Nov 20, 2022 · We find that conceptor post-processing achieves state-of-the-art (SoTA) debiasing results while maintaining LLMs' performance on the GLUE ...

However, such debiasing often fails to debias effectively and reduces language model performance in downstream tasks (Meade et al., 2022).

The paper use conceptors to reduce social biases in large language models (LLMs) like BERT and GPT, including two approaches: 1) post-processing to remove bias, ...

Oct 30, 2023 · We further show that cocneptor-aided debiasing is robust in differ- ent LLMs, various layers of models, and varied types of biases. Moreover, ...

Nov 20, 2022 · Pre-trained language models reflect the inherent social biases of their training corpus. Many methods have been proposed to mitigate this ...

Co-authors ; Conceptor-Aided Debiasing of Large Language Models. LS Yifei, L Ungar, J Sedoc. EMNLP, 2023. 1, 2023 ...

We propose two methods of applying conceptors (1) bias subspace projection by post-processing by the conceptor NOT operation; and (2) a new architecture, ...

Conceptor-Aided Debiasing of Large Language Models · Li Yifei | Lyle Ungar | João Sedoc · Proceedings of the 2023 Conference on Empirical Methods in Natural ...

Large Language models (LLMs), while powerful, exhibit harmful social biases. Debiasing is often challenging due to computational costs, data constraints, and ...

Conceptor-Aided Debiasing of Large Language Models Li Yifei, Lyle Ungar ... Language Understanding with Contrastive Reading Model and Frozen Large Language Models