Hongli Zhan ✈️@COLM’24
@honglizhan
PhD Student 🤘@UTAustin | ex- @IBMResearch @sjtu1896 | NLP, emotions, affective computing
ID: 1256575491634458624
http://honglizhan.github.io/ 02-05-2020 13:25:44
64 Tweet
565 Takipçi
863 Takip Edilen
Labeling preferences online for LLM alignment improves DPO vs using static prefs. We show we can use online prefs to train a reward model and label *even more* preferences to train the LLM. D2PO: discriminator-guided DPO Work w/ Nathan Lambert Scott Niekum Tanya Goyal Greg Durrett
I like the potential of LLMs to deliver specific functions, given the right training. Hongli Zhan ✈️@COLM’24 Desmond Ong et al have trained a model to help people think about their problems from alternative perspectives. Excited to see where this goes arxiv.org/abs/2404.01288
🥳Happy to share that our paper "Evaluating Short-Term Temporal Fluctuations of Social Biases in Social Media Data and Masked Language Models" has been accepted at #EMNLP2024 Congrats to my amazing co-authors: Jose Camacho-Collados and Prof. Danushka Bollegala 📜arxiv.org/pdf/2406.13556