GIFT: Guided Fine-Tuning and Transfer for Enhancing Instruction-Tuned Language Models

TL;DR

GIFT is a framework that guides task-specific fine-tuning of instruction-tuned language models by incorporating confidence signals from the instruction model, improving performance across benchmarks.

Contribution

It introduces a simple, efficient method for guiding fine-tuning using confidence signals, enhancing task adaptation without losing instruction-following abilities.

Findings

GIFT outperforms direct fine-tuning on mathematical and knowledge benchmarks.

It maintains robust generalization and test-time scaling behavior.

GIFT effectively merges learned adapters into instruction-tuned models.

Abstract

A promising paradigm for adapting instruction-tuned language models is to learn task-specific updates on a pretrained base model and subsequently merge them into the instruction-tuned model. However, existing approaches typically treat the instruction-tuned model as a passive target that is only involved at the final merging stage, without guiding the training process. We propose GIFT (Guided Fine-Tuning and Transfer), a simple and efficient framework that incorporates guidance from the instruction model into task adaptation. GIFT fine-tunes a low-rank adapter on the pretrained base model using confidence signals derived from the instruction-tuned model. The learned adapter is then merged into the instruction-tuned model, yielding task-specialized models that preserve general instruction-following behavior. We evaluate GIFT on mathematical and knowledge-intensive benchmarks across multiple model families and scales. Results show that GIFT consistently outperforms direct fine-tuning and representative transfer-based baselines, while maintaining robust generalization and favorable test-time scaling behavior.