Drift-Constrained Optimization: Only Direction Matters in Fine-Tuning Instruct Models
The paper argues that fine-tuning quality depends on the direction of model change, not just the size of the drift.
The authors set a fixed “behavioral drift” budget from a reference instruct model, then treat fine-tuning as choosing the most efficient update direction inside that limit. In QA-only experiments, a coarse layer-selective probe found directions that improved target performance while preserving reasoning and general abilities. On Qwen3-8B and Qwen3-14B, those updates improved scientific reasoning and multilingual translation, including results across more than 100 languages. HF Daily Papers' note
The authors set a fixed “behavioral drift” budget from a reference instruct model, then treat fine-tuning as choosing the most efficient update direction inside that limit. In QA-only experiments, a coarse layer-selective probe found directions that improved target performance while preserving reasoning and general abilities. On Qwen3-8B and Qwen3-14B, those updates improved scientific reasoning and multilingual translation, including results across more than 100 languages. HF Daily Papers' note
score 4