This is over Nan’s head but thought you might like to know.
(Ha Ha, my computer wouldn’t accept “computer morality” in my tags, insisting it should be computer mortality…)
arXiv Forum: How do we make accessible research papers a reality?
Can we truly call it “Open Science” when most research papers are not fully accessible? You are invited to join the forum on Monday April 17 to chart a path towards truly accessible research papers.
Learn More Skip to main content
Computer Science > Computation and Language [Submitted on 15 Feb 2023 (v1), last revised 18 Feb 2023 (this version, v2)]
The Capacity for Moral Self-Correction in Large Language Models
We test the hypothesis that language models trained with reinforcement learning from human feedback (RLHF) have the capability to “morally self-correct” — to avoid producing harmful outputs — if instructed to do so. We find strong evidence in support of this hypothesis across three different experiments, each of which reveal different facets of moral self-correction. We find that the capability for moral self-correction emerges at 22B model parameters, and typically improves with increasing model size and RLHF training. We believe that at this level of scale, language models obtain two capabilities that they can use for moral self-correction: (1) they can follow instructions and (2) they can learn complex normative concepts of harm like stereotyping, bias, and discrimination. As such, they can follow instructions to avoid certain kinds of morally harmful outputs. We believe our results are cause for cautious optimism regarding the ability to train language models to abide by ethical principles.
|Subjects:||Computation and Language (cs.CL)|
|Cite as:||arXiv:2302.07459 [cs.CL]|
|(or arXiv:2302.07459v2 [cs.CL]for this version)|
From <https://arxiv.org/abs/2302.07459> 49 scientists apparently plan for greater accessibility of research papers April 17 is the forum to discuss how this might happen. “You are invited to join the forum,” presumably via the internet.