iiinova — where your digital self keeps evolving

Reposted by Kawin Ethayarajh

Shion Honda @shionhonda.bsky.social · 24/11/2024

RLHF is not the only method for AI alignment. This post introduces modern algorithms like DPO, KTO, and DiscoPOP that offer simpler and more stable alternatives. Evolution of Preference Optimization Techniques | Hippocampus's Garden hippocampus-garden.com/preference_o...

hippocampus-garden.com

Evolution of Preference Optimization Techniques | Hippocampus's Garden

RLHF is not the only method for AI alignment. This article introduces modern algorithms like DPO and KTO that offer simpler and more stable alternatives.

061

Reposted by Kawin Ethayarajh

Elisa Kreiss @elisakreiss.bsky.social · 24/11/2024

I'm excited to kick off my Bluesky presence with wonderful news: Our paper "Reference-Based Metrics Are Biased Against Blind and Low-Vision Users' Image Description Preferences" won a Best Paper Award at the NLP for Positive Impact Workshop at EMNLP! Read it here: aclanthology.org/2024.nlp4pi-...

aclanthology.org

Reference-Based Metrics Are Biased Against Blind and Low-Vision Users’ Image Description Preferences

Rhea Kapur, Elisa Kreiss. Proceedings of the Third Workshop on NLP for Positive Impact. 2024.

419419

Kawin Ethayarajh @kawinethayarajh.bsky.social · 18/11/2024

Everyone is fixated on replicating o1 when there would be way more utility in figuring out what makes Claude so special.

040