Publications

Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Y. Dubois, Xinlei Chen, Xiaolong Wang, Sanmi Koyejo, Tatsunori Hashimoto, Carlos Guestrin

TLDR: A new language model layer that is more expressive than RNN but more efficient than attention.

Y. Dubois, B. Galambosi, P. Liang, T. Hashimoto

TLDR: We decrease the bias of AlpacaEval for longer outputs using regression analysis.

Y. Zhang, Y. Mai, J. Somerville Roberts, R. Bommasani, Y. Dubois, P. Liang

TLDR: Multidimensional evaluation of instruction following LLM with absolute scores.

Y. Ruan*, H. Dong*, A. Wang, S. Pitis, Y. Zhou, J. Ba, Y. Dubois, C. J. Maddison, T. Hashimoto

TLDR: An LM-based emulation framework for identifying the risks of LM agents at scale.

X. Li*, T. Zhang*, Y. Dubois*, R. Taori*, I. Gulrajani, C. Guestrin, P. Liang, T. Hashimoto

TLDR: A validated automatic evaluator for instruction-following language models. High-quality, cheap, and fast.

Y. Dubois*, X. Li*, R. Taori*, T. Zhang*, I. Gulrajani, J. Ba, C. Guestrin, P. Liang, T. Hashimoto

TLDR: AlpacaFarm replicates the RLHF process at a fraction of the time (<24h) and cost ($<200), enabling the research community to advance instruction following research.

R. Taori*, I. Gulrajani*, T. Zhang*, Y. Dubois*, X. Li*, C. Guestrin, P. Liang, T. Hashimoto

TLDR: We introduce Alpaca 7B, a instruction-following fine-tuned LLaMA model. On our preliminary evaluation, Alpaca behaves qualitatively similarly to OpenAI’s text-davinci-003, while being surprisingly small and easy/cheap to reproduce (<600$).

Y. Dubois, T. Hashimoto, P. Liang

TLDR: We derive a risk decomposition for self-supervised learning and use it to evaluate 169 pretrained models.

S. Santurkar, Y. Dubois, R. Taori, P. Liang, T. Hashimoto

TLDR: Our work performs a systematic investigation into whether additional language supervision (in CLIP) helps models learn more transferrable representations.

N. Miao, E. Mathieu, Y. Dubois, T. Rainforth, Y. W. Teh, A. Foster, H. Kim

TLDR: We introduce a method for automatically learning input-specific augmentations from data.

Y. Dubois, T. Hashimoto, S. Ermon, P. Liang

TLDR: We characterize idealized self-supervised representations, which leads to actionable insights for improving SSL algorithms.

Y Ruan*, Y. Dubois*, C. J. Maddison

TLDR: We give a simple variational objective whose optima are exactly the set of representations that are robust under covariate shift.

Y. Dubois, B. Bloem-Reddy, K. Ullrich, C. J. Maddison

TLDR: We formalize compression with respect to ML algorithms rather than human perception.

Y. Dubois, D. Kiela, D. J. Schwab, R. Vedantam

TLDR: We characterize and approximate optimal representations for supervised learning.

Y. Dubois*, J. Gordon*, A. Foong*

TLDR: A simple and unifying explanation of the neural process family, which are a collection of models that meta-learn a distribution over predictors.

A. Y. K. Foong*, W. P. Bruinsma*, J. Gordon*, Y. Dubois, J. Requeima, R. E. Turner

TLDR: We propose a translation equivariant (latent) neural process.

J. Gordon*, W. P. Bruinsma*, A. Y. K. Foong, J. Requeima,Y. Dubois, R. E. Turner

TLDR: We propose a translation equivariant conditional neural process.

Y. Dubois, Gautier Dagan, Dieuwke Hupkes, Elia Bruni

TLDR: We propose an attention that improves extrapolation capacity of neural NLP models.