Continual Learning Using Only Large Language Model Prompting

TL;DR

This paper tackles continual learning when only black-box LLM prompting is allowed, formulated as class-incremental learning ($CIL$) with no parameter updates. It introduces CLOB (Continual Learning Over Black-box LLMs) and CIS (in-context CL via Incremental Summarization), where knowledge for each class is stored as compact per-class summaries that are incrementally updated as new data arrives. Empirical results on four text-classification datasets show that CIS greatly outperforms strong baselines and approaches the upper bounds of joint prompting/fine-tuning, while exhibiting near-zero prompt-based forgetting. The work demonstrates a practical route to CL via LLM APIs and motivates extending summary-based representations to other modalities and longer-context models.

Abstract

We introduce CLOB, a novel continual learning (CL) paradigm wherein a large language model (LLM) is regarded as a black box. Learning is done incrementally via only verbal prompting. CLOB does not fine-tune any part of the LLM or add any trainable parameters to it. It is particularly suitable for LLMs that are accessible via APIs. We also propose a new CL technique, called CIS, based on incremental summarization that also overcomes the LLM's input length limit. Experiments show CIS outperforms baselines by a very large margin.

Figures (1)

• Figure 1: (1) Left: Overview of CIS in CLOB. (2) Right: Prompts used in each component of learning. Full prompts can be found in Appendix \ref{['sec:appendix.prompts']}. Some example summaries are given in Appendix \ref{['sec:appendix.examples']}.