Explain the counterproductive effect of instruct vectors after full fine-tuning

Determine why adding an instruct vector to a fully fine-tuned BonLM model is counterproductive, and establish whether instruct-vector compatibility varies across model families beyond the Llama and Qwen families.

Background

The study compares continued pre-training using full fine-tuning (FFT) and parameter-efficient low-rank adaptation (LoRA), with and without an instruct vector derived from an instruction-tuned version of the base model. Adding the instruct vector improves the LoRA-adapted model but decreases the performance of the FFT-adapted model, contrary to earlier findings.

The authors hypothesize that LoRA’s smaller parameter drift preserves greater compatibility with the instruction-tuned model, but they do not establish why the same vector is harmful after FFT. They also note that prior work has focused mainly on Llama and Qwen models, leaving the generality of instruct-vector compatibility across model families unresolved.

References

Why IV is counterproductive for the FFT model remains unclear. Because prior studies have focused primarily on Llama and Qwen models, IV compatibility may vary across model families, but further investigation is needed to confirm this.

— Reading the News: Adapting Large Language Models to Swedish Journalism Through Continued Pre-Training  (2608.30609 - Borggren et al., 31 Aug 2026) in Section 5, subsection “PEFT and Instruct-Vector Compatibility”