Efficiently Learning at Test-Time: Active Fine-Tuning of LLMs

Recent efforts in fine-tuning language models often rely on automatic data selection, commonly using Nearest Neighbors retrieval from large datasets. However, we theoretically show that this approach tends to select redundant data, limiting its effectiveness or even hurting performance. To address this, we introduce SIFT, a data selection algorithm designed to reduce uncertainty about the model’s response given a prompt, which unifies ideas from retrieval and active learning. Whereas Nearest Neighbor retrieval typically fails in the presence of information duplication, SIFT accounts for information duplication and optimizes the overall information gain of the selected examples. We focus our evaluations on fine-tuning at test-time for prompt-specific language modeling on the Pile dataset, and show that SIFT consistently outperforms Nearest Neighbor retrieval, with minimal computational overhead. Moreover, we show that our uncertainty estimates can predict the performance gain of test-time fine-tuning, and use this to develop an adaptive algorithm that invests test-time compute proportional to realized performance gains. We provide the activeft (Active Fine-Tuning) library which can be used as a drop-in replacement for Nearest Neighbor retrieval.

The standard paradigm of machine learning separates training and testing. Training aims to learn a model by inductively extracting general rules from data, and testing applies this model to new, unseen data. We investigate an alternative transductive paradigm where the model is fine-tuned at test-time specifically to the given task. Variations of this paradigm have been studied since the inception of machine learning as a field. Early examples are local learning (Cleveland, 1979; Cleveland & Devlin, 1988; Atkeson et al., 1997) and local fine-tuning (Bottou & Vapnik, 1992). More recently, with the advent of large pre-trained models which have good representations and are strong foundations for fine-tuning, the idea of test-time fine-tuning has re-gained attention (Krause et al., 2018; 2019; Sun et al., 2020). Hardt & Sun (2024) show that fine-tuning on data related to the prompt to a large language model (LLM) can significantly improve performance.