Dense Coordinate-List Fine-Tuning Induces a Controllable Interference Surface in Vision-Language Models

In three linesFine-tuning vision-language models on dense coordinate lists improves visual grounding but induces spurious repetitions. On Gemma 4 12B, high-capacity LoRA raises F1@0.3 from 0.007 to 0.448 but creates duplicate rate 0.080. Object-level control removes repetitions (rate 0.000) while preserving performance (F1 0.490).

Summary generated by Claude — human-verified