A New AI Research Explains How In-Context Instruction Learning (ICIL) Improves The Zero-Shot Task Generalization Performance For Both Pretrained And Instruction-Fine-Tuned Models

Large Language Models (LLMs) have shown they can adapt to target tasks during inference by a process known as few-shot demonstrations, sometimes known as in-context learning. This capability has become increasingly obvious as model sizes scale up, with LLMs displaying emerging features. One emerging talent is the capacity to generalize to unknown tasks by following directions. Instruction tuning, or RLHF, is one of the instruction learning approaches suggested to enhance this capability. Prior research, however, mostly concentrated on instruction-learning techniques based on fine-tuning. The model is multi-task fine-tuned on numerous tasks with instructions, necessitating many backpropagation procedures.

A group of researchers from KAIST and LG Research shows that In-Context Instruction Learning (ICIL), which entails learning to follow instructions during inference through in-context learning, is advantageous for both pretrained models that are readily available and models that have been specifically tuned to follow instructions, as shown in Figure 1. The prompt used by ICIL comprises many cross-task examples, each of which is an instance of a task’s education, input, and output. Since they completely exclude the functions used for demonstrations from the evaluation set and because they employ the same set of protests for all evaluation tasks, treating them as a single fixed prompt, as illustrated in Figure 2, ICIL is a zero-shot learning approach.

They create a fixed example set using a straightforward heuristic-based sampling method that works well for various downstream tasks and model sizes. They can evaluate and duplicate baseline zero-shot performance for new target tasks or models without depending on external tools by prepending the same fixed demonstration set for all jobs. Figure 1 shows that ICIL considerably improves the generalization performance on the zero-shot challenge of various pretrained LLMs that are not fine-tuned to obey instructions.

Their data demonstrate that the selection of classification tasks that feature clear response options in the instruction is what makes ICIL successful. Importantly, even smaller LLMs with ICIL perform better than larger language models without ICIL. For example, the 6B-sized ICIL GPT-J outperforms the 175B-sized Standard Zero-shot GPT-3 Davinci by 30. Second, they demonstrate how adding ICIL to instruction-fine-tuned LLMs enhances their capacity to follow zero-shot instructions, particularly for models with more than 100B elements. This suggests that the impact of ICIL is additive to the impact of instruction modification.

This is true even for generation target tasks, contrary to earlier research suggesting that few-shot in-context learning requires retrieving examples comparable to the target task. Even more surprisingly, they find that performance is not noticeably impacted when random phrases are substituted for the input instance distribution of each example. Based on this approach, they propose that LLMs, rather than depending on the complicated connection between instruction, input, and output, learn the correspondence between the response option provided in the instruction and the production of each demonstration during inference. The purpose of ICIL, according to this theory, is to assist LLMs in focusing on the target instruction to discover the signals for the response distribution of the target task.

Check out the Paper and Github. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 15k+ ML SubReddit, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.