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

Google DeepMind study: Training a model to argue moral dilemmas improves consistency, not judgmentMachine translation

Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.

AI introduction

Researchers used reinforcement learning to train Gemma-4 4B to argue moral dilemmas drawn from classic novels. On ambiguous MoralChoice cases, answer flips after swapping the options fell from 14.3% to 6.6%, while accuracy on moral fables fell from 56.5% to 53.4%. Evaluations of model advice should test consistency and correctness separately.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

谷歌 DeepMind 的一篇新论文发现,训练模型写出更好的道德论证会使它更果断,但不会提高它的道德判断能力。

人们越来越频繁地向大型语言模型寻求道德建议。当选项顺序改变后,模型仍坚持原来的答案,这通常被视为判断可靠的迹象。

他们使用强化学习训练 Gemma-4 4B,让它就取自经典小说的道德困境展开论证。模型将自己的论证两两比较并作出评判,胜出的论证决定奖励。

在含糊不清的 MoralChoice 案例中,交换选项后答案发生变化的比例从 14.3% 降至 6.6%。然而,模型附和用户的比例几乎没有变化,从 45.3% 变为 44.0%;它在道德寓言上的准确率则从 56.5% 降至 53.4%。

如果你评估模型在提供建议或作出判断的任务中的表现,应分别测试其一致性和正确性,因为表现更稳定的模型仍可能出错。

Source: Rohan Paul · x.com

Research
Found an error? Send a correction