GIFT framework guides fine-tuning of instruction-tuned language models using confidence signals from the base model

Researchers propose GIFT (Guided Fine-Tuning and Transfer), a framework that incorporates guidance from an instruction-tuned model into task adaptation by fine-tuning a low-rank adapter on a pretrained base model using confidence signals derived from the instruction model.

The learned adapter is merged into the instruction-tuned model to yield task-specialized models that preserve general instruction-following behavior, rather than treating the instruction-tuned model as a passive target only involved at the final merging stage.

GIFT consistently outperforms direct fine-tuning and representative transfer-based baselines on mathematical and knowledge-intensive benchmarks across multiple model families and scales, while maintaining robust generalization and favorable test-time scaling behavior.