Jiuding Sun

#LLMs are trained to predict the next word. But we found that they are overachievers! When GPT-J is busy predicting the next word, our Future Lens (future.baulab.info) sees that it's already thinking several words ahead – see "Back to the Future" below... #CoNLL2023 🧵 1/8

✨ Check out the full paper for details about our survey, error breakdown, and findings! arxiv.org/abs/2305.06299

We also share all the 319 valid instructions we collected over MMLU and BBL to facilitate future research!

An approach to addressing: we introduce a soft-prompt alignment method that adjusts rephrased instructions and improves the robustness of zero-shot models on a variety of tasks

How robust are the instructions in your instruction-tuned model? In our most recent work (w/ @ChantalShaib and @byron_c_wallace), we show that there is a considerable dip in performance on in-domain tasks when you slightly vary the instruction. arxiv.org/abs/2306.11270