Researchers reframe instruct-model fine-tuning as a direction-selection problem: rather than treating behavioral drift from the reference model as an uncontrolled side effect, they fix a drift budget in advance and study which update directions best improve target performance within that budget. Testing this in a QA-only fine-tuning setting where models must still produce multi-step reasoning at inference, they show a coarse layer-selective probe can reverse the usual failure of QA-only fine-tuning by finding directions that improve target performance while preserving reasoning ability. Across Qwen3-8B and Qwen3-14B and over 100 languages, models tuned along these directions improve scientific reasoning and multilingual translation, matching or beating dedicated translation systems.
