BBox-Adapter: Lightweight Adapting for Black-Box Large Language Models
This addresses the problem of adapting opaque, API-based LLMs for researchers and practitioners, offering a cost-effective and privacy-preserving solution, though it is incremental as it builds on adapter methods for black-box models.
The paper tackles the challenge of adapting black-box large language models (LLMs) like GPT-4 for specific tasks without access to internal parameters, by introducing BBox-Adapter, a lightweight adapter that improves model performance by up to 6.77% and reduces training and inference costs by 31.30x and 1.84x, respectively.
Adapting state-of-the-art Large Language Models (LLMs) like GPT-4 and Gemini for specific tasks is challenging. Due to the opacity in their parameters, embeddings, and even output probabilities, existing fine-tuning adaptation methods are inapplicable. Consequently, adapting these black-box LLMs is only possible through their API services, raising concerns about transparency, privacy, and cost. To address these challenges, we introduce BBox-Adapter, a novel lightweight adapter for black-box LLMs. BBox-Adapter distinguishes target and source domain data by treating target data as positive and source data as negative. It employs a ranking-based Noise Contrastive Estimation (NCE) loss to promote the likelihood of target domain data while penalizing that of the source domain. Furthermore, it features an online adaptation mechanism, which incorporates real-time positive data sampling from ground-truth, human, or AI feedback, coupled with negative data from previous adaptations. Extensive experiments demonstrate BBox-Adapter's effectiveness and cost efficiency. It improves model performance by up to 6.77% across diverse tasks and domains, while reducing training and inference costs by 31.30x and 1.84x, respectively.