比例上升了2.0至2.4个百分点。对于引发最长推理轨迹的人,调整后的一致程度降至约0.31至0.39。

对于涉及人的决策,应检查你的模型组合中是否有模型能给出正确答案,因为更高的一致程度也可能意味着更多共同的错误。

Source: Rohan Paul · x.com

Research
Found an error? Send a correction