A new AI study describes how in-context instructional learning (ICIL) improves zero-shot task generalization performance of both pre-trained and instruction-fine-tuning models

We have shown that large-scale language models (LLMs) can adapt to target tasks during inference by a process known as few-shot demonstration (also known as in-context learning). This feature becomes more and more apparent as the size of the model increases, and new features appear in LLM. One of the new talents is the ability to generalize unknown tasks according to instructions. Instruction Tuning (RLHF) is one instruction learning approach proposed to enhance this capability. However, previous research has mainly focused on fine-tuning-based teaching-learning techniques. The model has been multitask fine-tuned with a large number of tasks using instructions, requiring many backpropagation steps.

A group of researchers from KAIST and LG Research have shown that in-context instructional learning (ICIL), which learns to follow instructions during reasoning through in-context learning, as shown in Figure 1, favors both ready-to-use pretrained models and models specifically tuned to follow instructions. The prompts used in ICIL consist of many cross-tasking examples, each an instance of teaching, input, and output of a task. As shown in Figure 2, ICIL is a zero-shot learning approach because it completely excludes the function used for demonstration from the evaluation set and treats it as a single fixed prompt using the same set of protests for all evaluation tasks.

They create a fixed sample set using a simple heuristic-based sampling method suitable for different downstream tasks and model sizes. By adding the same fixed set of demonstrations to all jobs, you can evaluate and replicate the baseline zero-shot performance of new target tasks or models without relying on external tools. Figure 1 shows that ICIL significantly improves the generalization performance in the zero-shot challenge for various pre-trained LLMs that have not been fine-tuned to follow the instructions.

Their data show that choosing a classification task with clear response options to instructions leads to ICIL success. Importantly, even a small LLM with ICIL performs better than a large language model without ICIL. For example, the 6B size ICIL GPT-J has 30 better performance than the 175B size standard zero-shot GPT-3 Davinci. He then demonstrates how adding his ICIL to the LLM with fine-tuned instructions enhances the ability to follow zero-shot instructions, especially for models with over 100B elements. This suggests that the impact of ICIL is additive to that of instruction change.

This also applies to spawn target tasks. This is in contrast to previous work that suggested that small-shot in-context learning requires obtaining samples comparable to the target task. Even more surprisingly, we found that replacing the distribution of input instances for each example with random phrases did not noticeably affect performance. Based on this approach, they propose that LLM learns the correspondence between response options provided in instructions and the generation of each demonstration during inference, rather than relying on complex relationships between instructions, inputs, and outputs. According to this theory, the purpose of ICIL is to assist LLMs in finding signals in the target task’s response distribution, focusing on target instructions.

Please check paper and github. All credit for this research goes to the researchers of this project.Also, don’t forget to participate 15,000+ ML SubReddits, Discord channeland email newsletterShare the latest AI research news, cool AI projects, and more.